FV609-Simulation Using ProModel - Charles Harrell

T H I R D E D I T I O N
SIMULATION USING
PROMODEL
Dr. Charles Harrell

Professor, Brigham Young University, Provo, Utah
Director, PROMODEL Corporation, Orem, Utah
Dr. Biman K. Ghosh, Project Leader

Professor, California State Polytechnic University,
Pomona, California
Dr. Royce O. Bowden, Jr.

Professor and head, Mississippi State University,
Mississippi State, Mississippi
TM
har01307_fm_i-xviii.indd i 12/21/10 8:25 AM

TM
SIMULATION USING PROMODEL, THIRD EDITION

Published by McGraw-Hill, a business unit of The McGraw-Hill Companies, Inc., 1221 Avenue of the Americas,
New York, NY 10020. Copyright © 2012, 2004, 2000 by The McGraw-Hill Companies, Inc. All rights reserved.
No part of this publication may be reproduced or distributed in any form or by any means, or stored in a database
or retrieval system, without prior written consent of The McGraw-Hill Companies, Inc., including, but not limited
to, in any network or other electronic storage or transmission, or broadcast for distance learning.
Some ancillaries, including electronic and print components, may not be available to customers outside the
United States.
This book is printed on acid-free paper.
1 2 3 4 5 6 7 8 9 0 DOC/DOC 1 0 9 8 7 6 5 4 3 2 1
ISBN 0-07-340130-7
MHID 978-0-07-340130-0
Vice President & Editor-in-Chief: Marty Lange
Vice President EDP/Central Publishing Services: Kimberly Meriwether David
Publisher: Raghu Srinivasan
Sponsoring Editor: Peter E. Massar
Development Editor: Lora Neyens
Marketing Manager: Curt Reynolds
Project Manager: Melissa Leick
Design Coordinator: Brenda A. Rolwes
Cover Design: Studio Montage, St. Louis, Missouri
Cover Image: © Paul H. McGown for ProModel Corporation
Buyer: Nicole Baumgartner
Media Project Manager: Balaji Sundararaman
Compositor: MPS Limited, a Macmillan Company
Typeface: 10/12 Times Roman
Printer: R. R. Donnelley
Library of Congress Cataloging-in-Publication Data

Harrell, Charles, 1950-
Simulation using ProModel / Charles Harrell, Biman K. Ghosh, Royce O. Bowden, Jr.—3rd ed.
p. cm
Includes index.
ISBN 978-0-07-340130-0
1. Computer simulation. 2. ProModel. I. Ghosh, Biman K. II. Bowden, Royce. III. Title.
QA76.9.C65H355 2010
003'.35369—dc22
2010039566
www.mhhe.com
har01307_fm_i-xviii.indd ii 12/21/10 8:25 AM

Dedication
To Paula, Raja, and Pritha
B.K.G.
To Yvonne, Emily, Elizabeth, Richard, Ryan, and Polly
C.H.
To CeCelia, Taylor, and Robert
R.O.B.
har01307_fm_i-xviii.indd iii 12/21/10 8:25 AM

A B O U T T H E A U T H O R S
Charles Harrell
Charles Harrell is an associate professor of engineering and technology at Brigham
Young University and founder of PROMODEL Corporation. He received his
B.S. from Brigham Young University, M.S. in industrial engineering from the
University of Utah, and Ph.D. in manufacturing engineering from the Techni-
cal University of Denmark. His area of interest and expertise is in manufactur-
ing system design and simulation. He has worked in manufacturing and systems
engineering positions for Ford Motor Company and the Eaton Corporation. At
BYU he teaches courses in manufacturing systems, manufacturing simulation,
and manufacturing information systems. He is the author or coauthor of several
simulation books and has given numerous presentations on manufacturing system
design and simulation. He serves on the board of directors of PROMODEL Cor-
poration. He is a senior member of IIE and SME. He enjoys sports and playing
with his grandchildren.
Biman K. Ghosh
Biman K. Ghosh is a professor of industrial and manufacturing engineering at
the California State Polytechnic University, Pomona. He has been teaching and
consulting all over the United States for the past 25 years. Prior to that Dr. Ghosh
worked for such multinational companies as Tata, Mercedes Benz, Sandvik, and
Siemens. Dr. Ghosh has been a consultant to Northrop Grumman, McDonnell
Douglas, Boeing, Rockwell, Loral, Visteon, UPS, Sikorsky, Baxter, Paccar, John-
son Controls, Powersim, and Galorath. Dr. Ghosh has published extensively. His
teaching and consulting interests are in the areas of simulation, supply chain man-
agement, lean manufacturing, flexible manufacturing, cellular manufacturing,
computer-aided manufacturing, design for manufacturing, design of experiments,
total quality management, and work measurement. His passion is still and video
photography.
Royce O. Bowden
Royce Bowden is professor and head of Industrial and Systems Engineering
at Mississippi State University. Prior to joining Mississippi State, Dr. Bowden
worked in manufacturing engineering positions with Texas Instruments and Martin
Marietta Aerospace. His research program centers on the use, development, and
integration of simulation, artificial intelligence, and statistical techniques to solve
systems engineering problems and has been funded by organizations such as the
National Science Foundation, United States Department of Transportation, and
the National Space and Aeronautics Administration. Dr. Bowden’s research pro-
vided the foundation for the first widely used commercial simulation optimization
software. He is the author (or coauthor) of numerous research papers and two
simulation books. Dr. Bowden serves on the executive board of the Boy Scouts
of America Pushmataha Area Council and enjoys organizing and participating in
high adventure outdoor programs for the youth.
iv
har01307_fm_i-xviii.indd iv 12/21/10 8:25 AM

B R I E F C O N T E N T S
PART I
STUDY CHAPTERS
1 Introduction to Simulation 3
2 System Dynamics 33
3 Simulation Basics 57
4 Discrete-Event Simulation 87
5 Data Collection and Analysis 115
6 Model Building 161
7 Model Verification and Validation 193
8 Simulation Output Analysis 211
9 Comparing Systems 243
10 Simulation Optimization 273
11 Modeling Manufacturing Systems 299
12 Modeling Material Handling Systems 323
13 Modeling Service Systems 345
PART II
LABS
1 Introduction to ProModel 365
2 Building Your First Model 379
3 ProModel’s Output Viewer 405
4 Basic Modeling Concepts 419
5 Fitting Statistical Distributions to Input Data 465
6 Intermediate Model Building 475
9 Comparing Alternative Systems 549
10 Simulation Optimization with SimRunner 569
12 Material Handling Concepts 617
Appendix A Continuous and Discrete Distributions in ProModel 669
Appendix B Critical Values for Student’s t Distribution and Standard
Normal Distribution 674
Appendix C F Distribution for ␣ ⫽ 0.05 675
Appendix D Critical Values for Chi-Square Distribution 676
Index 677
har01307_fm_i-xviii.indd v 12/21/10 8:25 AM

C O N T E N T S
PART I 2.8 The Systems Approach 46

2.8.1 Identifying Problems and
STUDY CHAPTERS
Opportunities 47
2.8.2 Developing Alternative
1 Introduction to Simulation 3 Solutions 47
1.1 Introduction 3 2.8.3 Evaluating the
1.2 What Is Simulation? 5 Solutions 48
1.3 Why Simulate? 6 2.8.4 Selecting and Implementing
1.4 Doing Simulation 8 the Best Solution 48
1.5 Use of Simulation 10 2.9 Systems Analysis Techniques 48
1.6 When Simulation Is Appropriate 11 2.9.1 Hand Calculations 50
1.7 Qualifications for Doing 2.9.2 Spreadsheets 50
Simulation 13 2.9.3 Operations Research
1.8 Conducting a Simulation Study 14 Techniques 51
1.8.1 Defining the Objective 15 2.9.4 Special Computerized
1.8.2 Planning the Study 18 Tools 54
1.9 Economic Justification of 2.10 Summary 55
Simulation 18 2.11 Review Questions 55
1.10 Sources of Information on References 56
Simulation 22
1.11 How to Use This Book 22 3 Simulation Basics 57
1.12 Summary 23 3.1 Introduction 57
1.13 Review Questions 24 3.2 Types of Simulation 57
1.14 Case Studies 25 3.2.1 Static versus Dynamic
References 30 Simulation 58
3.2.2 Stochastic versus
2 System Dynamics 33 Deterministic
2.1 Introduction 33 Simulation 58
2.2 System Definition 34 3.2.3 Discrete-Event versus
2.3 System Elements 35 Continuous Simulation 59
2.3.1 Entities 36 3.3 Random Behavior 61
2.3.2 Activities 36 3.4 Simulating Random Behavior 63
2.3.3 Resources 36 3.4.1 Generating Random
2.3.4 Controls 37 Numbers 63
2.4 System Complexity 37 3.4.2 Generating Random
2.4.1 Interdependencies 38 Variates 68
2.4.2 Variability 39 3.4.3 Generating Random
2.5 System Performance Metrics 41 Variates from
2.6 System Variables 43 Common Continuous
2.6.1 Decision Variables 43 Distributions 71
2.6.2 Response Variables 44 3.4.4 Generating Random
2.6.3 State Variables 44 Variates from
2.7 System Optimization 44 Common Discrete
Distributions 73
har01307_fm_i-xviii.indd vi 12/21/10 8:25 AM

Contents vii
3.5 Simple Spreadsheet Simulation 74 4.8 Summary 112

3.5.1 Simulating Random 4.9 Review Questions 113
Variates 75 References 114
3.5.2 Simulating Dynamic,
Stochastic Systems 79 5 Data Collection and Analysis 115
3.5.3 Simulation Replications and 5.1 Introduction 115
Output Analysis 82 5.2 Guidelines for Data Gathering 116
3.6 Summary 83 5.3 Determining Data
3.7 Review Questions 83 Requirements 118
References 85 5.3.1 Structural Data 118
5.3.2 Operational Data 118
5.3.3 Numerical Data 119
4.1 Introduction 87 5.3.4 Use of a Questionnaire 119
4.2 How Discrete-Event Simulation 5.4 Identifying Data Sources 120
Works 88 5.5 Collecting the Data 121
4.3 A Manual Discrete-Event Simulation 5.5.1 Defining the Entity
Example 89 Flow 121
4.3.1 Simulation Model 5.5.2 Developing a Description
Assumptions 91 of Operation 122
4.3.2 Setting Up the 5.5.3 Defining Incidental
Simulation 91 Details and Refining Data
4.3.3 Running the Values 123
Simulation 94 5.6 Making Assumptions 124
4.3.4 Calculating Results 99 5.7 Statistical Analysis of Numerical
4.3.5 Issues 102 Data 125
4.4 Commercial Simulation 5.7.1 Tests for
Software 102 Independence 127
4.4.1 Modeling Interface 5.7.2 Tests for Identically
Module 102 Distributed Data 132
4.4.2 Model Processor 103 5.8 Distribution Fitting 134
4.4.3 Simulation Interface 5.8.1 Frequency
Module 103 Distributions 135
4.4.4 Simulation Processor 104 5.8.2 Theoretical
4.4.5 Animation Processor 104 Distributions 138
4.4.6 Output Processor 105 5.8.3 Fitting Theoretical
4.4.7 Output Interface Distributions to Data 142
Module 105 5.9 Selecting a Distribution in the
4.5 Simulation Using ProModel 106 Absence of Data 148
4.5.1 Building a Model 106 5.9.1 Most Likely or Mean
4.5.2 Running the Value 149
Simulation 106 5.9.2 Minimum and Maximum
4.5.3 Output Analysis 107 Values 149
4.6 Languages versus Simulators 109 5.9.3 Minimum, Most Likely, and
4.7 Future of Simulation 111 Maximum Values 149
har01307_fm_i-xviii.indd vii 12/21/10 8:25 AM

viii Contents
5.10 Bounded versus Boundless 7 Model Verification and

Distributions 151 Validation 193
5.11 Modeling Discrete Probabilities
7.1 Introduction 193
Using Continuous Distributions 151
7.2 Importance of Model Verification and
5.12 Data Documentation and
Validation 194
Approval 152
7.2.1 Reasons for Neglect 194
5.12.1 Data Documentation
7.2.2 Practices That Facilitate
Example 152
Verification and
5.13 Summary 155
Validation 195
5.14 Review Questions 155
7.3 Model Verification 196
5.15 Case Study 158
7.3.1 Preventive Measures 197
References 159
7.3.2 Establishing a Standard for
6 Model Building 161 Comparison 198
7.3.3 Verification
6.1 Introduction 161 Techniques 198
6.2 Converting a Conceptual Model to a 7.4 Model Validation 202
Simulation Model 162 7.4.1 Determining Model
6.2.1 Modeling Paradigms 162 Validity 203
6.2.2 Model Definition 164 7.4.2 Maintaining
6.3 Structural Elements 165 Validation 205
6.3.1 Entities 165 7.4.3 Validation Examples 205
6.3.2 Locations 167 7.5 Summary 209
6.3.3 Resources 169 7.6 Review Questions 210
6.3.4 Paths 171 References 210
6.4 Operational Elements 171
6.4.1 Routings 171 8 Simulation Output Analysis 211
6.4.2 Entity Operations 172
6.4.3 Entity Arrivals 175
8.2 Statistical Analysis of Simulation
6.4.4 Entity and Resource
Output 212
Movement 177
8.2.1 Simulation
6.4.5 Accessing Locations and
Replications 213
Resources 178
8.2.2 Performance
6.4.6 Resource Scheduling 180
Estimation 214
6.4.7 Downtimes and Repairs 181
8.2.3 Number of Replications
6.4.8 Use of Programming
(Sample Size) 218
Logic 185
8.2.4 Real-World Experiments
6.5 Miscellaneous Modeling Issues 187
versus Simulation
6.5.1 Modeling Rare
Experiments 221
Occurrences 187
8.3 Statistical Issues with Simulation
6.5.2 Large-Scale Modeling 188
Output 222
6.5.3 Cost Modeling 189
8.4 Terminating and Nonterminating
6.6 Summary 190
Simulations 225
8.4.1 Terminating
References 192
Simulations 226
har01307_fm_i-xviii.indd viii 12/21/10 8:25 AM

Contents ix
8.4.2 Nonterminating More Than Two Alternative

Simulations 226 Systems 258
8.5 Experimenting with Terminating 9.4.3 Factorial Design and
Simulations 227 Optimization 264
8.5.1 Selecting the Initial Model 9.5 Variance Reduction Techniques 265
State 228 9.5.1 Common Random
8.5.2 Selecting a Terminating Numbers 266
Event to Control Run 9.5.2 Example Use of Common
Length 228 Random Numbers 268
8.5.3 Determining the Number 9.5.3 Why Common Random
of Replications 228 Numbers Work 270
8.6 Experimenting with Nonterminating 9.6 Summary 271
Simulations 229 9.7 Review Questions 271
8.6.1 Determining the Warm-up References 272
Period 229
8.6.2 Obtaining Sample 10 Simulation Optimization 273
Observations 234
8.6.3 Determining Run
10.2 In Search of the Optimum 275
Length 239
10.3 Combining Direct Search Techniques
8.7 Summary 240
with Simulation 276
10.4 Evolutionary Algorithms 277
References 242
10.4.1 Combining Evolutionary
Algorithms with
Simulation 278
9.1 Introduction 243 10.4.2 Illustration of an
9.2 Hypothesis Testing 244 Evolutionary Algorithm’s
9.3 Comparing Two Alternative System Search of a Response
Designs 247 Surface 279
9.3.1 Welch Confidence Interval 10.5 Strategic and Tactical Issues of
for Comparing Two Simulation Optimization 281
Systems 248 10.5.1 Operational
9.3.2 Paired-t Confidence Efficiency 281
Interval for Comparing Two 10.5.2 Statistical Efficiency 282
Systems 250 10.5.3 General Optimization
9.3.3 Welch versus the Paired-t Procedure 282
Confidence Interval 252 10.6 Formulating an Example
9.4 Comparing More Than Two Optimization Problem 284
Alternative System Designs 253 10.6.1 Problem Description 285
9.4.1 The Bonferroni Approach 10.6.2 Demonstration of the
for Comparing More General Optimization
Than Two Alternative Procedure 286
Systems 253 10.7 Real-World Simulation Optimization
9.4.2 Advanced Statistical Project 289
Models for Comparing 10.7.1 Problem Description 289
har01307_fm_i-xviii.indd ix 12/21/10 8:25 AM

x Contents
10.7.2 Simulation Model and 11.7 Summary 320

Performance Measure 290 11.8 Review Questions 320
10.7.3 Toyota Solution References 320
Technique 291
10.7.4 Simulation Optimization 12 Modeling Material Handling
Technique 292 Systems 323
10.7.5 Comparison of Results 292 12.1 Introduction 323
10.8 Summary 294 12.2 Material Handling Principles 323
10.9 Review Questions 294 12.3 Material Handling
References 295 Classification 324
12.4 Conveyors 325
11 Modeling Manufacturing
12.4.1 Conveyor Types 325
Systems 299
12.4.2 Operational
11.1 Introduction 299 Characteristics 327
11.2 Characteristics of Manufacturing 12.4.3 Modeling Conveyor
Systems 300 Systems 328
11.3 Manufacturing Terminology 301 12.4.4 Modeling Single-Section
11.4 Use of Simulation in Conveyors 329
Manufacturing 303 12.4.5 Modeling Conveyor
11.5 Applications of Simulation in Networks 330
Manufacturing 304 12.5 Industrial Vehicles 330
11.5.1 Methods Analysis 305 12.5.1 Modeling Industrial
11.5.2 Plant Layout 306 Vehicles 331
11.5.3 Batch Sizing 308 12.6 Automated Storage/Retrieval
11.5.4 Production Control 309 Systems 331
11.5.5 Inventory Control 312 12.6.1 Configuring an
11.5.6 Supply Chain AS/RS 332
Management 313 12.6.2 Modeling AS/RSs 334
11.5.7 Production 12.7 Carousels 335
Scheduling 314 12.7.1 Carousel
11.5.8 Real-Time Control 315 Configurations 335
11.5.9 Emulation 315 12.7.2 Modeling Carousels 335
11.6 Manufacturing Modeling 12.8 Automatic Guided Vehicle
Techniques 316 Systems 336
11.6.1 Modeling Machine 12.8.1 Designing an AGVS 337
Setup 316 12.8.2 Controlling an AGVS 338
11.6.2 Modeling Machine Load 12.8.3 Modeling an AGVS 339
and Unload Time 316 12.9 Cranes and Hoists 340
11.6.3 Modeling Rework and 12.9.1 Crane Management 340
Scrap 317 12.9.2 Modeling Bridge
11.6.4 Modeling Transfer Cranes 340
Machines 317 12.10 Robots 341
11.6.5 Continuous Process 12.10.1 Robot Control 341
Systems 319 12.10.2 Modeling Robots 342
har01307_fm_i-xviii.indd x 12/21/10 8:25 AM

Contents xi
12.11 Summary 343 L1.4 Run-Time Menus and

12.12 Review Questions 343 Controls 368
References 344 L1.5 Simulation in Decision
Making 369
13 Modeling Service Systems 345 L1.5.1 California Cellular 370
13.1 Introduction 345 L1.5.2 ATM System 373
13.2 Characteristics of Service L1.6 Exercises 376
Systems 346
2 Building Your First Model 379
13.3 Performance Measures 347
13.4 Use of Simulation in Service L2.1 Building Your First Simulation
Systems 348 Model 379
13.5 Applications of Simulation in Service L2.2 Building the Bank of USA ATM
Industries 350 Model 387
13.5.1 Process Design 350 L2.3 Locations, Entities, Processing,
13.5.2 Method Selection 350 and Arrivals 392
13.5.3 System Layout 351 L2.4 Add Location 396
13.5.4 Staff Planning 351 L2.5 Effect of Variability on Model
13.5.5 Flow Control 352 Performance 397
13.6 Types of Service Systems 352 L2.6 Blocking 398
13.6.1 Service Factory 352 L2.7 Effect of Traffic Intensity on
13.6.2 Pure Service Shop 353 System Performance 401
13.6.3 Retail Service Store 353 L2.8 Exercises 402
13.6.4 Professional Service 354
13.6.5 Telephonic Service 354 3 ProModel’s Output Viewer 405
13.6.6 Delivery Service 355
L3.1 The Output Viewer 405
13.6.7 Transportation Service 355
L3.2 Report Tables 407
13.7 Simulation Example: A Help Desk
L3.3 Column Charts 408
Operation 355
L3.4 Utilization Charts 409
13.7.1 Background 356
L3.5 State Charts 409
13.7.2 Model Description 356
L3.6 Time-Series Charts 412
13.7.3 Results 359
L3.7 Dynamic Plots 413
13.8 Summary 360
L3.8 Exercises 415
References 360 4 Basic Modeling Concepts 419
L4.1 Multiple Locations, Multiple
PART II Entity Types 419
L4.2 Multiple Parallel Identical
LABS Locations 421
L4.3 Resources 425
1 Introduction to ProModel 365 L4.4 Routing Rules 426
L1.1 ProModel Opening Screen 366 L4.5 Variables 429
L1.2 ProModel Menu Bar 366 L4.6 Uncertainty in Routing—Track
L1.3 ProModel Tool Bars 367 Defects and Rework 432
har01307_fm_i-xviii.indd xi 12/21/10 8:25 AM

xii Contents
L4.7 Batching Multiple Entities of L6.2 Cycle Time 479

Similar Type 434 L6.3 Sorting, Inspecting a Sample, and
L4.7.1 Temporary Rework 480
Batching 434 L6.4 Preventive Maintenance and
L4.7.2 Permanent Machine Breakdowns 483
Batching 436 L6.4.1 Downtime Using
L4.8 Attaching One or More Entities MTBF and MTTR
to Another Entity 438 Data 484
L4.8.1 Permanent L6.4.2 Downtime Using
Attachment 438 MTTF and MTTR
L4.8.2 Temporary Data 484
Attachment 439 L6.5 Shift Working Schedule 488
L4.9 Accumulation of Entities 443 L6.6 Job Shop 491
L4.10 Splitting of One Entity into L6.7 Modeling Priorities 492
Multiple Entities 444 L6.7.1 Selecting among
L4.11 Decision Statements 445 Upstream
L4.11.1 IF-THEN-ELSE Processes 493
Statement 445 L6.8 Modeling a Pull System 494
L4.11.2 WHILE . . . DO L6.8.1 Pull Based on
Loop 447 Downstream
L4.11.3 DO . . . WHILE Demand 495
Loop 449 L6.8.2 Kanban System 497
L4.11.4 DO . . . UNTIL L6.9 Tracking Cost 499
Statement 449 L6.10 Importing a Background 502
L4.11.5 GOTO Statement 451 L6.11 Defining and Displaying
L4.11.6 WAIT UNTIL Views 503
Statement 452 L6.12 Creating a Model Package 506
L4.12 Periodic System Shutdown 453 L6.13 Exercises 508
L4.13 Exercises 456
7 Model Verification and
5 Fitting Statistical Distributions Validation 517
to Input Data 465
L7.1 Verification of an Inspection and
L5.1 An Introduction to Stat::Fit 465 Rework Model 517
L5.2 An Example Problem 467 L7.2 Verification by Tracing the
L5.3 Auto::Fit Input Data 470 Simulation Model 519
L5.4 Exercises 472 L7.3 Debugging the Simulation
Model 520
6 Intermediate Model Building 475 L7.3.1 Debugging ProModel
L6.1 Attributes 475 Logic 522
L6.1.1 Using Attributes to L7.3.2 Basic Debugger
Track Customer Options 523
Types 476 L7.3.3 Advanced Debugger
L6.1.2 Using Attributes and Options 524
Local Variables 478 L7.4 Exercises 525
har01307_fm_i-xviii.indd xii 12/21/10 8:25 AM

Contents xiii
8 Simulation Output Analysis 527 L10.2.2 Multiterm Objective

Functions 583
L8.1 Terminating versus
L10.2.3 Target Range Objective
Nonterminating Simulations 527
Functions 586
L8.2 Terminating Simulation 528
L10.3 Conclusions 588
L8.2.1 Starting and
L10.4 Exercises 590
Terminating Conditions
(Run Length) 529 11 Modeling Manufacturing
L8.2.2 Replications 530 Systems 595
L8.2.3 Required Number of
Replications 534 L11.1 Macros and Run-time
L8.2.4 Simulation Output Interface 595
Assumptions 534 L11.2 Generating Scenarios 599
L8.3 Nonterminating Simulation 536 L11.3 External Files 600
L8.3.1 Warm-up Time and Run L11.4 Arrays 604
Length 538 L11.5 Subroutines 607
L8.3.2 Replications or Batch L11.6 Random Number Streams 611
Intervals 543 L11.7 Merging a Submodel 611
L8.3.3 Required Batch Interval L11.8 Exercises 614
Length 545
L8.4 Exercises 547 12 Material Handling Concepts 617
L12.1 Conveyors 617
9 Comparing Alternative Systems 549 L12.1.1 Single Conveyor 618
L9.1 Overview of Statistical L12.1.2 Multiple
Methods 549 Conveyors 618
L9.2 Three Alternative Systems 550 L12.1.3 Recirculating
L9.3 Common Random Numbers 553 Conveyor 620
L9.4 Bonferroni Approach with L12.2 Resources, Path Networks, and
Paired-t Confidence Intervals 554 Interfaces 621
L9.5 Analysis of Designed L12.2.1 Resource as Material
Experiments 557 Handler 622
L9.5.1 Full Factorial L12.2.2 Resource as Machine
Design Simulation Operator
Experiment 558 and Material
L9.5.2 Fraction Factorial Handler 625
Experiment 561 L12.3 Crane Systems 626
L9.6 Exercises 563 L12.4 Exercises 628
Reference 637
10 Simulation Optimization with
SimRunner 569 13 Modeling Service Systems 639
L10.1 Introduction to SimRunner 569 L13.1 Balking of Customers 639
L10.2 SimRunner Projects 572 L13.2 Table Functions 641
L10.2.1 Single Term Objective L13.3 Arrival Cycles 643
Functions 574 L13.4 User Distribution 646
har01307_fm_i-xviii.indd xiii 12/21/10 8:25 AM

xiv Contents
L13.5 Modeling a University Appendix B Critical Values for Student’s t

Cafeteria 647 Distribution and Standard Normal
L13.6 Modeling a Call Distribution 674
Center—Outsource2US 651
Appendix C F Distribution for ␣ ⫽ 0.05 675
L13.7 Modeling a Triage—Los Angeles
County Hospital 654 Appendix D Critical Values for Chi-Square
L13.8 Modeling an Office (DMV) 657 Distribution 676
L13.9 Exercises 662
Appendix A Continuous and Discrete Index 677
Distributions in ProModel 669
har01307_fm_i-xviii.indd xiv 12/21/10 8:25 AM

F O R E W O R D
Simulation is a modeling and analysis technique used to evaluate and improve

dynamic systems of all types. It has grown from a relatively obscure technology
used by only a few specialists to a widely accepted tool used by decision makers
at all levels in an organization. Imagine being in a highly competitive industry
and managing a manufacturing or service facility that is burdened by outdated
technologies and inefficient management practices. In order to stay competitive,
you know that changes must be made, but you are not exactly sure what changes
would work best, or if certain changes will work at all. You would like to be able
to try out a few different ideas, but you recognize that this would be very time-
consuming, expensive, and disruptive to the current operation. Now, suppose
that there was some magical way you could make a duplicate of your system and
have unlimited freedom to rearrange activities, reallocate resources, or change
any operating procedures. What if you could even try out completely new tech-
nologies and radical new innovations all within just a matter of minutes or hours?
Suppose, further, that all of this experimentation could be done in compressed
time with automatic tracking and reporting of key performance measures. Not
only would you discover ways to improve your operation, but it could all be
achieved risk free—without committing any capital, wasting any time, or dis-
rupting the current system. This is precisely the kind of capability that simulation
provides. Simulation lets you experiment with a computer model of your system
in compressed time, giving you decision-making capability that is unattainable
in any other way.
This text is geared toward simulation courses taught at either a graduate or an
undergraduate level. It contains an ideal blend of theory and practice and covers
the use of simulation in both manufacturing and service systems. This makes it
well suited for use in courses in either an engineering or a business curriculum.
It is also suitable for simulation courses taught in statistics and computer science
programs. The strong focus on the practical aspects of simulation also makes it a
book that any practitioner of simulation would want to have on hand.
This text is designed to be used in conjunction with ProModel simulation
software, which accompanies the book. ProModel is one of the most powerful
and popular simulation packages used today for its ease of use and flexibility.
ProModel was the first fully commercial, Windows-based simulation package and
the first to introduce simulation optimization. ProModel is already being used in
thousands of organizations and taught in hundreds of universities and colleges
throughout the world. While many teaching aids have been developed to train
individuals in the use of ProModel, this is the first full-fledged textbook written
for teaching simulation using ProModel.
Simulation is definitely a learn-by-doing activity. The goal of this text is not
simply to introduce students to the topic of simulation, but to actually develop
competence in the use of simulation. To this end, the book contains plenty of
real-life examples, case studies, and lab exercises to give students actual experi-
ence in the use of simulation. Simulation texts often place too much emphasis on
the theory behind simulation and not enough emphasis on how it is used in actual
har01307_fm_i-xviii.indd xv 12/21/10 8:25 AM

xvi Foreword
problem-solving situations. In simulation courses we have taught over the years,

the strongest feedback we have received from students is that they wish they had
more hands-on time with simulation beginning from the very first week of the
semester.
This text is divided into two parts: a section on the general science and prac-
tice of simulation, and a lab section to train students in the use of ProModel.
While the book is intended for use with ProModel, the division of the book into
two parts permits a modular use of the book, allowing either part to be used inde-
pendently of any other part.
Part I consists of study chapters covering the science and technology of simu-
lation. The first four chapters introduce the topic of simulation, its application to
system design and improvement, and how simulation works. Chapters 5 through
10 present both the practical and theoretical aspects of conducting a simulation
project and applying simulation optimization. Chapters 11 through 13 cover spe-
cific applications of simulation to manufacturing, material handling, and service
systems.
Part II is the lab portion of the book containing exercises for developing sim-
ulation skills using ProModel. The labs are correlated with the reading chapters
in Part I so that Lab 1 should be completed along with Chapter 1 and so on. There
are 13 chapters and 13 labs. The labs are designed to be self-teaching. Students
are walked through the steps of modeling a particular situation and then are given
exercises to complete on their own.
This text focuses on the use of simulation to solve problems in the two
most common types of systems today: manufacturing and service systems.
Nearly 15 percent of the U.S. workforce is employed in manufacturing. In 1955,
about one-half of the U.S. workforce worked in the service sector. Today nearly
80 percent of the American workforce can be found in service-related occupa-
tions. Manufacturing and service systems share much in common. They both
consist of activities, resources, and controls for processing incoming entities.
The performance objectives in both instances relate to quality, efficiency, cost
reduction, process time reduction, and customer satisfaction. In addition to having
common elements and objectives, they are also often interrelated. Manufacturing
systems are supported by service activities such as product design, order manage-
ment, or maintenance. Service systems receive support from production activities
such as food production, check processing, or printing. Regardless of the industry
in which one ends up, an understanding of the modeling issues underlying both
systems will be helpful.
Electronic Textbook Options

Etextbooks are an innovative way for students to save money and create a greener
environment at the same time. An ebook can save students about half the cost of
a traditional textbook and offers unique features like a powerful search engine,
highlighting, and the ability to share notes with classmates using ebooks.
har01307_fm_i-xviii.indd xvi 12/21/10 8:25 AM

Foreword xvii
McGraw-Hill offers this text as an ebook. To talk about the ebook options,
contact your McGraw-Hill sales rep or visit the site www.coursesmart.com to
learn more.
Online Supplements
In addition to the printed text, numerous supplemental materials are available
on the McGraw-Hill Web site at www.mhhe.com/harrell3e. These include case
studies, PowerPoint slides and ProModel software downloads. The case studies
are taken mostly from actual scenarios that can be assigned as simulation proj-
ects. They are intended as capstone experiences to give students an opportunity to
bring together what they have learned in the course to solve a real-life problem.
McGraw-Hill Create™
Craft your teaching resources to match the way you teach! With McGraw-Hill
Create™, www.mcgrawhillcreate.com, you can easily rearrange chapters, com-
bine material from other content sources, and quickly upload content you have
written like your course syllabus or teaching notes. Find the content you need
in Create by searching through thousands of leading McGraw-Hill textbooks.
Arrange your book to fit your teaching style. Create even allows you to personal-
ize your book’s appearance by selecting the cover and adding your name, school,
and course information. Order a Create book and you’ll receive a complimentary
print review copy in 3–5 business days or a complimentary electronic review
copy (eComp) via email in minutes. Go to www.mcgrawhillcreate.com today and
register to experience how McGraw-Hill Create™ empowers you to teach your
students your way.
McGraw-Hill Higher Education and

Blackboard Have Teamed Up
Blackboard, the Web-based course-management system, has partnered with
McGraw-Hill to better allow students and faculty to use online materials and
activities to complement face-to-face teaching. Blackboard features exciting
social learning and teaching tools that foster more logical, visually impactful
and active learning opportunities for students. You’ll transform your closed-door
classrooms into communities where students remain connected to their educa-
tional experience 24 hours a day.
This partnership allows you and your students access to McGraw-Hill’s
Create™ right from within your Blackboard course—all with one single sign-on.
McGraw-Hill and Blackboard can now offer you easy access to industry leading
technology and content, whether your campus hosts it, or we do. Be sure to ask
your local McGraw-Hill representative for details.
har01307_fm_i-xviii.indd xvii 12/21/10 8:25 AM

A C K N O W L E D G M E N T S
No work of this magnitude is performed in a vacuum, independently of the help

and assistance of others. We are indebted to many colleagues, associates, and
other individuals who had a hand in this project. John Mauer (Geer Mountain
Software) provided valuable information on input modeling and the use of
Stat::Fit. Dr. John D. Hall (APT Research, Inc.) and Dr. Allen G. Greenwood
(Mississippi State University) helped to develop and refine the ANOVA material
in Chapter 9. Kerim Tumay provided valuable input on the issues associated with
service system simulation.
We are grateful to all the reviewers of past editions not only for their helpful
feedback, but also for their generous contributions and insights. For their work in
preparation of this third edition, we particularly want to thank: Krishna Krishnan,
Wichita State University; Robert H. Seidman, Southern New Hampshire Univer-
sity; Lee Tichenor, Western Illinois University; Hongyi Chen, University of Min-
nesota, Duluth; Anne Henriksen, James Madison University; Leonid Shnayder,
Stevens Institute of Technology; Bob Kolvoord, James Madison University; Dave
Keranen, University of Minnesota, Duluth; Wade H Shaw, Florida Institute of
Technology; and Marwa Hassan, Louisiana State University.
Many individuals were motivational and even inspirational in taking on this
project: Peter Kalish, Rob Bateman, Lou Keller, Averill Law, Richard Wysk,
Dennis Pegden, and Joyce Kupsh, to name a few. We would especially like to
thank our families for their encouragement and for so generously tolerating the
disruption of normal life caused by this project.
Thanks to all of the students who provided valuable feedback on the first and
second editions of the text. It is for the primary purpose of making simulation
interesting and worthwhile for students that we have written this book.
We are especially indebted to all the wonderful people at PROMODEL Cor-
poration who have been so cooperative in providing software and documentation.
Were it not for the excellent software tools and accommodating support staff at
PROMODEL, this book would not have been written.
Finally, we thank the editorial and production staff at McGraw-Hill: Peter
Massar, Lora Neyens, Melissa Leick, and Brittney Corrigan-McElroy. They have
been great to work with.
har01307_fm_i-xviii.indd xviii 12/21/10 8:25 AM

P A R T
I STUDY CHAPTERS
Chapters
1 Introduction to Simulation 3
2 System Dynamics 33
3 Simulation Basics 57
5 Data Collection and Analysis 115
6 Model Building 161
10 Simulation Optimization 273
12 Modeling Material Handling Systems 323
har01307_ch01_001-032.indd 1 12/17/10 9:30 AM

har01307_ch01_001-032.indd 2 12/17/10 9:30 AM
C H A P T E R
1 INTRODUCTION
TO SIMULATION
“Man is a tool using animal. . . . Without tools he is nothing, with tools he is all.”
—Thomas Carlyle
1.1 Introduction
On March 19, 1999, the following story appeared in The Wall Street Journal:
Captain Chet Rivers knew that his 747-400 was loaded to the limit. The giant plane,
weighing almost 450,000 pounds by itself, was carrying a full load of passengers
and baggage, plus 400,000 pounds of fuel for the long flight from San Francisco
to Australia. As he revved his four engines for takeoff, Capt. Rivers noticed that
San Francisco’s famous fog was creeping in, obscuring the hills to the north and west
of the airport.
At full throttle the plane began to roll ponderously down the runway, slowly at first
but building up to flight speed well within normal limits. Capt. Rivers pulled the throt-
tle back and the airplane took to the air, heading northwest across the San Francisco
peninsula towards the ocean. It looked like the start of another routine flight. Sud-
denly the plane began to shudder violently. Several loud explosions shook the craft
and smoke and flames, easily visible in the midnight sky, illuminated the right wing.
Although the plane was shaking so violently that it was hard to read the instruments,
Capt. Rivers was able to tell that the right inboard engine was malfunctioning, back-
firing violently. He immediately shut down the engine, stopping the explosions and
shaking.
However this introduced a new problem. With two engines on the left wing at
full power and only one on the right, the plane was pushed into a right turn, bringing
it directly towards San Bruno Mountain, located a few miles northwest of the air-
port. Capt. Rivers instinctively turned his control wheel to the left to bring the plane
back on course. That action extended the ailerons—control surfaces on the trailing
edges of the wings—to tilt the plane back to the left. However, it also extended the
spoilers—panels on the tops of the wings—increasing drag and lowering lift. With
har01307_ch01_001-032.indd 3 12/17/10 9:30 AM

4 Part I Study Chapters
the nose still pointed up, the heavy jet began to slow. As the plane neared stall speed,
the control stick began to shake to warn the pilot to bring the nose down to gain air
speed. Capt. Rivers immediately did so, removing that danger, but now San Bruno
Mountain was directly ahead. Capt. Rivers was unable to see the mountain due to
the thick fog that had rolled in, but the plane’s ground proximity sensor sounded
an automatic warning, calling “terrain, terrain, pull up, pull up.” Rivers frantically
pulled back on the stick to clear the peak, but with the spoilers up and the plane still
in a skidding right turn, it was too late. The plane and its full load of 100 tons of fuel
crashed with a sickening explosion into the hillside just above a densely populated
housing area.
“Hey Chet, that could ruin your whole day,” said Capt. Rivers’s supervisor, who
was sitting beside him watching the whole thing. “Let’s rewind the tape and see what
you did wrong.” “Sure Mel,” replied Chet as the two men stood up and stepped out-
side the 747 cockpit simulator. “I think I know my mistake already. I should have used
my rudder, not my wheel, to bring the plane back on course. Say, I need a breather
after that experience. I’m just glad that this wasn’t the real thing.”
The incident above was never reported in the nation’s newspapers, even though
it would have been one of the most tragic disasters in aviation history, because it
never really happened. It took place in a cockpit simulator, a device which uses com-
puter technology to predict and recreate an airplane’s behavior with gut-wrenching
realism.
The relief you undoubtedly felt to discover that this disastrous incident
was just a simulation gives you a sense of the impact that simulation can have
in averting real-world catastrophes. This story illustrates just one of the many
ways simulation is being used to help minimize the risk of making costly and
sometimes fatal mistakes in real life. Simulation technology is finding its way
into an increasing number of applications ranging from training for aircraft pi-
lots to the testing of new product prototypes. The one thing that these applica-
tions have in common is that they all provide a virtual environment that helps
prepare for real-life situations, resulting in significant savings in time, money,
and even lives.
One area where simulation is finding increased application is in manufactur-
ing and service system design and improvement. Its unique ability to accurately
predict the performance of complex systems makes it ideally suited for systems
planning. Just as a flight simulator reduces the risk of making costly errors in
actual flight, system simulation reduces the risk of having systems that operate
inefficiently or that fail to meet minimum performance requirements. While this
may not be life-threatening to an individual, it certainly places a company (not to
mention careers) in jeopardy.
In this chapter we introduce the topic of simulation and answer the following
questions:
• What is simulation?
• Why is simulation used?
• How is simulation performed?
• When and where should simulation be used?
har01307_ch01_001-032.indd 4 12/17/10 9:30 AM

Chapter 1 Introduction to Simulation 5
• What are the qualifications for doing simulation?

• How is simulation economically justified?
The purpose of this chapter is to create an awareness of how simulation is
used to visualize, analyze, and improve the performance of manufacturing and
service systems.
1.2 What Is Simulation?

The Oxford American Dictionary (1980) defines simulation as a way “to repro-
duce the conditions of a situation, as by means of a model, for study or testing or
training, etc.” For our purposes, we are interested in reproducing the operational
behavior of dynamic systems. The model that we will be using is a computer
model. Simulation in this context can be defined as the imitation of a dynamic
system using a computer model in order to evaluate and improve system perfor-
mance. According to Schriber (1987), simulation is “the modeling of a process or
system in such a way that the model mimics the response of the actual system to
events that take place over time.” By studying the behavior of the model, we can
gain insights about the behavior of the actual system.
Simulation is the imitation of a dynamic system

using a computer model in order to evaluate and
improve system performance.
In practice, simulation is usually performed using commercial simulation

software like ProModel that has modeling constructs specifically designed for cap-
turing the dynamic behavior of systems. Performance statistics are gathered dur-
ing the simulation and automatically summarized for analysis. Modern simulation
software provides a realistic, graphical animation of the system being modeled
(Figure 1.1). During the simulation, the user can interactively adjust the animation
speed and change model parameter values to do “what if” analysis on the fly. State-
of-the-art simulation technology even provides optimization capability—not that
simulation itself optimizes, but scenarios that satisfy defined feasibility constraints
can be automatically run and analyzed using special goal-seeking algorithms.
This book focuses primarily on discrete-event simulation, which models the
effects of the events in a system as they occur over time. Discrete-event simula-
tion employs statistical methods for generating random behavior and estimating
har01307_ch01_001-032.indd 5 12/17/10 9:30 AM

FIGURE 1.1
Simulation provides animation capability.
model performance. These methods are sometimes referred to as Monte Carlo

methods because of their similarity to the probabilistic outcomes found in games
of chance, and because Monte Carlo, a tourist resort in Monaco, was such a popu-
lar center for gambling.
1.3 Why Simulate?

Rather than leave design decisions to chance, simulation provides a way to validate
whether or not the best decisions are being made. Simulation avoids the expensive,
time-consuming, and disruptive nature of traditional trial-and-error techniques.
Trial-and-error approaches are

expensive, time consuming, and disruptive.
With the emphasis today on time-based competition, traditional trial-and-error

methods of decision making are no longer adequate. Regarding the shortcoming
har01307_ch01_001-032.indd 6 12/17/10 9:30 AM

of trial-and-error approaches in designing manufacturing systems, Solberg (1988)

notes,
The ability to apply trial-and-error learning to tune the performance of manufacturing
systems becomes almost useless in an environment in which changes occur faster than
the lessons can be learned. There is now a greater need for formal predictive method-
ology based on understanding of cause and effect.
The power of simulation lies in the fact that it provides a method of analysis
that is not only formal and predictive, but is capable of accurately predicting the
performance of even the most complex systems. Deming (1989) states, “Man-
agement of a system is action based on prediction. Rational prediction requires
systematic learning and comparisons of predictions of short-term and long-term
results from possible alternative courses of action.” The key to sound manage-
ment decisions lies in the ability to accurately predict the outcomes of alterna-
tive courses of action. Simulation provides precisely that kind of foresight. By
simulating alternative production schedules, operating policies, staffing levels,
job priorities, decision rules, and the like, a manager can more accurately predict
outcomes and therefore make more informed and effective management deci-
sions. With the importance in today’s competitive market of “getting it right the
first time,” the lesson is becoming clear: if at first you don’t succeed, you probably
should have simulated it.
By using a computer to model a system before it is built or to test operating
policies before they are actually implemented, many of the pitfalls that are often
encountered in the start-up of a new system or the modification of an existing
system can be avoided. Improvements that traditionally took months and even
years of fine-tuning to achieve can be attained in a matter of days or even hours.
Because simulation runs in compressed time, weeks of system operation can be
simulated in only a few minutes or even seconds. The characteristics of simula-
tion that make it such a powerful planning and decision-making tool can be sum-
marized as follows:
• Captures system interdependencies.
• Accounts for variability in the system.
• Is versatile enough to model any system.
• Shows behavior over time.
• Is less costly, time consuming, and disruptive than experimenting on the
actual system.
• Provides information on multiple performance measures.
• Is visually appealing and engages people’s interest.
• Provides results that are easy to understand and communicate.
• Runs in compressed, real, or even delayed time.
• Forces attention to detail in a design.
Because simulation accounts for interdependencies and variation, it provides
insights into the complex dynamics of a system that cannot be obtained using
har01307_ch01_001-032.indd 7 12/17/10 9:30 AM

other analysis techniques. Simulation gives systems planners unlimited freedom

to try out different ideas for improvement, risk free—with virtually no cost, no
waste of time, and no disruption to the current system. Furthermore, the results are
both visual and quantitative with performance statistics automatically reported on
all measures of interest.
Even if no problems are found when analyzing the output of simulation, the
exercise of developing a model is, in itself, beneficial in that it forces one to think
through the operational details of the process. Simulation can work with inaccu-
rate information, but it can’t work with incomplete information. Often solutions
present themselves as the model is built—before any simulation run is made. It
is a human tendency to ignore the operational details of a design or plan until the
implementation phase, when it is too late for decisions to have a significant im-
pact. As the philosopher Alfred North Whitehead observed, “We think in generali-
ties; we live detail” (Audon 1964). System planners often gloss over the details of
how a system will operate and then get tripped up during implementation by all of
the loose ends. The expression “the devil is in the details” has definite application
to systems planning. Simulation forces decisions on critical details so they are not
left to chance or to the last minute, when it may be too late.
Simulation promotes a try-it-and-see attitude that stimulates innovation and
encourages thinking “outside the box.” It helps one get into the system with sticks
and beat the bushes to flush out problems and find solutions. It also puts an end
to fruitless debates over what solution will work best and by how much. Simula-
tion takes the emotion out of the decision-making process by providing objective
evidence that is difficult to refute.
1.4 Doing Simulation

Simulation is nearly always performed as part of a larger process of system
design or process improvement. A design problem presents itself or a need for
improvement exists. Alternative solutions are generated and evaluated, and the
best solution is selected and implemented. Simulation comes into play during
the evaluation phase. First, a model is developed for an alternative solution. As
the model is run, it is put into operation for the period of interest. Performance
statistics (utilization, processing time, and so on) are gathered and reported at the
end of the run. Usually several replications (independent runs) of the simulation
are made. Averages and variances across the replications are calculated to provide
statistical estimates of model performance. Through an iterative process of mod-
eling, simulation, and analysis, alternative configurations and operating policies
can be tested to determine which solution works the best.
Simulation is essentially an experimentation tool in which a computer model
of a new or existing system is created for the purpose of conducting experiments.
The model acts as a surrogate for the actual or real-world system. Knowledge
gained from experimenting on the model can be transferred to the real system.
When we speak of doing simulation, we are talking about “the process of de-
signing a model of a real system and conducting experiments with this model”
har01307_ch01_001-032.indd 8 12/17/10 9:30 AM

(Shannon 1998). Conducting experiments on a model reduces the time, cost, and
disruption of experimenting on the actual system. In this respect, simulation can
be thought of as a virtual prototyping tool for demonstrating proof of concept.
The procedure for doing simulation follows the scientific method of (1) for-
mulating a hypothesis, (2) setting up an experiment, (3) testing the hypothesis
through experimentation, and (4) drawing conclusions about the validity of the
hypothesis. In simulation, we formulate a hypothesis about what design or operat-
ing policies work best. We then set up an experiment in the form of a simulation
model to test the hypothesis. With the model, we conduct multiple replications of
the experiment or simulation. Finally, we analyze the simulation results and draw
conclusions about our hypothesis. If our hypothesis was correct, we can confi-
dently move ahead in making the design or operational changes (assuming time
and other implementation constraints are satisfied). As shown in Figure 1.2, this
process is repeated until we are satisfied with the results.
By now it should be obvious that simulation itself is not a solution tool but
rather an evaluation tool. It describes how a defined system will behave; it does
FIGURE 1.2
The process Start
of simulation
experimentation.
Formulate a
hypothesis
Develop a
simulation model
No
Run simulation
experiment
Hypothesis
correct?
Yes
End
har01307_ch01_001-032.indd 9 12/17/10 9:30 AM

not prescribe how it should be designed. Simulation doesn’t compensate for one’s
ignorance of how a system is supposed to operate. Neither does it excuse one from
being careful and responsible in the handling of input data and the interpretation
of output results. Rather than being perceived as a substitute for thinking, simula-
tion should be viewed as an extension of the mind that enables one to understand
the complex dynamics of a system.
1.5 Use of Simulation

Simulation began to be used in commercial applications in the 1960s. Initial mod-
els were usually programmed in FORTRAN and often consisted of thousands of
lines of code. Not only was model building an arduous task, but extensive debug-
ging was required before models ran correctly. Models frequently took a year or
more to build and debug so that, unfortunately, useful results were not obtained
until after a decision and monetary commitment had already been made. Lengthy
simulations were run in batch mode on expensive mainframe computers where
CPU time was at a premium. Long development cycles prohibited major changes
from being made once a model was built.
Only in the last couple of decades has simulation gained popularity as a
decision-making tool in manufacturing and service industries. Much of the
growth in the use of simulation is due to the increased availability and ease of use
of simulation software that runs on standard PCs. For many companies, simula-
tion has become a standard practice when a new facility is being planned or a
process change is being evaluated. It is fast becoming to systems planners what
spreadsheet software has become to financial planners.
The primary use of simulation continues to be in the area of manufacturing and
logistics, which include warehousing and distribution systems. These areas tend to
have clearly defined relationships and formalized procedures that are well suited
to simulation modeling. They are also the systems that stand to benefit the most
from such an analysis tool since capital investments are so high and changes are so
disruptive. In the service sector, healthcare systems are also a prime candidate for
simulation. Recent trends to standardize and systematize other business processes
such as order processing, invoicing, and customer support are boosting the appli-
cation of simulation in these areas as well. It has been observed that 80 percent of
all business processes are repetitive and can benefit from the same analysis tech-
niques used to improve manufacturing systems (Harrington 1991). With this being
the case, the use of simulation in designing and improving business processes of
every kind will likely continue to grow.
While the primary use of simulation is in decision support, it is by no means
limited to applications requiring a decision. An increasing use of simulation is
in the area of communication and visualization. Modern simulation software
incorporates visual animation that stimulates interest in the model and effectively
communicates complex system dynamics. A proposal for a new system design
can be sold much easier if its operation can actually be demonstrated.
har01307_ch01_001-032.indd 10 12/17/10 9:30 AM

On a smaller scale, simulation is being used to provide interactive, computer-

based training in which a management trainee is given the opportunity to practice
decision-making skills by interacting with the model during the simulation. It is
also being used in real-time control applications where the model interacts with
the real system to monitor progress and provide master control. The power of
simulation to capture system dynamics both visually and functionally opens up
numerous opportunities for its use in an integrated environment.
Since the primary use of simulation is in decision support, most of our discus-
sion will focus on the use of simulation to make system design and operational
decisions. As a decision support tool, simulation has been used to help plan and
make improvements in many areas of both manufacturing and service industries.
Typical applications of simulation include
• Work-flow planning. • Throughput analysis.
• Capacity planning. • Productivity improvement.
• Cycle time reduction. • Layout analysis.
• Staff and resource planning. • Line balancing.
• Work prioritization. • Batch size optimization.
• Bottleneck analysis. • Production scheduling.
• Quality improvement. • Resource scheduling.
• Cost reduction. • Maintenance scheduling.
• Inventory reduction. • Control system design.
1.6 When Simulation Is Appropriate

Not all system problems that could be solved with the aid of simulation should
be solved using simulation. It is important to select the right tool for the task. For
some problems, simulation may be overkill—like using a shotgun to kill a fly.
Simulation has certain limitations of which one should be aware before making a
decision to apply it to a given situation. It is not a panacea for all system-related
problems and should be used only if the shoe fits. As a general guideline, simula-
tion is appropriate if the following criteria hold true:
• An operational (logical or quantitative) decision is being made.
• The process being analyzed is well defined and repetitive.
• Activities and events are interdependent and variable.
• The cost impact of the decision is greater than the cost of doing the
simulation.
• The cost to experiment on the actual system is greater than the cost of
simulation.
Decisions should be of an operational nature. Perhaps the most significant
limitation of simulation is its restriction to the operational issues associated with
systems planning in which a logical or quantitative solution is being sought. It is
har01307_ch01_001-032.indd 11 12/17/10 9:30 AM

not very useful in solving qualitative problems such as those involving technical
or sociological issues. For example, it can’t tell you how to improve machine
reliability or how to motivate workers to do a better job (although it can assess
the impact that a given level of reliability or personal performance can have on
overall system performance). Qualitative issues such as these are better addressed
using other engineering and behavioral science techniques.
Processes should be well defined and repetitive. Simulation is useful only
if the process being modeled is well structured and repetitive. If the process
doesn’t follow a logical sequence and adhere to defined rules, it may be difficult
to model. Simulation applies only if you can describe how the process operates.
This does not mean that there can be no uncertainty in the system. If random
behavior can be described using probability expressions and distributions, they
can be simulated. It is only when it isn’t even possible to make reasonable as-
sumptions of how a system operates (because either no information is available
or behavior is totally erratic) that simulation (or any other analysis tool for that
matter) becomes useless. Likewise, one-time projects or processes that are never
repeated the same way twice are poor candidates for simulation. If the scenario
you are modeling is likely never going to happen again, it is of little benefit to
do a simulation.
Activities and events should be interdependent and variable. A system
may have lots of activities, but if they never interfere with each other or are
deterministic (that is, they have no variation), then using simulation is prob-
ably unnecessary. It isn’t the number of activities that makes a system difficult
to analyze. It is the number of interdependent, random activities. The effect
of simple interdependencies is easy to predict if there is no variability in the
activities. Determining the flow rate for a system consisting of 10 processing
activities is very straightforward if all activity times are constant and activities
are never interrupted. Likewise, random activities that operate independently
of each other are usually easy to analyze. For example, 10 machines operating
in isolation from each other can be expected to produce at a rate that is based
on the average cycle time of each machine less any anticipated downtime.
It is the combination of interdependencies and random behavior that really
produces the unpredictable results. Simpler analytical methods such as math-
ematical calculations and spreadsheet software become less adequate as the
number of activities that are both interdependent and random increases. For
this reason, simulation is primarily suited to systems involving both interde-
pendencies and variability.
The cost impact of the decision should be greater than the cost of doing the
simulation. Sometimes the impact of the decision itself is so insignificant that it
doesn’t warrant the time and effort to conduct a simulation. Suppose, for example,
you are trying to decide whether a worker should repair rejects as they occur or
wait until four or five accumulate before making repairs. If you are certain that
the next downstream activity is relatively insensitive to whether repairs are done
sooner rather than later, the decision becomes inconsequential and simulation is a
wasted effort.
har01307_ch01_001-032.indd 12 12/17/10 9:30 AM

The cost to experiment on the actual system should be greater than the cost
of simulation. While simulation avoids the time delay and cost associated with
experimenting on the real system, in some situations it may actually be quicker
and more economical to experiment on the real system. For example, the decision
in a customer mailing process of whether to seal envelopes before or after they
are addressed can easily be made by simply trying each method and comparing
the results. The rule of thumb here is that if a question can be answered through
direct experimentation quickly, inexpensively, and with minimal impact to the
current operation, then don’t use simulation. Experimenting on the actual system
also eliminates some of the drawbacks associated with simulation, such as prov-
ing model validity.
There may be other situations where simulation is appropriate independent
of the criteria just listed (see Banks and Gibson 1997). This is certainly true in the
case of models built purely for visualization purposes. If you are trying to sell a
system design or simply communicate how a system works, a realistic animation
created using simulation can be very useful, even though nonbeneficial from an
analysis point of view.
1.7 Qualifications for Doing Simulation

Many individuals are reluctant to use simulation because they feel unqualified.
Certainly some training is required to use simulation, but it doesn’t mean that only
statisticians or operations research specialists can learn how to use it. Decision
support tools are always more effective when they involve the decision maker,
especially when the decision maker is also the domain expert or person who is
most familiar with the design and operation of the system. The process owner
or manager, for example, is usually intimately familiar with the intricacies and
idiosyncrasies of the system and is in the best position to know what elements to
include in the model and be able to recommend alternative design solutions. When
performing a simulation, often improvements suggest themselves in the very activ-
ity of building the model that the decision maker might never discover if someone
else is doing the modeling. This reinforces the argument that the decision maker
should be heavily involved in, if not actually conducting, the simulation project.
To make simulation more accessible to non–simulation experts, products
have been developed that can be used at a basic level with very little training.
Unfortunately, there is always a potential danger that a tool will be used in a way
that exceeds one’s skill level. While simulation continues to become more user-
friendly, this does not absolve the user from acquiring the needed skills to make
intelligent use of it. Many aspects of simulation will continue to require some
training. Hoover and Perry (1989) note, “The subtleties and nuances of model
validation and output analysis have not yet been reduced to such a level of rote
that they can be completely embodied in simulation software.”
Modelers should be aware of their own inabilities in dealing with the sta-
tistical issues associated with simulation. Such awareness, however, should not
har01307_ch01_001-032.indd 13 12/17/10 9:30 AM

prevent one from using simulation within the realm of one’s expertise. There
are both a basic as well as an advanced level at which simulation can be benefi-
cially used. Rough-cut modeling to gain fundamental insights, for example, can
be achieved with only a rudimentary understanding of simulation. One need not
have extensive simulation training to go after the low-hanging fruit. Simulation
follows the 80–20 rule, where 80 percent of the benefit can be obtained from
knowing only 20 percent of the science involved (just make sure you know the
right 20 percent). It isn’t until more precise analysis is required that additional
statistical training and knowledge of experimental design are needed.
To reap the greatest benefits from simulation, a certain degree of knowledge
and skill in the following areas is useful:
• Project management.
• Communication.
• Systems engineering.
• Statistical analysis and design of experiments.
• Modeling principles and concepts.
• Basic programming and computer skills.
• Training on one or more simulation products.
• Familiarity with the system being investigated.
Experience has shown that some people learn simulation more rapidly
and become more adept at it than others. People who are good abstract think-
ers yet also pay close attention to detail seem to be the best suited for doing
simulation. Such individuals are able to see the forest while still keeping an
eye on the trees (these are people who tend to be good at putting together
1,000-piece puzzles). They are able to quickly scope a project, gather the
pertinent data, and get a useful model up and running without lots of starts
and stops. A good modeler is somewhat of a sleuth, eager yet methodical and
discriminating in piecing together all of the evidence that will help put the
model pieces together.
If short on time, talent, resources, or interest, the decision maker need not
despair. Plenty of consultants who are professionally trained and experienced can
provide simulation services. A competitive bid will help get the best price, but one
should be sure that the individual assigned to the project has good credentials. If
the use of simulation is only occasional, relying on a consultant may be the pre-
ferred approach.
1.8 Conducting a Simulation Study

Once a suitable application has been selected and appropriate tools and trained
personnel are in place, the simulation study can begin. Simulation is much
more than building and running a model of the process. Successful simula-
tion projects are well planned and coordinated. While there are no strict rules
har01307_ch01_001-032.indd 14 12/17/10 9:30 AM

on how to conduct a simulation project, the following steps are generally

recommended:
Step 1: Define objective and plan the study. Define the purpose of the
simulation project and what the scope of the project will be. A project
plan needs to be developed to determine the resources, time, and budget
requirements for carrying out the project.
Step 2: Collect and analyze system data. Identify, gather, and analyze the
data defining the system to be modeled. This step results in a conceptual
model and a data document on which all can agree.
Step 3: Build the model. Develop a simulation model of the system.
Step 4: Validate the model. Debug the model and make sure it is a credible
representation of the real system.
Step 5: Conduct experiments. Run the simulation for each of the scenarios
to be evaluated and analyze the results.
Step 6: Present the results. Present the findings and make
recommendations so that an informed decision can be made.
Each step need not be completed in its entirety before moving to the next
step. The procedure for doing a simulation is an iterative one in which activities
are refined and sometimes redefined with each iteration. The decision to push
toward further refinement should be dictated by the objectives and constraints of
the study as well as by sensitivity analysis, which determines whether additional
refinement will yield meaningful results. Even after the results are presented,
there are often requests to conduct additional experiments.
Figure 1.3 illustrates this iterative process.
Here we will briefly look at defining the objective and planning the study,
which is the first step in a simulation study. The remaining steps will be discussed
at length in subsequent chapters.
1.8.1 Defining the Objective

The objective of a simulation defines the purpose or reason for conducting the
simulation study. It should be realistic and achievable, given the time and resource
constraints of the study. Simulation objectives can be grouped into the following
general categories:
• Performance analysis—What is the all-around performance of the system
in terms of resource utilization, flow time, output rate, and so on?
• Capacity/constraint analysis—When pushed to the maximum, what is
the processing or production capacity of the system and where are the
bottlenecks?
• Configuration comparison—How well does one system or operational
configuration meet performance objectives compared to another?
• Optimization—What settings for particular decision variables best
achieve desired performance goals?
har01307_ch01_001-032.indd 15 12/17/10 9:30 AM

FIGURE 1.3
Iterative nature Define objective
of simulation. and plan study
Collect and
analyze system
data
Build model
Validate model
Conduct
experiments
Present
results
• Sensitivity analysis—Which decision variables are the most influential on

performance measures, and how influential are they?
• Visualization—How can the system operation be most effectively visualized?
Following is a list of sample design and operational questions that simulation
can help answer. They are intended as examples of specific objectives that might
be defined for a simulation study.
1. How many operating personnel are needed to meet required production
or service levels?
2. What level of automation is the most cost-effective?
3. How many machines, tools, fixtures, or containers are needed to meet
throughput requirements?
har01307_ch01_001-032.indd 16 12/17/10 9:30 AM

4. What is the least-cost method of material handling or transportation

that meets processing requirements?
5. What are the optimum number and size of waiting areas, storage areas,
queues, and buffers?
6. Where are the bottlenecks in the system, and how can they be eliminated?
7. What is the best way to route material, customers, or calls through the
system?
8. What is the best way to allocate personnel to specific tasks?
9. How much raw material and work-in-process inventory should be
maintained?
10. What is the best production control method (kanban, for example)?
When the goal is to analyze some aspect of system performance, the tendency
is to think in terms of the mean or expected value of the performance metric. For
example, we are frequently interested in the average contents of a queue or the
average utilization of a resource. There are other metrics that may have equal or
even greater meaning that can be obtained from a simulation study. For example,
we might be interested in variation as a metric, such as the standard deviation in
waiting times. Extreme values can also be informative, such as the minimum and
maximum number of contents in a storage area. We might also be interested in a
percentile such as the percentage of time that the utilization of a machine is less
than a particular value, say, 80 percent. While frequently we speak of designing
systems to be able to handle peak periods, it often makes more sense to design for
a value above which values only occur less than 5 or 10 percent of the time. It is
more economical, for example, to design a staging area on a shipping dock based
on 90 percent of peak time usage rather than based on the highest usage during
peak time. Sometimes a single measure is not as descriptive as a trend or pattern
of performance. Perhaps a measure has increasing and decreasing periods such as
the activity in a restaurant. In these situations, a detailed time series report would
be the most meaningful.
While well-defined and clearly stated objectives are important to guide the
simulation effort, they should not restrict the simulation or inhibit creativity.
Michael Schrage (1999) observes that “the real value of a model or simulation may
stem less from its ability to test a hypothesis than from its power to generate useful
surprise. Louis Pasteur once remarked that ‘chance favors the prepared mind.’ It
holds equally true that chance favors the prepared prototype: models and simula-
tions can and should be media to create and capture surprise and serendipity. . . .
The challenge is to devise transparent models that also make people shake their
heads and say ‘Wow!’ ” The right “experts” can be “hypervulnerable to surprise
but well situated to turn surprise to their advantage. That’s why Alexander Fleming
recognized the importance of a mold on an agar plate and discovered penicillin.”
Finally, he says, “A prototype should be an invitation to play. You know you have a
successful prototype when people who see it make useful suggestions about how it
can be improved.”
har01307_ch01_001-032.indd 17 12/17/10 9:30 AM

1.8.2 Planning the Study

With a realistic, meaningful, and well-defined objective established, a scope of
work and schedule can be developed for achieving the stated objective. The scope
of work is important for guiding the study as well as providing a specification of
the work to be done upon which all can agree. The scope is essentially a project
specification that helps set expectations by clarifying to others exactly what the
simulation will include and exclude. Such a specification is especially important
if an outside consultant is performing the simulation so that there is mutual under-
standing of the deliverables required.
An important part of the scope is a specification of the models that will be
built. When evaluating improvements to an existing system, it is often desirable
to model the current system first. This is called an “as-is” model. Results from the
as-is model are statistically compared with outputs of the real-world system to
validate the simulation model. This as-is model can then be used as a benchmark
or baseline to compare the results of “to-be” models. For reengineering or process
improvement studies, this two-phase modeling approach is recommended. For
entirely new facilities or processes, there will be no as-is model. There may, how-
ever, be several to-be models to compare.
To ensure that the scope is complete and the schedule is realistic, a determi-
nation should be made of
• The models to be built and experiments to be made.
• The software tools and personnel that will be used to build the models.
• Who will be responsible for gathering the data for building the models.
• How the models will be verified and validated.
• How the results will be presented.
Once these issues have been settled, a project schedule can be developed
showing each of the tasks to be performed and the time to complete each task.
Remember to include sufficient time for documenting the model and adding any
final touches to the animation for presentation purposes. Any additional resources,
activities (travel, etc.) and their associated costs should also be identified for bud-
geting purposes.
1.9 Economic Justification of Simulation

Cost is always an important issue when considering the use of any software tool,
and simulation is no exception. Simulation should not be used if the cost exceeds
the expected benefits. This means that both the costs and the benefits should be
carefully assessed. The use of simulation is often prematurely dismissed due to
the failure to recognize the potential benefits and savings it can produce. Much of
the reluctance in using simulation stems from the mistaken notion that simulation
is costly and very time consuming. This perception is shortsighted and ignores
the fact that in the long run simulation usually saves much more time and cost
har01307_ch01_001-032.indd 18 12/17/10 9:30 AM

than it consumes. It is true that the initial investment, including training and start-
up costs, may be between $10,000 and $30,000 (simulation products themselves
generally range between $1,000 and $20,000). However, this cost is often recov-
ered after the first one or two projects. The ongoing expense of using simulation
for individual projects is estimated to be between 1 and 3 percent of the total
project cost (Glenney and Mackulak 1985). With respect to the time commitment
involved in doing simulation, much of the effort that goes into building the model
is in arriving at a clear definition of how the system operates, which needs to be
done anyway. With the advanced modeling tools that are now available, the actual
model development and running of simulations take only a small fraction (often
less than 5 percent) of the overall system design time.
Savings from simulation are realized by identifying and eliminating problems
and inefficiencies that would have gone unnoticed until system implementation.
Cost is also reduced by eliminating overdesign and removing excessive safety
factors that are added when performance projections are uncertain. By identifying
and eliminating unnecessary capital investments, and discovering and correcting
operating inefficiencies, it is not uncommon for companies to report hundreds of
thousands of dollars in savings on a single project through the use of simulation.
The return on investment (ROI) for simulation often exceeds 1,000 percent, with
payback periods frequently being only a few months or the time it takes to com-
plete a simulation project.
One of the difficulties in developing an economic justification for simulation
is the fact that it is usually not known in advance how much savings will be real-
ized until it is actually used. Most applications in which simulation has been used
have resulted in savings that, had the savings been known in advance, would have
looked very good in an ROI or payback analysis.
One way to assess in advance the economic benefit of simulation is to assess
the risk of making poor design and operational decisions. One need only ask what
the potential cost would be if a misjudgment in systems planning were to occur.
Suppose, for example, that a decision is made to add another machine to solve a
capacity problem in a production or service system. What are the cost and prob-
ability associated with this being the wrong decision? If the cost associated with
a wrong decision is $100,000 and the decision maker is only 70 percent confident
that the decision is correct, then there is a 30 percent chance of incurring a cost of
$100,000. This results in a probable cost of $30,000 (.3 ⫻ $100,000). Using this
approach, many decision makers recognize that they can’t afford not to use simu-
lation because the risk associated with making the wrong decision is too high.
Tying the benefits of simulation to management and organizational goals
also provides justification for its use. For example, a company committed to con-
tinuous improvement or, more specifically, to lead time or cost reduction can be
sold on simulation if it can be shown to be historically effective in these areas.
Simulation has gained the reputation as a best practice for helping companies
achieve organizational goals. Companies that profess to be serious about per-
formance improvement will invest in simulation if they believe it can help them
achieve their goals.
har01307_ch01_001-032.indd 19 12/17/10 9:30 AM

FIGURE 1.4 Concept Design Installation Operation

Cost of making
changes at subsequent
stages of system
development.
Cost
System stage
FIGURE 1.5
Comparison of
Cost without
cumulative system
simulation
costs with and without
System costs
simulation.
Cost with
simulation
Design Implementation Operation

phase phase phase
The real savings from simulation come from allowing designers to make mis-
takes and work out design errors on the model rather than on the actual system.
The concept of reducing costs through working out problems in the design phase
rather than after a system has been implemented is best illustrated by the rule of
tens. This principle states that the cost to correct a problem increases by a factor
of 10 for every design stage through which it passes without being detected (see
Figure 1.4).
Simulation helps avoid many of the downstream costs associated with poor
decisions that are made up front. Figure 1.5 illustrates how the cumulative cost
resulting from systems designed using simulation can compare with the cost of
designing and operating systems without the use of simulation. Note that while
har01307_ch01_001-032.indd 20 12/17/10 9:30 AM

the short-term cost may be slightly higher due to the added labor and software
costs associated with simulation, the long-term costs associated with capital in-
vestments and system operation are considerably lower due to better efficien-
cies realized through simulation. Dismissing the use of simulation on the basis
of sticker price is myopic and shows a lack of understanding of the long-term
savings that come from having well-designed, efficiently operating systems.
Many examples can be cited to show how simulation has been used to avoid
costly errors in the start-up of a new system. Simulation prevented an unnecessary
expenditure when a Fortune 500 company was designing a facility for produc-
ing and storing subassemblies and needed to determine the number of containers
required for holding the subassemblies. It was initially felt that 3,000 containers
were needed until a simulation study showed that throughput did not improve sig-
nificantly when the number of containers was increased from 2,250 to 3,000. By
purchasing 2,250 containers instead of 3,000, a savings of $528,375 was expected
in the first year, with annual savings thereafter of over $200,000 due to the savings
in floor space and storage resulting from having 750 fewer containers (Law and
McComas 1988).
Even if dramatic savings are not realized each time a model is built, simula-
tion at least inspires confidence that a particular system design is capable of meet-
ing required performance objectives and thus minimizes the risk often associated
with new start-ups. The economic benefit associated with instilling confidence
was evidenced when an entrepreneur, who was attempting to secure bank financ-
ing to start a blanket factory, used a simulation model to show the feasibility of
the proposed factory. Based on the processing times and equipment lists supplied
by industry experts, the model showed that the output projections in the business
plan were well within the capability of the proposed facility. Although unfamiliar
with the blanket business, bank officials felt more secure in agreeing to support
the venture (Bateman et al. 1997).
Often simulation can help improve productivity by exposing ways of making
better use of existing assets. By looking at a system holistically, long-standing
problems such as bottlenecks, redundancies, and inefficiencies that previously
went unnoticed start to become more apparent and can be eliminated. “The trick
is to find waste, or muda,” advises Shingo; “after all, the most damaging kind of
waste is the waste we don’t recognize” (Shingo 1992). Consider the following ac-
tual examples where simulation helped uncover and eliminate wasteful practices:
• GE Nuclear Energy was seeking ways to improve productivity without
investing large amounts of capital. Through the use of simulation, the
company was able to increase the output of highly specialized reactor
parts by 80 percent. The cycle time required for production of each part
was reduced by an average of 50 percent. These results were obtained
by running a series of models, each one solving production problems
highlighted by the previous model (Bateman et al. 1997).
• A large manufacturing company with stamping plants located throughout
the world produced stamped aluminum and brass parts on order according
har01307_ch01_001-032.indd 21 12/17/10 9:30 AM

to customer specifications. Each plant had from 20 to 50 stamping presses

that were utilized anywhere from 20 to 85 percent. A simulation study
was conducted to experiment with possible ways of increasing capacity
utilization. As a result of the study, machine utilization improved from an
average of 37 to 60 percent (Hancock, Dissen, and Merten 1977).
• A diagnostic radiology department in a community hospital was modeled
to evaluate patient and staff scheduling, and to assist in expansion
planning over the next five years. Analysis using the simulation model
enabled improvements to be discovered in operating procedures that
precluded the necessity for any major expansions in department size
(Perry and Baum 1976).
In each of these examples, significant productivity improvements were real-
ized without the need for making major investments. The improvements came
through finding ways to operate more efficiently and utilize existing resources
more effectively. These capacity improvement opportunities were brought to light
through the use of simulation.
1.10 Sources of Information on Simulation

Simulation is a rapidly growing technology. While the basic science and theory
remain the same, new and better software is continually being developed to make
simulation more powerful and easier to use. It will require ongoing education for
those using simulation to stay abreast of these new developments. There are many
sources of information to which one can turn to learn the latest developments in
simulation technology. Some of the sources that are available include
• Conferences and workshops sponsored by vendors and professional
societies (such as, Winter Simulation Conference and the IIE Conference).
• Professional magazines and journals (IIE Solutions, International Journal
of Modeling and Simulation, and the like).
• Websites of vendors and professional societies (www.promodel.com,
www.scs.org, and so on).
• Demonstrations and tutorials provided by vendors.
• Textbooks (like this one).
1.11 How to Use This Book

This book is divided into two parts. Part I contains chapters describing the science
and practice of simulation. The emphasis is deliberately oriented more toward the
practice than the science. Simulation is a powerful decision support tool that has
a broad range of applications. While a fundamental understanding of how simula-
tion works is presented, the aim has been to focus more on how to use simulation
har01307_ch01_001-032.indd 22 12/17/10 9:30 AM

to solve real-world problems. Review questions at the end of each chapter help
reinforce the concepts presented.
Part II contains ProModel lab exercises that help develop simulation skills.
ProModel is a simulation package designed specifically for ease of use, yet it
provides the flexibility to model any discrete event or continuous flow process.
It is similar to other simulation products in that it provides a set of basic model-
ing constructs and a language for defining the logical decisions that are made in
a system. Basic modeling objects in ProModel include entities (the objects being
processed), locations (the places where processing occurs), resources (the agents
used to process the entities), and paths (the course of travel for entities and re-
sources in moving between locations such as aisles or conveyors). Logical behav-
ior such as the way entities arrive and their routings can be defined with little, if
any, programming using the data entry tables that are provided. ProModel is used
by thousands of professionals in manufacturing and service-related industries and
is taught in hundreds of institutions of higher learning.
It is recommended that students be assigned at least one simulation project
during the course. Preferably this is a project performed for a nearby company
or institution so it will be meaningful. If such a project cannot be found, or as
an additional practice exercise. An assortment of case problems are posted on
the McGraw-Hill Website (www.mhhe.com/harrell3e) that can be assigned. Stu-
dent projects should be selected early in the course so that data gathering can get
started and the project completed within the allotted time. The chapters in Part I
are sequenced to parallel an actual simulation project.
1.12 Summary
Businesses today face the challenge of quickly designing and implementing complex
production and service systems that are capable of meeting growing demands for
quality, delivery, affordability, and service. With recent advances in computing and
software technology, simulation tools are now available to help meet this challenge.
Simulation is a powerful technology that is being used with increasing frequency to
improve system performance by providing a way to make better design and manage-
ment decisions. When used properly, simulation can reduce the risks associated with
starting up a new operation or making improvements to existing operations.
Because simulation accounts for interdependencies and variability, it pro-
vides insights that cannot be obtained any other way. Where important system
decisions are being made of an operational nature, simulation is an invaluable
decision-making tool. Its usefulness increases as variability and interdependency
increase and the importance of the decision becomes greater.
Lastly, simulation actually makes designing systems fun! Not only can a
designer try out new design concepts to see what works best, but the visualiza-
tion makes it take on a realism that is like watching an actual system in operation.
Through simulation, decision makers can play what-if games with a new system
or modified process before it actually gets implemented. This engaging process
stimulates creative thinking and leads to good design decisions.
har01307_ch01_001-032.indd 23 12/17/10 9:30 AM

1.13 Review Questions

1. Define simulation.
2. What reasons are there for the increased popularity of computer
simulation?
3. What are two specific questions that simulation might help answer in a
bank? In a manufacturing facility? In a dental office?
4. What are three advantages that simulation has over alternative
approaches to systems design?
5. Does simulation itself optimize a system design? Explain.
6. How does simulation follow the scientific method?
7. A restaurant gets extremely busy during lunch (11:00 A.M. to 2:00 P.M.)
and is trying to decide whether it should increase the number of
waitresses from two to three. What considerations would you look at to
determine whether simulation should be used to make this decision?
8. How would you develop an economic justification for using
simulation?
9. Is a simulation exercise wasted if it exposes no problems in a system
design? Explain.
10. A simulation run was made showing that a modeled factory could
produce 130 parts per hour. What information would you want to know
about the simulation study before placing any confidence in the results?
11. A PC board manufacturer has high work-in-process (WIP) inventories,
yet machines and equipment seem underutilized. How could simulation
help solve this problem?
12. How important is a statistical background for doing simulation?
13. How can a programming background be useful in doing simulation?
14. Why are good project management and communication skills important
in simulation?
15. Why should the process owner be heavily involved in a simulation
project?
16. For which of the following problems would simulation likely be
useful?
a. Increasing the throughput of a production line.
b. Increasing the pace of a worker on an assembly line.
c. Decreasing the time that patrons at an amusement park spend
waiting in line.
d. Determining the percentage defective from a particular machine.
e. Determining where to place inspection points in a process.
f. Finding the most efficient way to fill out an order form.
17. Why is it important to have clearly defined objectives for a simulation
that everyone understands and can agree on?
har01307_ch01_001-032.indd 24 12/17/10 9:30 AM

18. For each of the following systems, define one possible simulation
objective:
a. A manufacturing cell.
b. A fast-food restaurant.
c. An emergency room of a hospital.
d. A tire distribution center.
e. A public bus transportation system.
f. A post office.
g. A bank of elevators in a high-rise office building.
h. A time-shared, multitasking computer system consisting of a server,
many slave terminals, a high-speed disk, and a processor.
i. A rental car agency at a major airport.
1.14 Case Studies
CASE STUDY A
AST COMPUTES BIG BENEFITS USING SIMULATION
John Perry
Manufacturing Engineer
Source: AST Research Inc. Used with permission by ProModel. 1999.
In the world of computer manufacturing and assembly, success is measured in terms of sec-
onds. The PC business poses some unique challenges: it is intensely competitive with thin
profit margins and is characterized by constantly changing technologies and products. The
race is to excel in rapid delivery to the marketplace through high-velocity manufacturing.
A manufacturer that can minimize the time needed to assemble a computer, move nimbly
to respond to change, and get maximum productivity out of its capital investment has an
edge. At AST Research, we use computer simulation to realize these goals and, so far, have
achieved significant benefits. The application area, which is product assembly, is an area
where almost anyone can benefit from the use of simulation.
har01307_ch01_001-032.indd 25 12/17/10 9:30 AM

The Problem
AST Research Inc., founded in 1980, has become a multibillion-dollar PC manufacturer.
We assemble personal computers and servers in Fort Worth, Texas, and offshore. For a
long time, we “optimized” (planned) our assembly procedures using traditional methods—
gathering time-and-motion data via stopwatches and videotape, performing simple arith-
metic calculations to obtain information about the operation and performance of the
assembly line, and using seat-of-the-pants guesstimates to “optimize” assembly line output
and labor utilization.
The Model
In December 1994 a new vice president joined AST. Management had long been commit-
ted to increasing the plant’s efficiency and output, and our new vice president had experi-
ence in using simulation to improve production. We began using ProModel simulation
as a tool for optimizing our assembly lines, and to improve our confidence that changes
proposed in the assembly process would work out in practice. The results have been sig-
nificant. They include
• Reduced labor input to each unit.
• Shortened production cycle times.
• Reduced reject rate.
• Increased ability to change assembly instructions quickly when changes become
necessary.
Now when we implement changes, we are confident that those changes in fact will improve
things.
The first thing we did was learn how to use the simulation software. Then we attempted
to construct a model of our current operations. This would serve two important functions: it
would tell us whether we really understood how to use the software, and it would validate
our understanding of our own assembly lines. If we could not construct a model of our
assembly line that agreed with the real one, that would mean that there was a major flaw
either in our ability to model the system or in our understanding of our own operations.
Building a model of our own operations sounded simple enough. After all, we had an
exhaustive catalog of all the steps needed to assemble a computer, and (from previous data
collection efforts) information on how long each operation took. But building a model
that agreed with our real-world situation turned out to be a challenging yet tremendously
educational activity.
For example, one of our early models showed that we were producing several thousand
units in a few hours. Since we were not quite that good—we were off by at least a factor
of 10—we concluded that we had a major problem in the model we had built, so we went
back to study things in more detail. In every case, our early models failed because we had
overlooked or misunderstood how things actually worked on our assembly lines.
Eventually, we built a model that worked and agreed reasonably well with our real-
world system out in the factory. To make use of the model, we generated ideas that we
thought would cut down our assembly time and then simulated them in our model.
We examined a number of changes proposed by our engineers and others, and then
simulated the ones that looked most promising. Some proposed changes were counter-
productive according to our simulation results. We also did a detailed investigation of our
testing stations to determine whether it was more efficient to move computers to be tested
har01307_ch01_001-032.indd 26 12/17/10 9:30 AM

into the testing station via FIFO (first-in, first-out) or LIFO (last-in, first-out). Modeling
showed us that FIFO was more efficient. When we implemented that change, we realized
the gain we had predicted.
Simulation helped us avoid buying more expensive equipment. Some of our material-
handling specialists predicted, based on their experience, that if we increased throughput
by 30 percent, we would have to add some additional, special equipment to the assembly
floor or risk some serious blockages. Simulation showed us that was not true and in prac-
tice the simulation turned out to be correct.
We determined that we could move material faster if we gave material movers a specific
pattern to follow instead of just doing things sequentially. For example, in moving certain
items from our testing area, we determined that the most time-efficient way would be to
move shelf 1 first, followed by shelf 4, then shelf 3, and so on.
After our first round of making “serious” changes to our operation and simulating them,
our actual production was within a few percentage points of our predicted production.
Also, by combining some tasks, we were able to reduce our head count on each assembly
line significantly.
We have completed several rounds of changes, and today, encouraged by the experience
of our new investor, Samsung, we have made a significant advance that we call Vision 5.
The idea of Vision 5 is to have only five people in each cell assembling computers. Although
there was initially some skepticism about whether this concept would work, our simulations
showed that it would, so today we have converted one of our “focused factories” to this
concept and have experienced additional benefits. Seeing the benefits from that effort has
caused our management to increase its commitment to simulation.
The Results
Simulation has proven its effectiveness at AST Research. We have achieved a number of
useful, measurable goals. For competitive reasons, specific numbers cannot be provided;
however, in order of importance, the benefits we have achieved are
• Reduced the reject rate.
• Reduced blockage by 25 percent.
• Increased operator efficiency by 20 percent.
• Increased overall output by 15 percent.
• Reduced the labor cost of each computer.
Other benefits included increased ability to explain and justify proposed changes to
management through the use of the graphic animation. Simulation helped us make fewer
missteps in terms of implementing changes that could have impaired our output. We were
able to try multiple scenarios in our efforts to improve productivity and efficiency at com-
paratively low cost and risk. We also learned that the best simulation efforts invite partici-
pation by more disciplines in the factory, which helps in terms of team-building. All of
these benefits were accomplished at minimal cost. These gains have also caused a cultural
shift at AST, and because we have a tool that facilitates production changes, the company
is now committed to continuous improvement of our assembly practices.
Our use of simulation has convinced us that it produces real, measurable results—and
equally important, it has helped us avoid making changes that we thought made common
sense, but when simulated turned out to be ineffective. Because of that demonstrable pay-
off, simulation has become a key element of our toolkit in optimizing production.
har01307_ch01_001-032.indd 27 12/17/10 9:30 AM

Questions
1. What were the objectives for using simulation at AST?
2. Why was simulation better than the traditional methods they were using to achieve
these objectives?
3. What common-sense solution was disproved by using simulation?
4. What were some of the unexpected side benefits from using simulation?
5. What insights on the use of simulation did you gain from this case study?
CASE STUDY B
DURHAM REGIONAL HOSPITAL SAVES $150,000
ANNUALLY USING SIMULATION TOOLS
Bonnie Lowder
Management Engineer, Premier
Used with permission by ProModel. 1999.
Durham Regional Hospital, a 450-bed facility located in Durham, North Carolina, has been
serving Durham County for 25 years. This public, full-service, acute-care facility is facing
the same competition that is now a part of the entire health care industry. With that in mind,
Durham Regional Hospital is making a conscious effort to provide the highest quality of
care while also controlling costs.
To assist with cost control efforts, Durham Regional Hospital uses Premier’s Customer-
Based Management Engineering program. Premier’s management engineers are very
har01307_ch01_001-032.indd 28 12/17/10 9:30 AM

involved with the hospital’s reengineering and work redesign projects. Simulation is one
of the tools the management engineers use to assist in the redesign of hospital services and
processes. Since the hospital was preparing to add an outpatient services area that was to
open in May 1997, a MedModel simulation project was requested by Durham Regional
Hospital to see how this Express Services area would impact their other outpatient areas.
This project involved the addition of an outpatient Express Services addition. The
Express Services area is made up of two radiology rooms, four phlebotomy lab stations, a
patient interview room, and an EKG room. The model was set up to examine which kind of
patients would best be serviced in that area, what hours the clinic would operate, and what
staffing levels would be necessary to provide optimum care.
The Model
Data were collected from each department with potential Express Services patients.
The new Express Services area would eliminate the current reception desk in the main
radiology department; all radiology outpatients would have their order entry at Express
Services. In fiscal year 1996, the radiology department registered 21,159 outpatients. Of
those, one-third could have had their procedure performed in Express Services. An aver-
age of 18 outpatient surgery patients are seen each week for their preadmission testing. All
these patients could have their preadmission tests performed in the Express Services area.
The laboratory sees approximately 14 walk-in phlebotomy patients per week. Of those,
10 patients are simple collections and 4 are considered complex. The simple collections
can be performed by anyone trained in phlebotomy. The complex collections should be
done by skilled lab personnel. The collections for all of these patients could be performed
in Express Services. Based on the data, 25 patients a day from the Convenient Care Clinic
will need simple X rays and could also use the Express Services area. Procedure times for
each patient were determined from previous data collection and observation.
The model was built in two months. Durham Regional Hospital had used simulation in
the past for both Emergency Department and Ambulatory Care Unit redesign projects and
thus the management team was convinced of its efficacy. After the model was completed,
it was presented to department managers from all affected areas. The model was presented
to the assembled group in order to validate the data and assumptions. To test for validity,
the model was run for 30 replications, with each replication lasting a period of two weeks.
The results were measured against known values.
The Results
The model showed that routing all Convenient Care Clinic patients through Express Ser-
vices would create a bottleneck in the imaging rooms. This would create unacceptable
wait times for the radiology patients and the Convenient Care patients. Creating a model
scenario where Convenient Care patients were accepted only after 5:00 P.M. showed that
the anticipated problem could be eliminated. The model also showed that the weekend
volume would be very low. Even at minimum staffing levels, the radiology technicians
and clerks would be underutilized. The recommendation was made to close the Express
Services area on the weekends. Finally, the model showed that the staffing levels could be
lower than had been planned. For example, the workload for the outpatient lab tech drops
off after 6:00 P.M. The recommendation was to eliminate outpatient lab techs after 6:00 P.M.
Further savings could be achieved by cross-training the radiology technicians and pos-
sibly the clerks to perform simple phlebotomy. This would also provide for backup during
har01307_ch01_001-032.indd 29 12/17/10 9:30 AM

busy times. The savings for the simulation efforts were projected to be $148,762 annu-
ally. These savings were identified from the difference in staffing levels initially requested
for Express Services and the levels that were validated after the simulation model results
were analyzed, as well as the closing of the clinic on weekends. This model would also be
used in the future to test possible changes to Express Services. Durham Regional Hospital
would be able to make minor adjustments to the area and visualize the outcome before
implementation. Since this was a new area, they would also be able to test minor changes
before the area was opened.
The results of the model allowed the hospital to avoid potential bottlenecks in the
radiology department, reduce the proposed staffing levels in the Express Services area, and
validate that the clinic should be closed on weekends. As stated by Dottie Hughes, director
of Radiology Services: “The simulation model allowed us to see what changes needed to
be made before the area is opened. By making these changes now, we anticipate a shorter
wait time for our patients than if the simulation had not been used.” The simulation results
were able to show that an annual savings of $148,762 could be expected by altering some
of the preconceived Express Services plan.
Future Applications
Larry Suitt, senior vice president, explains, “Simulation has proved to be a valuable tool for
our hospital. It has allowed us to evaluate changes in processes before money is spent on
construction. It has also helped us to redesign existing services to better meet the needs of
our patients.” Durham Regional Hospital will continue to use simulation in new projects to
improve its health care processes. The hospital is responsible for the ambulance service for
the entire county. After the 911 call is received, the hospital’s ambulance service picks up
the patient and takes him or her to the nearest hospital. Durham Regional Hospital is plan-
ning to use simulation to evaluate how relocating some of the ambulances to other stations
will affect the response time to the 911 calls.
Questions
1. Why was this a good application for simulation?
2. What key elements of the study made the project successful?
3. What specific decisions were made as a result of the simulation study?
4. What economic benefit was able to be shown from the project?
5. What insights did you gain from this case study about the way simulation is used?
References
Audon, Wyston Hugh, and L. Kronenberger. The Faber Book of Aphorisms. London: Faber
and Faber, 1964.
Banks, Jerry, and R. Gibson. “10 Rules for Determining When Simulation Is Not Appropri-
ate.” IIE Solutions, September 1997, pp. 30–32.
Banks, Jerry, and R. Gibson. “Selecting Simulation Software.” IIE Solutions, May 1997,
pp. 29–32.
Bateman, Robert E.; Royce O. Bowden; Thomas J. Gogg; Charles R. Harrell; and Jack
R. A. Mott. System Improvement Using Simulation. Utah: PROMODEL Corp., 1997.
har01307_ch01_001-032.indd 30 12/17/10 9:30 AM

Deming, W. E. Foundation for Management of Quality in the Western World. Paper read
at a meeting of the Institute of Management Sciences, Osaka, Japan, 24 July 1989.
Glenney, Neil E., and Gerald T. Mackulak. “Modeling & Simulation Provide Key to CIM
Implementation Philosophy.” Industrial Engineering, May 1985.
Hancock, Walton; R. Dissen; and A. Merten. “An Example of Simulation to Improve Plant
Productivity.” AIIE Transactions, March 1977, pp. 2–10.
Harrell, Charles R., and Donald Hicks. “Simulation Software Component Architec-
ture for Simulation-Based Enterprise Applications.” In Proceedings of the 1998
Winter Simulation Conference, ed. D. J. Medeiros, E. F. Watson, J. S. Carson, and
M. S. Manivannan, pp. 1717–21. Institute of Electrical and Electronics Engineers,
Piscataway, New Jersey.
Harrington, H. James. Business Process Improvement. New York: McGraw-Hill, 1991.
Hoover, Stewart V., and Ronald F. Perry. Simulation: A Problem-Solving Approach.
Reading, MA: Addison-Wesley, 1989.
Kelton, W. D. “Statistical Issues in Simulation.” In Proceedings of the 1996 Winter
Simulation Conference, ed. J. Charnes, D. Morrice, D. Brunner, and J. Swain, 1996,
pp. 47–54.
Law, A. M., and M. G. McComas. “How Simulation Pays Off.” Manufacturing Engineer-
ing, February 1988, pp. 37–39.
Mott, Jack, and Kerim Tumay. “Developing a Strategy for Justifying Simulation.” Indus-
trial Engineering, July 1992, pp. 38–42.
Oxford American Dictionary. New York: Oxford University Press, 1980. [compiled by]
Eugene Enrich et al.
Perry, R. F., and R. F. Baum. “Resource Allocation and Scheduling for a Radiology
Department.” In Cost Control in Hospitals. Ann Arbor, MI: Health Administration
Press, 1976.
Rohrer, Matt, and Jerry Banks. “Required Skills of a Simulation Analyst.” IIE Solutions,
May 1998, pp. 7–23.
Schrage, Michael. Serious Play: How the World’s Best Companies Simulate to Innovate.
Cambridge, MA: Harvard Business School Press, 1999.
Schriber, T. J. “The Nature and Role of Simulation in the Design of Manufacturing
Systems.” Simulation in CIM and Artificial Intelligence Techniques, ed. J. Retti and
K. E. Wichmann. S.D., CA.: Society for Computer Simulation, 1987, pp. 5–8.
Shannon, Robert E. “Introduction to the Art and Science of Simulation.” In Proceed-
ings of the 1998 Winter Simulation Conference, ed. D. J. Medeiros, E. F. Watson,
J. S. Carson, and M. S. Manivannan, pp. 7–14. Piscataway, NJ: Institute of Electrical
and Electronics Engineers.
Shingo, Shigeo. The Shingo Production Management System—Improving Process Func-
tions. Trans. Andrew P. Dillon. Cambridge, MA: Productivity Press, 1992.
Solberg, James. “Design and Analysis of Integrated Manufacturing Systems.” In W. Dale
Compton. Washington, D.C.: National Academy Press, 1988, p. 4.
The Wall Street Journal, March 19, 1999. “United 747’s Near Miss Sparks a Widespread
Review of Pilot Skills,” p. A1.
har01307_ch01_001-032.indd 31 12/17/10 9:30 AM

har01307_ch01_001-032.indd 32 12/17/10 9:30 AM
C H A P T E R
2 SYSTEM DYNAMICS
“A fool with a tool is still a fool.”

—Unknown
2.1 Introduction
Knowing how to do simulation doesn’t make someone a good systems designer
any more than knowing how to use a CAD system makes one a good product
designer. Simulation is a tool that is useful only if one understands the nature of
the problem to be solved. It is designed to help solve systemic problems that are
operational in nature. Simulation exercises fail to produce useful results more
often because of a lack of understanding of system dynamics than a lack of know-
ing how to use the simulation software. The challenge is in understanding how
the system operates, knowing what you want to achieve with the system, and
being able to identify key leverage points for best achieving desired objectives.
To illustrate the nature of this challenge, consider the following actual scenario:
The pipe mill for the XYZ Steel Corporation was an important profit center, turning
steel slabs selling for under $200/ton into a product with virtually unlimited demand
selling for well over $450/ton. The mill took coils of steel of the proper thickness
and width through a series of machines that trimmed the edges, bent the steel into a
cylinder, welded the seam, and cut the resulting pipe into appropriate lengths, all on
a continuously running line. The line was even designed to weld the end of one coil
to the beginning of the next one “on the fly,” allowing the line to run continually for
days on end.
Unfortunately the mill was able to run only about 50 percent of its theoretical
capacity over the long term, costing the company tens of millions of dollars a year in
lost revenue. In an effort to improve the mill’s productivity, management studied each
step in the process. It was fairly easy to find the slowest step in the line, but additional
study showed that only a small percentage of lost production was due to problems at
33
har01307_ch02_033-056.indd 33 12/17/10 10:13 AM

this “bottleneck” operation. Sometimes a step upstream from the bottleneck would
have a problem, causing the bottleneck to run out of work, or a downstream step
would go down temporarily, causing work to back up and stop the bottleneck. Some-
times the bottleneck would get so far behind that there was no place to put incoming,
newly made pipe. In this case the workers would stop the entire pipe-making process
until the bottleneck was able to catch up. Often the bottleneck would then be idle
waiting until the newly started line was functioning properly again and the new pipe
had a chance to reach it. Sometimes problems at the bottleneck were actually caused
by improper work at a previous location.
In short, there was no single cause for the poor productivity seen at this plant.
Rather, several separate causes all contributed to the problem in complex ways. Man-
agement was at a loss to know which of several possible improvements (additional or
faster capacity at the bottleneck operation, additional storage space between stations,
better rules for when to shut down and start up the pipe-forming section of the mill,
better quality control, or better training at certain critical locations) would have the
most impact for the least cost. Yet the poor performance of the mill was costing
enormous amounts of money. Management was under pressure to do something, but
what should it be?
This example illustrates the nature and difficulty of the decisions that an op-
erations manager faces. Managers need to make decisions that are the “best” in
some sense. To do so, however, requires that they have clearly defined goals and
understand the system well enough to identify cause-and-effect relationships.
While every system is different, just as every product design is different,
the basic elements and types of relationships are the same. Knowing how the
elements of a system interact and how overall performance can be improved are
essential to the effective use of simulation. This chapter reviews basic system
dynamics and answers the following questions:
• What is a system?
• What are the elements of a system?
• What makes systems so complex?
• What are useful system metrics?
• What is a systems approach to systems planning?
• How do traditional systems analysis techniques compare with simulation?
2.2 System Definition

We live in a society that is composed of complex, human-made systems that we de-
pend on for our safety, convenience, and livelihood. Routinely we rely on transpor-
tation, health care, production, and distribution systems to provide needed goods
and services. Furthermore, we place high demands on the quality, convenience,
timeliness, and cost of the goods and services that are provided by these systems.
Remember the last time you were caught in a traffic jam, or waited for what seemed
like an eternity in a restaurant or doctor’s office? Contrast that experience with
the satisfaction that comes when you find a store that sells quality merchandise at
har01307_ch02_033-056.indd 34 12/17/10 10:13 AM

Chapter 2 System Dynamics 35
discount prices or when you locate a health care organization that provides prompt
and professional service. The difference is between a system that has been well
designed and operates smoothly, and one that is poorly planned and managed.
A system, as used here, is defined as a collection of elements that function
together to achieve a desired goal (Blanchard 1991). Key points in this definition
include the fact that (1) a system consists of multiple elements, (2) these elements
are interrelated and work in cooperation, and (3) a system exists for the purpose
of achieving specific objectives. Examples of systems are traffic systems, politi-
cal systems, economic systems, manufacturing systems, and service systems. Our
main focus will be on manufacturing and service systems that process materials,
information, and people.
Manufacturing systems can be small job shops and machining cells or large
production facilities and assembly lines. Warehousing and distribution as well as
entire supply chain systems will be included in our discussions of manufacturing
systems. Service systems cover a wide variety of systems including health care fa-
cilities, call centers, amusement parks, public transportation systems, restaurants,
banks, and so forth.
Both manufacturing and service systems may be termed processing systems be-
cause they process items through a series of activities. In a manufacturing system,
raw materials are transformed into finished products. For example, a bicycle manu-
facturer starts with tube stock that is then cut, welded, and painted to produce bicycle
frames. In service systems, customers enter with some service need and depart as
serviced (and, we hope, satisfied) customers. In a hospital emergency room, for ex-
ample, nurses, doctors, and other staff personnel admit and treat incoming patients
who may undergo tests and possibly even surgical procedures before finally being re-
leased. Processing systems are artificial (they are human-made), dynamic (elements
interact over time), and usually stochastic (they exhibit random behavior).
2.3 System Elements

From a simulation perspective, a system can be said to consist of entities, ac-
tivities, resources, and controls (see Figure 2.1). These elements define the who,
what, where, when, and how of entity processing. This model for describing a
FIGURE 2.1
Incoming entities Outgoing entities
Elements of a system. Activities
Resources Controls
System
har01307_ch02_033-056.indd 35 12/17/10 10:13 AM

system corresponds closely to the well-established ICAM definition (IDEF) pro-

cess model developed by the defense industry (ICAM stands for an early Air
Force program in integrated computer-aided manufacturing). The IDEF modeling
paradigm views a system as consisting of inputs and outputs (that is, entities),
activities, mechanisms (that is, resources), and controls.
2.3.1 Entities
Entities are the items processed through the system such as products, customers,
and documents. Different entities may have unique characteristics such as cost,
shape, priority, quality, or condition. Entities may be further subdivided into the
following types:
• Human or animate (customers, patients, etc.).
• Inanimate (parts, documents, bins, etc.).
• Intangible (calls, electronic mail, etc.).
For most manufacturing and service systems, the entities are discrete items. This
is the case for discrete part manufacturing and is certainly the case for nearly all
service systems that process customers, documents, and others. For some production
systems, called continuous systems, a nondiscrete substance is processed rather than
discrete entities. Examples of continuous systems are oil refineries and paper mills.
2.3.2 Activities
Activities are the tasks performed in the system that are either directly or indi-
rectly involved in the processing of entities. Examples of activities include ser-
vicing a customer, cutting a part on a machine, or repairing a piece of equipment.
Activities usually consume time and often involve the use of resources. Activities
may be classified as
• Entity processing (check-in, treatment, inspection, fabrication, etc.).
• Entity and resource movement (forklift travel, riding in an elevator, etc.).
• Resource adjustments, maintenance, and repairs (machine setups, copy
machine repair, etc.).
2.3.3 Resources
Resources are the means by which activities are performed. They provide the support-
ing facilities, equipment, and personnel for carrying out activities. While resources
facilitate entity processing, inadequate resources can constrain processing by limiting
the rate at which processing can take place. Resources have characteristics such as ca-
pacity, speed, cycle time, and reliability. Like entities, resources can be categorized as
• Human or animate (operators, doctors, maintenance personnel, etc.).
• Inanimate (equipment, tooling, floor space, etc.).
• Intangible (information, electrical power, etc.).
har01307_ch02_033-056.indd 36 12/17/10 10:13 AM

Resources can also be classified as being dedicated or shared, permanent or con-

sumable, and mobile or stationary.
2.3.4 Controls
Controls dictate how, when, and where activities are performed. Controls impose
order on the system. At the highest level, controls consist of schedules, plans,
and policies. At the lowest level, controls take the form of written procedures and
machine control logic. At all levels, controls provide the information and decision
logic for how the system should operate. Examples of controls include
• Routing sequences.
• Production plans.
• Work schedules.
• Task prioritization.
• Control software.
• Instruction sheets.
2.4 System Complexity

Elements of a system operate in concert with one another in ways that often re-
sult in complex interactions. The word complex comes from the Latin complexus,
meaning entwined or connected together. Unfortunately, unaided human intuition
is not very good at analyzing and understanding complex systems. Economist
Herbert Simon called this inability of the human mind to grasp real-world com-
plexity “the principle of bounded rationality.” This principle states that “the ca-
pacity of the human mind for formulating and solving complex problems is very
small compared with the size of the problem whose solution is required for objec-
tively rational behavior in the real world, or even for a reasonable approximation
to such objective rationality” (Simon 1957).
Bounded rationality—our limited ability to grasp

real-world complexity.
While the sheer number of elements in a system can stagger the mind
(the number of different entities, activities, resources, and controls can easily
har01307_ch02_033-056.indd 37 12/17/10 10:13 AM

exceed 100), the interactions of these elements are what make systems so com-
plex and difficult to analyze. System complexity is primarily a function of the
following two factors:
1. Interdependencies between elements so that each element affects other
elements.
2. Variability in element behavior that produces uncertainty.
Interdependencies Variability Complexity
These two factors characterize virtually all human-made systems and make
system behavior difficult to analyze and predict. As shown in Figure 2.2, the de-
gree of analytical difficulty increases exponentially as the number of interdepen-
dencies and random variables increases.
2.4.1 Interdependencies
Interdependencies cause the behavior of one element to affect other elements in
the system. For example, if a machine breaks down, repair personnel are put into
action while downstream operations become idle for lack of parts. Upstream op-
erations may even be forced to shut down due to a logjam in the entity flow caus-
ing a blockage of activities. Another place where this chain reaction or domino
effect manifests itself is in situations where resources are shared between two
or more activities. A doctor treating one patient, for example, may be unable to
FIGURE 2.2
Analytical difficulty
Degree of analytical
as a function of
the number of
difficulty
interdependencies and
random variables.
Number of interdependencies
and random variables
har01307_ch02_033-056.indd 38 12/17/10 10:13 AM

immediately respond to another patient needing his or her attention. This delay in
response may also set other forces in motion.
It should be clear that the complexity of a system has less to do with the
number of elements in the system than with the number of interdependent rela-
tionships. Even interdependent relationships can vary in degree, causing more
or less impact on overall system behavior. System interdependency may be
either tight or loose depending on how closely elements are linked. Elements
that are tightly coupled have a greater impact on system operation and perfor-
mance than elements that are only loosely connected. When an element such
as a worker or machine is delayed in a tightly coupled system, the impact is
immediately felt by other elements in the system and the entire process may be
brought to a screeching halt.
In a loosely coupled system, activities have only a minor, and often delayed,
impact on other elements in the system. Systems guru Peter Senge (1990) notes
that for many systems, “Cause and effect are not closely related in time and space.”
Sometimes the distance in time and space between cause-and-effect relationships
becomes quite sizable. If enough reserve inventory has been stockpiled, a truck-
ers’ strike cutting off the delivery of raw materials to a transmission plant in one
part of the world may not affect automobile assembly in another part of the world
for weeks. Cause-and-effect relationships are like a ripple of water that diminishes
in impact as the distance in time and space increases.
Obviously, the preferred approach to dealing with interdependencies is to
eliminate them altogether. Unfortunately, this is not entirely possible for most
situations and actually defeats the purpose of having systems in the first place.
The whole idea of a system is to achieve a synergy that otherwise would be un-
attainable if every component were to function in complete isolation. Several
methods are used to decouple system elements or at least isolate their influence
so disruptions are not felt so easily. These include providing buffer inventories,
implementing redundant or backup measures, and dedicating resources to single
tasks. The downside to these mitigating techniques is that they often lead to ex-
cessive inventories and underutilized resources. The point to be made here is that
interdependencies, though they may be minimized somewhat, are simply a fact of
life and are best dealt with through effective coordination and management.
2.4.2 Variability
Variability is a characteristic inherent in any system involving humans and ma-
chinery. Uncertainty in supplier deliveries, random equipment failures, unpre-
dictable absenteeism, and fluctuating demand all combine to create havoc in
planning system operations. Variability compounds the already unpredictable
effect of interdependencies, making systems even more complex and unpredict-
able. Variability propagates in a system so that “highly variable outputs from
one workstation become highly variable inputs to another” (Hopp and Spearman
2000). Table 2.1 identifies the types of random variability that are typical of most
manufacturing and service systems.
har01307_ch02_033-056.indd 39 12/17/10 10:13 AM

TABLE 2.1 Examples of System Variability

Type of Variability Examples
Activity times Operation times, repair times, setup times, move times
Decisions To accept or reject a part, where to direct a particular customer,
which task to perform next
Quantities Lot sizes, arrival quantities, number of workers absent
Event intervals Time between arrivals, time between equipment failures
Attributes Customer preference, part size, skill level
The tendency in systems planning is to ignore variability and calculate system

capacity and performance based on average values. Many commercial scheduling
packages such as MRP (material requirements planning) software work this way.
Ignoring variability distorts the true picture and leads to inaccurate performance
predictions. Designing systems based on average requirements is like deciding
whether to wear a coat based on the average annual temperature or prescribing
the same eyeglasses for everyone based on average eyesight. Adults have been
known to drown in water that was only four feet deep—on the average! Wherever
variability occurs, an attempt should be made to describe the nature or pattern of
the variability and assess the range of the impact that variability might have on
system performance.
Perhaps the most illustrative example of the impact that variability can have
on system behavior is the simple situation where entities enter into a single queue
to wait for a single server. An example of this might be customers lining up in
front of an ATM. Suppose that the time between customer arrivals is exponen-
tially distributed with an average time of one minute and that they take an average
time of one minute, exponentially distributed, to transact their business. In queu-
ing theory, this is called an M/M/1 queuing system. If we calculate system perfor-
mance based solely on average time, there will never be any customers waiting in
the queue. Every minute that a customer arrives the previous customer finishes his
or her transaction. Now if we calculate the number of customers waiting in line,
taking into account the variation, we will discover that the waiting line grows to
infinity (the technical term is that the system “explodes”). Who would guess that
in a situation involving only one interdependent relationship that variation alone
would make the difference between zero items waiting in a queue and an infinite
number in the queue?
By all means, variability should be reduced and even eliminated wherever
possible. System planning is much easier if you don’t have to contend with it.
Where it is inevitable, however, simulation can help predict the impact it will
have on system performance. Likewise, simulation can help identify the degree
of improvement that can be realized if variability is reduced or even eliminated.
For example, it can tell you how much reduction in overall flow time and flow
time variation can be achieved if operation time variation can be reduced by, say,
20 percent.
har01307_ch02_033-056.indd 40 12/17/10 10:13 AM

2.5 System Performance Metrics

Metrics are measures used to assess the performance of a system. At the high-
est level of an organization or business, metrics measure overall performance in
terms of profits, revenues, costs relative to budget, return on assets, and so on.
These metrics are typically financial in nature and show bottom-line performance.
Unfortunately, such metrics are inherently lagging, disguise low-level operational
performance, and are reported only periodically. From an operational standpoint,
it is more beneficial to track such factors as time, quality, quantity, efficiency, and
utilization. These operational metrics reflect immediate activity and are directly
controllable. They also drive the higher financially related metrics. Key opera-
tional metrics that describe the effectiveness and efficiency of manufacturing and
service systems include the following:
• Flow time—the average time it takes for an item or customer to be
processed through the system. Synonyms include cycle time, throughput
time, and manufacturing lead time. For order fulfillment systems, flow
time may also be viewed as customer response time or turnaround time.
A closely related term in manufacturing is makespan, which is the time
to process a given set of jobs. Flow time can be shortened by reducing
activity times that contribute to flow time such as setup, move, operation,
and inspection time. It can also be reduced by decreasing work-in-process
or average number of entities in the system. Since over 80 percent of
cycle time is often spent waiting in storage or queues, elimination of
buffers tends to produce the greatest reduction in cycle time. Another
solution is to add more resources, but this can be costly.
• Utilization—the percentage of scheduled time that personnel, equipment,
and other resources are in productive use. If a resource is not being
utilized, it may be because it is idle, blocked, or down. To increase
productive utilization, you can increase the demand on the resource or
reduce resource count or capacity. It also helps to balance work loads. In
a system with high variability in activity times, it is difficult to achieve
high utilization of resources. Job shops, for example, tend to have low
machine utilization. Increasing utilization for the sake of utilization is not
a good objective. Increasing the utilization of nonbottleneck resources,
for example, often only creates excessive inventories without creating
additional throughput.
• Value-added time—the amount of time material, customers, and so forth
spend actually receiving value, where value is defined as anything for
which the customer is willing to pay. From an operational standpoint,
value-added time is considered the same as processing time or time spent
actually undergoing some physical transformation or servicing. Inspection
time and waiting time are considered non-value-added time.
• Waiting time—the amount of time that material, customers, and so
on spend waiting to be processed. Waiting time is by far the greatest
har01307_ch02_033-056.indd 41 12/17/10 10:13 AM

component of non-value-added time. Waiting time can be decreased by

reducing the number of items (such as customers or inventory levels) in
the system. Reducing variation and interdependencies in the system can
also reduce waiting times. Additional resources can always be added, but
the trade-off between the cost of adding the resources and the savings of
reduced waiting time needs to be evaluated.
• Flow rate—the number of items produced or customers serviced per
unit of time (such as parts or customers per hour). Synonyms include
production rate, processing rate, or throughput rate. Flow rate can be
increased by better management and utilization of resources, especially
the limiting or bottleneck resource. This is done by ensuring that the
bottleneck operation or resource is never starved or blocked. Once system
throughput matches the bottleneck throughput, additional processing
or throughput capacity can be achieved by speeding up the bottleneck
operation, reducing downtimes and setup times at the bottleneck
operation, adding more resources to the bottleneck operation, or off-
loading work from the bottleneck operation.
• Inventory or queue levels—the number of items or customers in storage
or waiting areas. It is desirable to keep queue levels to a minimum while
still achieving target throughput and response time requirements. Where
queue levels fluctuate, it is sometimes desirable to control the minimum
or maximum queue level. Queuing occurs when resources are unavailable
when needed. Inventory or queue levels can be controlled either by
balancing flow or by restricting production at nonbottleneck operations.
JIT (just-in-time) production is one way to control inventory levels.
• Yield—from a production standpoint, the percentage of products
completed that conform to product specifications as a percentage of the
total number of products that entered the system as raw materials. If 95
out of 100 items are nondefective, the yield is 95 percent. Yield can also
be measured by its complement—reject or scrap rate.
• Customer responsiveness—the ability of the system to deliver products
in a timely fashion to minimize customer waiting time. It might be
measured as fill rate, which is the number of customer orders that can
be filled immediately from inventory. In minimizing job lateness, it may
be desirable to minimize the overall late time, minimize the number or
percentage of jobs that are late, or minimize the maximum tardiness
of jobs. In make-to-stock operations, customer responsiveness can be
assured by maintaining adequate inventory levels. In make-to-order,
customer responsiveness is improved by lowering inventory levels so that
cycle times can be reduced.
• Variance—the degree of fluctuation that can and often does occur in any
of the preceding metrics. Variance introduces uncertainty, and therefore
risk, in achieving desired performance goals. Manufacturers and service
providers are often interested in reducing variance in delivery and service
har01307_ch02_033-056.indd 42 12/17/10 10:13 AM

times. For example, cycle times and throughput rates are going to have
some variance associated with them. Variance is reduced by controlling
activity times, improving resource reliability, and adhering to schedules.
These metrics can be given for the entire system, or they can be broken down by
individual resource, entity type, or some other characteristic. By relating these
metrics to other factors, additional meaningful metrics can be derived that are
useful for benchmarking or other comparative analysis. Typical relational metrics
include minimum theoretical flow time divided by actual flow time (flow time
efficiency), cost per unit produced (unit cost), annual inventory sold divided by
average inventory (inventory turns or turnover ratio), or units produced per cost
or labor input (productivity).
2.6 System Variables

Designing a new system or improving an existing system requires more than sim-
ply identifying the elements and performance goals of the system. It requires an
understanding of how system elements affect each other and overall performance
objectives. To comprehend these relationships, you must understand three types
of system variables:
1. Decision variables
2. Response variables
3. State variables
2.6.1 Decision Variables

Decision variables (also called input factors in SimRunner) are sometimes re-
ferred to as the independent variables in an experiment. Changing the values
of a system’s independent variables affects the behavior of the system. Inde-
pendent variables may be either controllable or uncontrollable depending on
whether the experimenter is able to manipulate them. An example of a con-
trollable variable is the number of operators to assign to a production line or
whether to work one or two shifts. Controllable variables are called decision
variables because the decision maker (experimenter) controls the values of the
variables. An uncontrollable variable might be the time to service a customer or
the reject rate of an operation. When defining the system, controllable variables
are the information about the system that is more prescriptive than descriptive
(see section 2.9.3).
Obviously, all independent variables in an experiment are ultimately
controllable—but at a cost. The important point here is that some variables are
easier to change than others. When conducting experiments, the final solution is
often based on whether the cost to implement a change produces a higher return
in performance.
har01307_ch02_033-056.indd 43 12/17/10 10:13 AM

2.6.2 Response Variables

Response variables (sometimes called performance or output variables) measure
the performance of the system in response to particular decision variable settings.
A response variable might be the number of entities processed for a given period,
the average utilization of a resource, or any of the other system performance met-
rics described in section 2.5.
In an experiment, the response variable is the dependent variable, which
depends on the particular value settings of the independent variables. The ex-
perimenter doesn’t manipulate dependent variables, only independent or decision
variables. Obviously, the goal in systems planning is to find the right values or
settings of the decision variables that give the desired response values.
2.6.3 State Variables

State variables indicate the status of the system at any specific point in time.
Examples of state variables are the current number of entities waiting to be pro-
cessed or the current status (busy, idle, down) of a particular resource. Response
variables are often summaries of state variable changes over time. For example,
the individual times that a machine is in a busy state can be summed over a
particular period and divided by the total available time to report the machine
utilization for that period.
State variables are dependent variables like response variables in that they
depend on the setting of the independent variables. State variables are often ig-
nored in experiments since they are not directly controlled like decision variables
and are not of as much interest as the summary behavior reported by response
variables.
Sometimes reference is made to the state of a system as though a system
itself can be in a particular state such as busy or idle. The state of a system actu-
ally consists of “that collection of variables necessary to describe a system at a
particular time, relative to the objectives of the study” (Law and Kelton 2007). If
we study the flow of customers in a bank, for example, the state of the bank for a
given point in time would include the current number of customers in the bank,
the current status of each teller (busy, idle, or whatever), and perhaps the time that
each customer has been in the system thus far.
2.7 System Optimization

Finding the right setting for decision variables that best meets performance objec-
tives is called optimization. Specifically, optimization seeks the best combination
of decision variable values that either minimizes or maximizes some objective
function such as costs or profits. An objective function is simply a response vari-
able of the system. A typical objective in an optimization problem for a manu-
facturing or service system might be minimizing costs or maximizing flow rate.
har01307_ch02_033-056.indd 44 12/17/10 10:13 AM

For example, we might be interested in finding the optimum number of personnel

for staffing a customer support activity that minimizes costs yet handles the call
volume. In a manufacturing concern, we might be interested in maximizing the
throughput that can be achieved for a given system configuration. Optimization
problems often include constraints, limits to the values that the decision variables
can take on. For example, in finding the optimum speed of a conveyor such that
production cost is minimized, there would undoubtedly be physical limits to how
slow or fast the conveyor can operate. Constraints can also apply to response
variables. An example of this might be an objective to maximize throughput but
subject to the constraint that average waiting time cannot exceed 15 minutes.
In some instances, we may find ourselves trying to achieve conflicting objec-
tives. For example, minimizing production or service costs often conflicts with
minimizing waiting costs. In system optimization, one must be careful to weigh
priorities and make sure the right objective function is driving the decisions. If,
for example, the goal is to minimize production or service costs, the obvious solu-
tion is to maintain only a sufficient workforce to meet processing requirements.
Unfortunately, in manufacturing systems this builds up work-in-process and re-
sults in high inventory carrying costs. In service systems, long queues result in
long waiting times, hence dissatisfied customers. At the other extreme, one might
feel that reducing inventory or waiting costs should be the overriding goal and,
therefore, decide to employ more than an adequate number of resources so that
work-in-process or customer waiting time is virtually eliminated. It should be
obvious that there is a point at which the cost of adding another resource can no
longer be justified by the diminishing incremental savings in waiting costs that are
realized. For this reason, it is generally conceded that a better strategy is to find
the right trade-off or balance between the number of resources and waiting times
so that the total cost is minimized (see Figure 2.3).
FIGURE 2.3
Cost curves showing
optimum number of
resources to minimize
total cost.
Total cost
Optimum Resource costs
Cost
Waiting costs
Number of resources
har01307_ch02_033-056.indd 45 12/17/10 10:13 AM

As shown in Figure 2.3, the number of resources at which the sum of the
resource costs and waiting costs is at a minimum is the optimum number of re-
sources to have. It also becomes the optimum acceptable waiting time.
In systems design, arriving at an optimum system design is not always re-
alistic, given the almost endless configurations that are sometimes possible and
limited time that is available. From a practical standpoint, the best that can be
expected is a near optimum solution that gets us close enough to our objective,
given the time constraints for making the decision.
2.8 The Systems Approach

Due to departmentalization and specialization, decisions in the real world are
often made without regard to overall system performance. With everyone busy
minding his or her own area of responsibility, often no one is paying attention
to the big picture. One manager in a large manufacturing corporation noted that
as many as 99 percent of the system improvement recommendations made in his
company failed to look at the system holistically. He further estimated that nearly
80 percent of the suggested changes resulted in no improvement at all, and many
of the suggestions actually hurt overall performance. When attempting to make
system improvements, it is often discovered that localized changes fail to pro-
duce the overall improvement that is desired. Put in technical language: Achiev-
ing a local optimum often results in a global suboptimum. In simpler terms: It’s
okay to act locally as long as one is thinking globally. The elimination of a prob-
lem in one area may only uncover, and sometimes even exacerbate, problems in
other areas.
Approaching system design with overall objectives in mind and consider-
ing how each element relates to each other and to the whole is called a systems
or holistic approach to systems design. Because systems are composed of inter-
dependent elements, it is not possible to accurately predict how a system will
perform simply by examining each system element in isolation from the whole.
To presume otherwise is to take a reductionist approach to systems design, which
focuses on the parts rather than the whole. While structurally a system may be
divisible, functionally it is indivisible and therefore requires a holistic approach
to systems thinking. Kofman and Senge (1995) observe
The defining characteristic of a system is that it cannot be understood as a function

of its isolated components. First, the behavior of the system doesn’t depend on what
each part is doing but on how each part is interacting with the rest. . . . Second, to
understand a system we need to understand how it fits into the larger system of which
it is a part. . . . Third, and most important, what we call the parts need not be taken as
primary. In fact, how we define the parts is fundamentally a matter of perspective and
purpose, not intrinsic in the nature of the “real thing” we are looking at.
Whether designing a new system or improving an existing system, it is

important to follow sound design principles that take into account all relevant
har01307_ch02_033-056.indd 46 12/17/10 10:13 AM

FIGURE 2.4
Four-step iterative Identify problems
approach to systems and opportunities.
improvement.
Select and Develop

implement the alternative
best solution. solutions.
Evaluate the
solutions.
variables. The activity of systems design and process improvement, also called
systems engineering, has been defined as
The effective application of scientific and engineering efforts to transform an opera-
tional need into a defined system configuration through the top-down iterative process
of requirements definition, functional analysis, synthesis, optimization, design, test
and evaluation (Blanchard 1991).
To state it simply, systems engineering is the process of identifying problems or

other opportunities for improvement, developing alternative solutions, evaluating
the solutions, and selecting and implementing the best solutions (see Figure 2.4).
All of this should be done from a systems point of view.
2.8.1 Identifying Problems and Opportunities

The importance of identifying the most significant problem areas and recognizing
opportunities for improvement cannot be overstated. Performance standards should
be set high in order to look for the greatest improvement opportunities. Companies
making the greatest strides are setting goals of 100 to 500 percent improvement
in many areas such as inventory reduction or customer lead time reduction. Set-
ting high standards pushes people to think creatively and often results in break-
through improvements that would otherwise never be considered. Contrast this way
of thinking with one hospital whose standard for whether a patient had a quality
experience was whether the patient left alive! Such lack of vision will never inspire
the level of improvement needed to meet ever-increasing customer expectations.
2.8.2 Developing Alternative Solutions

We usually begin developing a solution to a problem by understanding the prob-
lem, identifying key variables, and describing important relationships. This helps
har01307_ch02_033-056.indd 47 12/17/10 10:13 AM

identify possible areas of focus and leverage points for applying a solution. Tech-
niques such as cause-and-effect analysis and pareto analysis are useful here.
Once a problem or opportunity has been identified and key decision variables
isolated, alternative solutions can be explored. This is where most of the design and
engineering expertise comes into play. Knowledge of best practices for common types
of processes can also be helpful. The designer should be open to all possible alterna-
tive feasible solutions so that the best possible solutions don’t get overlooked.
Generating alternative solutions requires creativity as well as organizational
and engineering skills. Brainstorming sessions, in which designers exhaust every
conceivably possible solution idea, are particularly useful. Designers should
use every stretch of the imagination and not be stifled by conventional solutions
alone. The best ideas come when system planners begin to think innovatively and
break from traditional ways of doing things. Simulation is particularly helpful in
this process in that it encourages thinking in radical new ways.
2.8.3 Evaluating the Solutions

Alternative solutions should be evaluated based on their ability to meet the criteria
established for the evaluation. These criteria often include performance goals,
cost of implementation, impact on the sociotechnical infrastructure, and consis-
tency with organizational strategies. Many of these criteria are difficult to mea-
sure in absolute terms, although most design options can be easily assessed in
terms of relative merit.
After narrowing the list to two or three of the most promising solutions using
common sense and rough-cut analysis, more precise evaluation techniques may
need to be used. This is where simulation and other formal analysis tools come
into play.
2.8.4 Selecting and Implementing the Best Solution

Often the final selection of what solution to implement is not left to the analyst,
but rather is a management decision. The analyst’s role is to present his or her
evaluation in the clearest way possible so that an informed decision can be made.
Even after a solution is selected, additional modeling and analysis are often
needed for fine-tuning the solution. Implementers should then be careful to make
sure that the system is implemented as designed, documenting reasons for any
modifications.
2.9 Systems Analysis Techniques

While simulation is perhaps the most versatile and powerful systems analysis
tool, other available techniques also can be useful in systems planning. These al-
ternative techniques are usually computational methods that work well for simple
systems with little interdependency and variability. For more complex systems,
these techniques still can provide rough estimates but fall short in producing the
har01307_ch02_033-056.indd 48 12/17/10 10:13 AM

insights and accurate answers that simulation provides. Systems implemented

using these techniques usually require some adjustments after implementation to
compensate for inaccurate calculations. For example, if after implementing a sys-
tem it is discovered that the number of resources initially calculated is insufficient
to meet processing requirements, additional resources are added. This adjustment
can create extensive delays and costly modifications if special personnel training
or custom equipment is involved. As a precautionary measure, a safety factor is
sometimes added to resource and space calculations to ensure they are adequate.
Overdesigning a system, however, also can be costly and wasteful.
As system interdependency and variability increase, not only does system
performance decrease, but the ability to accurately predict system performance
decreases as well (Lloyd and Melton 1997). As Robert Pool explains, “Complex-
ity creates unpredictability. The more complex a system, the more difficult it is to
understand all the different ways the system may behave—and, in particular, to
anticipate all the different ways it may fail. Interdependence among parts creates
entirely new ways that things can go wrong, ways that engineers often overlook
or ignore.” Simulation enables a planner to accurately predict the expected perfor-
mance of a system design and ultimately make better design decisions.
Systems analysis tools, in addition to simulation, include simple calculations,
spreadsheets, operations research techniques (such as linear programming and
queuing theory), and special computerized tools for scheduling, layout, and so
forth. While these tools can provide quick and approximate solutions, they tend
to make oversimplifying assumptions, perform only static calculations, and are
limited to narrow classes of problems. Additionally, they fail to fully account for
interdependencies and variability of complex systems and therefore are not as ac-
curate as simulation in predicting complex system performance (see Figure 2.5).
They all lack the power, versatility, and visual appeal of simulation. They do pro-
vide quick solutions, however, and for certain situations produce adequate results.
FIGURE 2.5
Simulation improves 100%
performance With
predictability. simulation
System predictability
50%
Without
simulation
Call centers Banks Airports

Doctor's offices Emergency rooms Hospitals
Machining cells Production lines Factories
0%
Low Medium High
System complexity
har01307_ch02_033-056.indd 49 12/17/10 10:13 AM

They are important to cover here, not only because they sometimes provide a
good alternative to simulation, but also because they can complement simulation
by providing initial design estimates for input to the simulation model. They also
can be useful to help validate the results of a simulation by comparing them with
results obtained using an analytic model.
2.9.1 Hand Calculations

Quick-and-dirty, pencil-and-paper sketches and calculations can be remarkably
helpful in understanding basic requirements for a system. Many important deci-
sions have been made as the result of sketches drawn and calculations performed
on a napkin or the back of an envelope. Some decisions may be so basic that
a quick mental calculation yields the needed results. Most of these calculations
involve simple algebra, such as finding the number of resource units (such as ma-
chines or service agents) to process a particular workload knowing the capacity
per resource unit. For example, if a requirement exists to process 200 items per
hour and the processing capacity of a single resource unit is 75 work items per
hour, three units of the resource, most likely, are going to be needed.
The obvious drawback to hand calculations is the inability to manually per-
form complex calculations or to take into account tens or potentially even hun-
dreds of complex relationships simultaneously.
2.9.2 Spreadsheets
Spreadsheet software comes in handy when calculations, sometimes involving hun-
dreds of values, need to be made. Manipulating rows and columns of numbers on
a computer is much easier than doing it on paper, even with a calculator handy.
Spreadsheets can be used to perform rough-cut analysis such as calculating aver-
age throughput or estimating machine requirements. The drawback to spreadsheet
software is the inability (or, at least, limited ability) to include variability in activity
times, arrival rates, and so on, and to account for the effects of interdependencies.
What-if experiments can be run on spreadsheets based on expected values
(average customer arrivals, average activity times, mean time between equipment
failures) and simple interactions (activity A must be performed before activity B).
This type of spreadsheet simulation can be very useful for getting rough per-
formance estimates. For some applications with little variability and component
interaction, a spreadsheet simulation may be adequate. However, calculations
based on only average values and oversimplified interdependencies potentially
can be misleading and result in poor decisions. As one ProModel user reported,
“We just completed our final presentation of a simulation project and successfully
saved approximately $600,000. Our management was prepared to purchase an
additional overhead crane based on spreadsheet analysis. We subsequently built a
ProModel simulation that demonstrated an additional crane will not be necessary.
The simulation also illustrated some potential problems that were not readily ap-
parent with spreadsheet analysis.”
har01307_ch02_033-056.indd 50 12/17/10 10:13 AM

Another weakness of spreadsheet modeling is the fact that all behavior is

assumed to be period-driven rather than event-driven. Imagine how difficult it
would be to figure out how your bank account balance fluctuated during a par-
ticular period when all you had to go on was your ending monthly balance. Using
ending balances does not reflect changes as they occurred during the period. You
can know the current state of the system at any point in time only by updating
the state variables of the system each time an event or transaction occurs. When
it comes to creating dynamic models, spreadsheet simulation suffers from the
“curse of dimensionality” because the size of the model becomes unmanageable.
2.9.3 Operations Research Techniques

Traditional operations research (OR) techniques utilize mathematical models to
solve problems involving simple to moderately complex relationships. These
mathematical models include both deterministic models such as mathematical
programming, routing, or network flows and probabilistic models such as queu-
ing and decision trees. These OR techniques provide quick, quantitative answers
without going through the guesswork process of trial and error. OR techniques
can be divided into two general classes: prescriptive and descriptive.
Prescriptive Techniques
Prescriptive OR techniques provide an optimum solution to a problem, such as the
optimum amount of resource capacity to minimize costs, or the optimum product
mix that will maximize profits. Examples of prescriptive OR optimization tech-
niques include linear programming and dynamic programming. These techniques
are generally applicable when only a single goal is desired for minimizing or maxi-
mizing some objective function—such as maximizing profits or minimizing costs.
Because optimization techniques are generally limited to optimizing for a single
goal, secondary goals get sacrificed that may also be important. Additionally, these
techniques do not allow random variables to be defined as input data, thereby forcing
the analyst to use average process times and arrival rates that can produce mislead-
ing results. They also usually assume that conditions are constant over the period of
study. In contrast, simulation is capable of analyzing much more complex relation-
ships and time-varying circumstances. With optimization capabilities now provided
in simulation, simulation software has even taken on a prescriptive roll.
Descriptive Techniques
Descriptive techniques such as queuing theory are static analysis techniques that
provide good estimates for basic problems such as determining the expected av-
erage number of entities in a queue or the average waiting times for entities in a
queuing system. Queuing theory is of particular interest from a simulation per-
spective because it looks at many of the same system characteristics and issues
that are addressed in simulation.
Queuing theory is essentially the science of waiting lines (in the United
Kingdom, people wait in queues rather than lines). A queuing system consists of
har01307_ch02_033-056.indd 51 12/17/10 10:13 AM

FIGURE 2.6 Arriving entities Queue Server Departing entities

Queuing system
configuration.
.
.
.
one or more queues and one or more servers (see Figure 2.6). Entities, referred to
in queuing theory as the calling population, enter the queuing system and either
are immediately served if a server is available or wait in a queue until a server
becomes available. Entities may be serviced using one of several queuing disci-
plines: first-in, first-out (FIFO); last-in, first-out (LIFO); priority; and others. The
system capacity, or number of entities allowed in the system at any one time, may
be either finite or, as is often the case, infinite. Several different entity queuing
behaviors can be analyzed such as balking (rejecting entry), reneging (abandon-
ing the queue), or jockeying (switching queues). Different interarrival time dis-
tributions (such as constant or exponential) may also be analyzed, coming from
either a finite or infinite population. Service times may also follow one of several
distributions such as exponential or constant.
Kendall (1953) devised a simple system for classifying queuing systems in
the form A/B/s, where A is the type of interarrival distribution, B is the type of
service time distribution, and s is the number of servers. Typical distribution
types for A and B are as follows:
M for Markovian or exponential distribution
G for a general distribution
D for a deterministic or constant value
An M/D/1 queuing system, for example, is a system in which interarrival times are
exponentially distributed, service times are constant, and there is a single server.
The arrival rate in a queuing system is usually represented by the Greek letter
lambda () and the service rate by the Greek letter mu (). The mean interarrival
time then becomes 1/ and the mean service time is 1/. A traffic intensity factor
/ is a parameter used in many of the queuing equations and is represented by
the Greek letter rho ( ).
Common performance measures of interest in a queuing system are based on
steady-state or long-term expected values and include the following:
L ⫽ expected number of entities in the system (number in the queue and in
service)
Lq ⫽ expected number of entities in the queue (queue length)
har01307_ch02_033-056.indd 52 12/17/10 10:13 AM

W ⫽ expected time in the system (flow time)

Wq ⫽ expected time in the queue (waiting time)
Pn ⫽ probability of exactly n customers in the system (n = 0, 1, . . .)
The M/M/1 system with infinite capacity and a FIFO queue discipline is perhaps
the most basic queuing problem and sufficiently conveys the procedure for ana-
lyzing queuing systems and understanding how the analysis is performed. The
equations for calculating the common performance measures in an M/M/1 are

L ⫽ _______ ⫽ _______
(1 ⫺ ) ( ⫺ )
2
Lq ⫽ L ⫺ ⫽ _______
(1 ⫺ )
1
W ⫽ ______
⫺

Wq ⫽ ________
( ⫺ )
Pn ⫽ (1 ⫺ )n n ⫽ 0, 1, . . .
If either the expected number of entities in the system or the expected waiting
time is known, the other can be calculated easily using Little’s law (1961):
L ⫽ W
Little’s law also can be applied to the queue length and waiting time:
Lq ⫽ Wq
Example: Suppose customers arrive to use an automatic teller machine (ATM) at
an interarrival time of 3 minutes exponentially distributed and spend an average of
2.4 minutes exponentially distributed at the machine. What is the expected number
of customers the system and in the queue? What is the expected waiting time for cus-
tomers in the system and in the queue?
⫽ 20 per hour
⫽ 25 per hour

⫽ __
⫽ .8
Solving for L:

L ⫽ _______
( ⫺ )
20
⫽ _________
(25 ⫺ 20)
20 ⫽ 4
⫽ ___
5
har01307_ch02_033-056.indd 53 12/17/10 10:13 AM

Solving for Lq:

2
Lq ⫽ _______
(1 ⫺ )
.82
⫽ _______
(1 ⫺ .8)
.64
⫽ ___
.2
⫽ 3.2
Solving for W using Little’s formula:
L
W ⫽ __

4
⫽ ___
20
⫽ .20 hrs
⫽ 12 minutes
Solving for Wq using Little’s formula:

Lq
Wq ⫽ __

3.2
⫽ ___
20
⫽ .16 hrs
⫽ 9.6 minutes
Descriptive OR techniques such as queuing theory are useful for the most
basic problems, but as systems become even moderately complex, the prob-
lems get very complicated and quickly become mathematically intractable.
In contrast, simulation provides close estimates for even the most complex
systems (assuming the model is valid). In addition, the statistical output of
simulation is not limited to only one or two metrics but instead provides
information on all performance measures. Furthermore, while OR techniques
give only average performance measures, simulation can generate detailed
time-series data and histograms providing a complete picture of performance
over time.
2.9.4 Special Computerized Tools

Many special computerized tools have been developed for forecasting, schedul-
ing, layout, staffing, and so on. These tools are designed to be used for narrowly
focused problems and are extremely effective for the kinds of problems they
are intended to solve. They are usually based on constant input values and are
computed using static calculations. The main benefit of special-purpose decision
tools is that they are usually easy to use because they are designed to solve a
specific type of problem.
har01307_ch02_033-056.indd 54 12/17/10 10:13 AM

2.10 Summary
An understanding of system dynamics is essential to using any tool for planning
system operations. Manufacturing and service systems consist of interrelated ele-
ments (personnel, equipment, and so forth) that interactively function to produce
a specified outcome (an end product, a satisfied customer, and so on). Systems are
made up of entities (the objects being processed), resources (the personnel, equip-
ment, and facilities used to process the entities), activities (the process steps),
and controls (the rules specifying the who, what, where, when, and how of entity
processing).
The two characteristics of systems that make them so difficult to analyze are
interdependencies and variability. Interdependencies cause the behavior of one
element to affect other elements in the system either directly or indirectly. Vari-
ability compounds the effect of interdependencies in the system, making system
behavior nearly impossible to predict without the use of simulation.
The variables of interest in systems analysis are decision, response, and state
variables. Decision variables define how a system works; response variables in-
dicate how a system performs; and state variables indicate system conditions at
specific points in time. System performance metrics or response variables are
generally time, utilization, inventory, quality, or cost related. Improving system
performance requires the correct manipulation of decision variables. System op-
timization seeks to find the best overall setting of decision variable values that
maximizes or minimizes a particular response variable value.
Given the complex nature of system elements and the requirement to make
good design decisions in the shortest time possible, it is evident that simulation
can play a vital role in systems planning. Traditional systems analysis techniques
are effective in providing quick but often rough solutions to dynamic systems
problems. They generally fall short in their ability to deal with the complexity and
dynamically changing conditions in manufacturing and service systems. Simula-
tion is capable of imitating complex systems of nearly any size and to nearly any
level of detail. It gives accurate estimates of multiple performance metrics and
leads designers toward good design decisions.

1. Why is an understanding of system dynamics important to the use of
simulation?
2. What is a system?
3. What are the elements of a system from a simulation perspective? Give
an example of each.
4. What are two characteristics of systems that make them so complex?
5. What is the difference between a decision variable and a response variable?
6. Identify five input variables of a manufacturing or service system that
tend to be random.
har01307_ch02_033-056.indd 55 12/17/10 10:13 AM

7. Give two examples of state variables.

8. List three performance metrics that you feel would be important for a
computer assembly line.
9. List three performance metrics you feel would be useful for a hospital
emergency room.
10. Define optimization in terms of decision variables and response variables.
11. Is maximizing resource utilization a good overriding performance
objective for a manufacturing system? Explain.
12. What is a systems approach to problem solving?
13. Where does simulation fit into the overall approach of systems engineering?
14. In what situations would you use analytical techniques (like hand
calculations or spreadsheet modeling) over simulation?
15. Assuming you decided to use simulation to determine how many lift
trucks were needed in a distribution center, how might analytical models
be used to complement the simulation study both before and after?
16. What advantages does simulation have over traditional OR techniques
used in systems analysis?
17. Students come to a professor’s office to receive help on a homework
assignment every 10 minutes exponentially distributed. The time to help
a student is exponentially distributed with a mean of 7 minutes. What are
the expected number of students waiting to be helped and the average
time waiting before being helped? For what percentage of time is it
expected there will be more than two students at the professor’s office?
References
Blanchard, Benjamin S. System Engineering Management. New York: John Wiley & Sons,
1991.
Hopp, Wallace J., and M. Spearman. Factory Physics. New York: Irwin/McGraw-Hill,
2000, p. 282.
Kendall, D. G. “Stochastic Processes Occurring in the Theory of Queues and Their
Analysis by the Method of Imbedded Markov Chains.” Annals of Mathematical
Statistics 24 (1953), pp. 338–54.
Kofman, Fred, and P. Senge. Communities of Commitment: The Heart of Learning
Organizations. Sarita Chawla and John Renesch, (eds.), Portland, OR. Productivity
Press, 1995.
Law, Averill M. Simulation Modeling and Analysis 4th ed. New York: McGraw-Hill, 2007.
Little, J. D. C. “A Proof for the Queuing Formula: L ⫽ W.” Operations Research 9, no. 3
(1961), pp. 383–87.
Lloyd, S., and K. Melton. “Using Statistical Process Control to Obtain More Precise
Distribution Fitting Using Distribution Fitting Software.” Simulators International
XIV 29, no. 3 (April 1997), pp. 193–98.
Pool, Robert. Beyond Engineering. Oxford University Press, 1997, p. 131.
Senge, Peter. The Fifth Discipline. New York: Doubleday, 1990.
Simon, Herbert A. Models of Man. New York: John Wiley & Sons, 1957, p. 198.
har01307_ch02_033-056.indd 56 12/17/10 10:13 AM

C H A P T E R
3 SIMULATION BASICS
“Zeal without knowledge is fire without light.”

—Dr. Thomas Fuller
3.1 Introduction
Simulation is much more meaningful when we understand what it is actually
doing. Understanding how simulation works helps us to know whether we are
applying it correctly and what the output results mean. Many books have been
written that give thorough and detailed discussions of the science of simulation
(see Banks et al. 2001; Hoover and Perry 1989; Law 2007; Pooch and Wall 1993;
Ross 1990; Shannon 1975; Thesen and Travis 1992; and Widman, Loparo, and
Nielsen 1989). This chapter attempts to summarize the basic technical issues
related to simulation that are essential to understand in order to get the greatest
benefit from the tool. The chapter discusses the different types of simulation and
how random behavior is simulated. A spreadsheet simulation example is given
in this chapter to illustrate how various techniques are combined to simulate the
behavior of a common system.
3.2 Types of Simulation

The way simulation works is based largely on the type of simulation used. There
are many ways to categorize simulation. Some of the most common include
• Static or dynamic.
• Stochastic or deterministic.
• Discrete event or continuous.
57
har01307_ch03_057-086.indd 57 20/12/10 1:49 PM

The type of simulation we focus on in this book can be classified as dynamic,

stochastic, discrete-event simulation. To better understand this classification, we
will look at what these three different simulation categories mean.
3.2.1 Static versus Dynamic Simulation

A static simulation is one that is not based on time. It often involves drawing
random samples to generate a statistical outcome, so it is sometimes called Monte
Carlo simulation. In finance, Monte Carlo simulation is used to select a portfolio
of stocks and bonds. Given a portfolio, with different probabilistic payouts, it is
possible to generate an expected yield. One material handling system supplier
developed a static simulation model to calculate the expected time to travel from
one rack location in a storage system to any other rack location. A random sample
of 100 from–to relationships were used to estimate an average travel time. Had
every from–to trip been calculated, a 1,000-location rack would have involved
1,000 factorial calculations.
Dynamic simulation includes the passage of time. It looks at state changes
as they occur over time. A clock mechanism moves forward in time and state
variables are updated as time advances. Dynamic simulation is well suited for
analyzing manufacturing and service systems since they operate over time.
3.2.2 Stochastic versus Deterministic Simulation

Simulations in which one or more input variables are random are referred to as
stochastic or probabilistic simulations. A stochastic simulation produces output
that is itself random and therefore gives only one data point of how the system
might behave.
Simulations having no input components that are random are said to be
deterministic. Deterministic simulation models are built the same way as stochastic
models except that they contain no randomness. In a deterministic simulation, all
future states are determined once the input data and initial state have been defined.
As shown in Figure 3.1, deterministic simulations have constant inputs and
produce constant outputs. Stochastic simulations have random inputs and produce
random outputs. Inputs might include activity times, arrival intervals, and rout-
ing sequences. Outputs include metrics such as average flow time, flow rate, and
resource utilization. Any output impacted by a random input variable is going to
also be a random variable. That is why the random inputs and random outputs of
Figure 3.1(b) are shown as statistical distributions.
A deterministic simulation will always produce the exact same outcome no
matter how many times it is run. In stochastic simulation, several randomized runs
or replications must be made to get an accurate performance estimate because
each run varies statistically. Performance estimates for stochastic simulations are
obtained by calculating the average value of the performance metric across all of
the replications. In contrast, deterministic simulations need to be run only once to
get precise results because the results are always the same.
har01307_ch03_057-086.indd 58 20/12/10 1:49 PM

Chapter 3 Simulation Basics 59
FIGURE 3.1
Examples of (a) a deterministic simulation and (b) a stochastic simulation.
Constant Constant Random Random

inputs outputs inputs outputs
7
12.3
3.4 Simulation Simulation
106
5
(a) (b)
3.2.3 Discrete-Event versus Continuous Simulation

A discrete-event simulation is one in which state changes occur at discrete points
in time as triggered by events. Typical simulation events might include
• The arrival of an entity to a workstation.
• The failure of a resource.
• The completion of an activity.
• The end of a shift.
State changes in a model occur when some event happens. The state of the model
becomes the collective state of all the elements in the model at a particular point
in time. State variables in a discrete-event simulation are referred to as discrete-
change state variables. A restaurant simulation is an example of a discrete-event
simulation because all of the state variables in the model, such as the number of
customers in the restaurant, are discrete-change state variables (see Figure 3.2).
Most manufacturing and service systems are typically modeled using discrete-
event simulation.
In continuous simulation, state variables change continuously with respect to
time and are therefore referred to as continuous-change state variables. An ex-
ample of a continuous-change state variable is the level of oil in an oil tanker that is
being either loaded or unloaded, or the temperature of a building that is controlled
by a heating and cooling system. Some readers are perhaps familiar with a set
of nonlinear partial differential equations called Navier-Stokes equations used to
model the continuous flow of liquids and other substances. Figure 3.3 compares a
discrete-change state variable and a continuous-change state variable as they vary
over time.
Continuous simulation products use either differential equations or differ-
ence equations to define the rates of change in state variables over time.
har01307_ch03_057-086.indd 59 20/12/10 1:49 PM

FIGURE 3.2 Number of

Discrete events cause customers
in restaurant
discrete state changes.
1 …
Time
Start Event 1 Event 2 Event 3
Simulation (customer (customer (customer
arrives) arrives) departs)
FIGURE 3.3
Continuous-change
Comparison of a state variable
discrete-change
state variable and a
continuous-change Value
Discrete-change
state variable.
state variable
Time
Differential Equations
The change that occurs in some continuous-change state variables is expressed in
terms of the derivatives of the state variables. Equations involving derivatives of
a state variable are referred to as differential equations. The state variable , for
example, might change as a function of both and time t:
d (t)
_____ 2(t) t2
dt
We then need a second equation to define the initial condition of :
(0) K
On a computer, numerical integration is used to calculate the change in a par-
ticular response variable over time. Numerical integration is performed at the end
of successive small time increments referred to as steps. Numerical analysis tech-
niques, such as Runge-Kutta integration, are used to integrate the differential equa-
tions numerically for each incremental time step. One or more threshold values for
each continuous-change state variable are usually defined that determine when some
action is to be triggered, such as shutting off a valve or turning on a pump.
har01307_ch03_057-086.indd 60 20/12/10 1:49 PM

Difference Equations
Sometimes a continuous-change state variable can be modeled using difference
equations. In such instances, the time is decomposed into periods of length t. An
algebraic expression is then used to calculate the value of the state variable at the
end of period k 1 based on the value of the state variable at the end of period k.
For example, the following difference equation might be used to express the rate
of change in the state variable as a function of the current value of , a rate of
change (r), and the length of the time period (t):
(k 1) (k) rt
Batch processing in which fluids are pumped into and out of tanks can often
be modeled using difference equations.
Combined Continuous and Discrete Simulation

Many simulation software products provide both discrete-event and continu-
ous simulation capabilities. This enables systems that have both discrete-event
and continuous characteristics to be modeled, resulting in a hybrid simulation.
Most processing systems that have continuous-change state variables also have
discrete-change state variables. For example, a truck or tanker arrives at a fill sta-
tion (a discrete event) and begins filling a tank (a continuous process).
Four basic interactions occur between discrete- and continuous-change variables:
1. A continuous variable value may suddenly increase or decrease as the
result of a discrete event (like the replenishment of inventory in an
inventory model).
2. The initiation of a discrete event may occur as the result of reaching a
threshold value in a continuous variable (like reaching a reorder point in
an inventory model).
3. The change rate of a continuous variable may be altered as the result of a
discrete event (a change in inventory usage rate as the result of a sudden
change in demand).
4. An initiation or interruption of change in a continuous variable may
occur as the result of a discrete event (the replenishment or depletion of
inventory initiates or terminates a continuous change of the continuous
variable).
3.3 Random Behavior

Stochastic systems frequently have time or quantity values that vary within a
given range and according to specified density, as defined by a probability distri-
bution. Probability distributions are useful for predicting the next time, distance,
quantity, and so forth when these values are random variables. For example, if an
operation time varies between 2.2 minutes and 4.5 minutes, it would be defined
in the model as a probability distribution. Probability distributions are defined
by specifying the type of distribution (normal, exponential, or another type) and
har01307_ch03_057-086.indd 61 20/12/10 1:49 PM

FIGURE 3.4 p(x) 1.0

Examples of (a) a
discrete probability .8
distribution and (b) a
continuous probability
distribution. .6 (a )
.4
.2
0
0 1 2 3 4 5 6 7 x
Discrete Value
f (x) 1.0
.8
.6
(b )
.4
.2
0
0 1 2 3 4 5 6 7 x
Continuous Value
the parameters that describe the shape or density and range of the distribution.
For example, we might describe the time for a check-in operation to be normally
distributed with a mean of 5.2 minutes and a standard deviation of 0.4 minute.
During the simulation, values are obtained from this distribution for successive
operation times. The shape and range of time values generated for this activity
will correspond to the parameters used to define the distribution. When we gener-
ate a value from a distribution, we call that value a random variate.
Probability distributions from which we obtain random variates may be either
discrete (they describe the likelihood of specific values occurring) or continuous
(they describe the likelihood of a value being within a given range). Figure 3.4
shows graphical examples of a discrete distribution and a continuous distribution.
A discrete distribution represents a finite or countable number of possible
values. An example of a discrete distribution is the number of items in a lot or
individuals in a group of people. A continuous distribution represents a continuum
har01307_ch03_057-086.indd 62 20/12/10 1:49 PM

of values. An example of a continuous distribution is a machine with a cycle

time that is uniformly distributed between 1.2 minutes and 1.8 minutes. An infi-
nite number of possible values exist within this range. Discrete and continuous
distributions are further defined in Chapter 5. Appendix A describes many of the
distributions used in simulation.
3.4 Simulating Random Behavior

One of the most powerful features of simulation is its ability to mimic random be-
havior or variation that is characteristic of stochastic systems. Simulating random
behavior requires that a method be provided to generate random numbers as well
as routines for generating random variates based on a given probability distribu-
tion. Random numbers and random variates are defined in the next sections along
with the routines that are commonly used to generate them.
3.4.1 Generating Random Numbers

Random behavior is imitated in simulation by using a random number genera-
tor. The random number generator operates deep within the heart of a simulation
model, pumping out a stream of random numbers. It provides the foundation for
simulating “random” events occurring in the simulated system such as the ar-
rival time of cars to a restaurant’s drive-through window; the time it takes the
driver to place an order; the number of hamburgers, drinks, and fries ordered; and
the time it takes the restaurant to prepare the order. The input to the procedures
used to generate these types of events is a stream of numbers that are uniformly
distributed between zero and one (0 x 1). The random number generator is
responsible for producing this stream of independent and uniformly distributed
numbers (Figure 3.5).
Before continuing, it should be pointed out that the numbers produced by
a random number generator are not “random” in the truest sense. For example,
the generator can reproduce the same sequence of numbers again and again,
which is not indicative of random behavior. Therefore, they are often referred
FIGURE 3.5 f(x)

The uniform(0, 1)
distribution of a
random number for 0 ⱕ x ⱕ 1
generator. f(x) ⫽ { 01 elsewhere
1.0
Mean ⫽ ␮ ⫽ 1
2
Variance ⫽ ␴ 2 ⫽ 1
12
0 1.0 x
har01307_ch03_057-086.indd 63 20/12/10 1:49 PM

to as pseudo-random number generators (pseudo comes from Greek and means

false or fake). Practically speaking, however, “good” pseudo-random number
generators can pump out long sequences of numbers that pass statistical tests for
randomness (the numbers are independent and uniformly distributed). Thus the
numbers approximate real-world randomness for purposes of simulation, and the
fact that they are reproducible helps us in two ways. It would be difficult to debug
a simulation program if we could not “regenerate” the same sequence of random
numbers to reproduce the conditions that exposed an error in our program. We
will also learn in Chapter 9 how reproducing the same sequence of random num-
bers is useful when comparing different simulation models. For brevity, we will
drop the pseudo prefix as we discuss how to design and keep our random number
generator healthy.
Linear Congruential Generators

There are many types of established random number generators, and researchers
are actively pursuing the development of new ones (L’Ecuyer 1998). However,
most simulation software is based on linear congruential generators (LCG). The
LCG is efficient in that it quickly produces a sequence of random numbers with-
out requiring a great deal of computational resources. Using the LCG, a sequence
of integers Z1, Z2, Z3, . . . is defined by the recursive formula
Zi (aZi1 c) mod (m)
where the constant a is called the multiplier, the constant c the increment, and
the constant m the modulus (Law 2007). The user must provide a seed or starting
value, denoted Z0, to begin generating the sequence of integer values. Z0, a, c, and
m are all nonnegative integers. The value of Zi is computed by dividing (aZi1 c)
by m and setting Zi equal to the remainder part of the division, which is the result
returned by the mod function. Therefore, the Zi values are bounded by 0 Zi
m 1 and are uniformly distributed in the discrete case. However, we desire the
continuous version of the uniform distribution with values ranging between a low
of zero and a high of one, which we will denote as Ui for i 1, 2, 3, . . . . Accord-
ingly, the value of Ui is computed by dividing Zi by m.
In a moment, we will consider some requirements for selecting the values for
a, c, and m to ensure that the random number generator produces a long sequence
of numbers before it begins to repeat them. For now, however, let’s assign the
following values a 21, c 3, and m 16 and generate a few pseudo-random
numbers. Table 3.1 contains a sequence of 20 random numbers generated from the
recursive formula
Zi (21Zi1 3) mod (16)
An integer value of 13 was somewhat arbitrarily selected between 0 and
m 1 16 1 15 as the seed (Z0 13) to begin generating the sequence of
numbers in Table 3.1. The value of Z1 is obtained as
Z1 (aZ0 c) mod (m) (21(13) 3) mod (16) (276) mod (16) 4
har01307_ch03_057-086.indd 64 20/12/10 1:49 PM

TABLE 3.1 Example LCG Zi ⴝ (21Ziⴚ1 ⴙ 3) mod (16),

with Z0 13
i 21Zi1 3 Zi Ui Zi /16
0 13
1 276 4 0.2500
2 87 7 0.4375
3 150 6 0.3750
4 129 1 0.0625
5 24 8 0.5000
6 171 11 0.6875
7 234 10 0.6250
8 213 5 0.3125
9 108 12 0.7500
10 255 15 0.9375
11 318 14 0.8750
12 297 9 0.5625
13 192 0 0.0000
14 3 3 0.1875
15 66 2 0.1250
16 45 13 0.8125
17 276 4 0.2500
18 87 7 0.4375
19 150 6 0.3750
20 129 1 0.0625
Note that 4 is the remainder term from dividing 16 into 276. The value of U1 is
computed as
U1 Z116 416 0.2500
The process continues using the value of Z1 to compute Z2 and then U2.
Changing the value of Z0 produces a different sequence of uniform(0, 1) num-
bers. However, the original sequence can always be reproduced by setting the
seed value back to 13 (Z0 13). The ability to repeat the simulation experiment
under the exact same “random” conditions is very useful, as will be demonstrated
in Chapter 9 with a technique called common random numbers.
Note that in ProModel the sequence of random numbers is not generated
in advance and then read from a table as we have done. Instead, the only value
saved is the last Zi value that was generated. When the next random number in the
sequence is needed, the saved value is fed back to the generator to produce the
next random number in the sequence. In this way, the random number generator
is called each time a new random event is scheduled for the simulation.
Due to computational constraints, the random number generator cannot go on
indefinitely before it begins to repeat the same sequence of numbers. The LCG in
Table 3.1 will generate a sequence of 16 numbers before it begins to repeat itself.
You can see that it began repeating itself starting in the 17th position. The value
har01307_ch03_057-086.indd 65 20/12/10 1:49 PM

of 16 is referred to as the cycle length of the random number generator, which is

disturbingly short in this case. A long cycle length is desirable so that each replica-
tion of a simulation is based on a different segment of random numbers. This is
how we collect independent observations of the model’s output.
Let’s say, for example, that to run a certain simulation model for one replica-
tion requires that the random number generator be called 1,000 times during the
simulation and we wish to execute five replications of the simulation. The first
replication would use the first 1,000 random numbers in the sequence, the second
replication would use the next 1,000 numbers in the sequence (the number that
would appear in positions 1,001 to 2,000 if a table were generated in advance),
and so on. In all, the random number generator would be called 5,000 times. Thus
we would need a random number generator with a cycle length of at least 5,000.
The maximum cycle length that an LCG can achieve is m. To realize the
maximum cycle length, the values of a, c, and m have to be carefully selected. A
guideline for the selection is to assign (Pritsker 1995)
1. m 2b, where b is determined based on the number of bits per word on
the computer being used. Many computers use 32 bits per word, making
31 a good choice for b.
2. c and m such that their greatest common factor is 1 (the only positive
integer that exactly divides both m and c is 1).
3. a 1 4k, where k is an integer.
Following this guideline, the LCG can achieve a full cycle length of over 2.1 bil-
lion (231 to be exact) random numbers.
Frequently, the long sequence of random numbers is subdivided into smaller
segments. These subsegments are referred to as streams. For example, Stream 1
could begin with the random number in the first position of the sequence and
continue down to the random number in the 200,000th position of the sequence.
Stream 2, then, would start with the random number in the 200,001st position of
the sequence and end at the 400,000th position, and so on. Using this approach,
each type of random event in the simulation model can be controlled by a unique
stream of random numbers. For example, Stream 1 could be used to generate the
arrival pattern of cars to a restaurant’s drive-through window and Stream 2 could
be used to generate the time required for the driver of the car to place an order.
This assumes that no more than 200,000 random numbers are needed to simulate
each type of event. The practical and statistical advantages of assigning unique
streams to each type of event in the model are described in Chapter 9.
To subdivide the generator’s sequence of random numbers into streams, you
first need to decide how many random numbers to place in each stream. Next, you
begin generating the entire sequence of random numbers (cycle length) produced
by the generator and recording the Zi values that mark the beginning of each
stream. Therefore, each stream has its own starting or seed value. When using the
random number generator to drive different events in a simulation model, the pre-
viously generated random number from a particular stream is used as input to the
generator to generate the next random number from that stream. For convenience,
har01307_ch03_057-086.indd 66 20/12/10 1:49 PM

you may want to think of each stream as a separate random number generator to
be used in different places in the model. For example, see Figure 9.5 in Chapter 9.
There are two types of linear congruential generators: the mixed congruential
generator and the multiplicative congruential generator. Mixed congruential gen-
erators are designed by assigning c > 0. Multiplicative congruential generators are
designed by assigning c 0. The multiplicative generator is more efficient than the
mixed generator because it does not require the addition of c. The maximum cycle
length for a multiplicative generator can be set within one unit of the maximum
cycle length of the mixed generator by carefully selecting values for a and m. From
a practical standpoint, the difference in cycle length is insignificant considering
that both types of generators can boast cycle lengths of more than 2.1 billion.
ProModel uses the following multiplicative generator:
Zi (630,360,016Zi1) mod (231 1)
Specifically, it is a prime modulus multiplicative linear congruential generator
(PMMLCG) with a 630,360,016, c 0, and m 231 1. It has been exten-
sively tested and is known to be a reliable random number generator for simu-
lation (Law and Kelton 2000). The ProModel implementation of this generator
divides the cycle length of 231 1 2,147,483,647 into 100 unique streams.
Testing Random Number Generators

When faced with using a random number generator about which you know very
little, it is wise to verify that the numbers emanating from it satisfy the two impor-
tant properties defined at the beginning of this section. The numbers produced by
the random number generator must be (1) independent and (2) uniformly distrib-
uted between zero and one (uniform(0, 1)). To verify that the generator satisfies
these properties, you first generate a sequence of random numbers U1, U2, U3, . . .
and then subject them to an appropriate test of hypothesis.
The hypotheses for testing the independence property are
H0: Ui values from the generator are independent
H1: Ui values from the generator are not independent
Several statistical methods have been developed for testing these hypotheses at
a specified significance level . One of the most commonly used methods is the
runs test. Banks et al. (2001) review three different versions of the runs test for
conducting this independence test. Additionally, two runs tests are implemented in
Stat::Fit—the Runs Above and Below the Median Test and the Runs Up and Runs
Down Test. Chapter 5 contains additional material on tests for independence.
The hypotheses for testing the uniformity property are
H0: Ui values are uniform(0, 1)
H1: Ui values are not uniform(0, 1)
Several statistical methods have also been developed for testing these hypoth-
eses at a specified significance level . The Kolmogorov-Smirnov test and the
chi-square test are perhaps the most frequently used tests. (See Chapter 5 for
har01307_ch03_057-086.indd 67 20/12/10 1:49 PM

a description of the chi-square test.) The objective is to determine if the uni-

form(0, 1) distribution fits or describes the sequence of random numbers produced
by the random number generator. These tests are included in the Stat::Fit software
and are further described in many introductory textbooks on probability and sta-
tistics (see, for example, Johnson 1994).
3.4.2 Generating Random Variates

This section introduces common methods for generating observations (random
variates) from probability distributions other than the uniform(0, 1) distribution.
For example, the time between arrivals of cars to a restaurant’s drive-through
window may be exponentially distributed, and the time required for drivers to
place orders at the window might follow the lognormal distribution. Observations
from these and other commonly used distributions are obtained by transform-
ing the observations generated by the random number generator to the desired
distribution. The transformed values are referred to as variates from the specified
distribution.
There are several methods for generating random variates from a de-
sired distribution, and the selection of a particular method is somewhat de-
pendent on the characteristics of the desired distribution. Methods include the
inverse transformation method, the acceptance/rejection method, the com-
position method, the convolution method, and the methods employing special
properties. The inverse transformation method, which is commonly used to
generate variates from both discrete and continuous distributions, is described
starting first with the continuous case. For a review of the other methods,
see Law and Kelton (2007).
Continuous Distributions
The application of the inverse transformation method to generate random variates
from continuous distributions is straightforward and efficient for many continu-
ous distributions. For a given probability density function f(x), find the cumula-
tive distribution function of X. That is, F(x) P(X x). Next, set U F(x),
where U is uniform(0, 1), and solve for x. Solving for x yields x F1(U). The
equation x F1(U) transforms U into a value for x that conforms to the given
distribution f(x).
As an example, suppose that we need to generate variates from the expo-
nential distribution with mean . The probability density function f(x) and cor-
responding cumulative distribution function F(x) are

1 ex/
__ for x 0
f(x)
0 elsewhere
1
x/
F(x)
e for x 0
0 elsewhere
har01307_ch03_057-086.indd 68 20/12/10 1:49 PM

Setting U F(x) and solving for x yields
U 1 ex/
e x 1 U

ln (e ) ln (1 U)
x
where ln is the natural logarithm
x ln (1 U)
x ln (1 U)
The random variate x in the above equation is exponentially distributed with mean
provided U is uniform(0, 1).
Suppose three observations of an exponentially distributed random variable
with mean 2 are desired. The next three numbers generated by the random
number generator are U1 0.27, U2 0.89, and U3 0.13. The three numbers
are transformed into variates x1, x2, and x3 from the exponential distribution with
mean 2 as follows:
x1 2 ln (1 U1) 2 ln (1 0.27) 0.63

x2 2 ln (1 U2) 2 ln (1 0.89) 4.41
x3 2 ln (1 U3) 2 ln (1 0.13) 0.28
Figure 3.6 provides a graphical representation of the inverse transformation

method in the context of this example. The first step is to generate U, where U is
uniform(0, 1). Next, locate U on the y axis and draw a horizontal line from that
point to the cumulative distribution function [F(x) 1 ex/2]. From this point
of intersection with F(x), a vertical line is dropped down to the x axis to obtain
the corresponding value of the variate. This process is illustrated in Figure 3.6 for
generating variates x1 and x2 given U1 0.27 and U2 0.89.
Application of the inverse transformation method is straightforward as long
as there is a closed-form formula for the cumulative distribution function, which
FIGURE 3.6 F(x)

Graphical
representation of 1.00
inverse transformation U2 = 1 – e –x2/ 2 = 0.89
method for continuous
variates.
0.50
U1 = 1 – e –x1/ 2 = 0.27
x1 = –2 ln (1 – 0.27) = 0.63 x2 = –2 ln (1 – 0.89) = 4.41
har01307_ch03_057-086.indd 69 20/12/10 1:49 PM

is the case for many continuous distributions. However, the normal distribution
is one exception. Thus it is not possible to solve for a simple equation to gener-
ate normally distributed variates. For these cases, there are other methods that
can be used to generate the random variates. See, for example, Law (2007) for a
description of additional methods for generating random variates from continuous
distributions.
Discrete Distributions
The application of the inverse transformation method to generate variates from
discrete distributions is basically the same as for the continuous case. The differ-
ence is in how it is implemented. For example, consider the following probability
mass function:

0.10 for x 1
p(x) P(X x) 0.30 for x 2
0.60 for x 3
The random variate x has three possible values. The probability that x is equal to 1
is 0.10, P(X 1) 0.10; P(X 2) 0.30; and P(X 3) 0.60. The cumulative
distribution function F(x) is shown in Figure 3.7. The random variable x could be
used in a simulation to represent the number of defective components on a circuit
board or the number of drinks ordered from a drive-through window, for example.
Suppose that an observation from the above discrete distribution is desired.
The first step is to generate U, where U is uniform(0, 1). Using Figure 3.7, the
value of the random variate is determined by locating U on the y axis, drawing
a horizontal line from that point until it intersects with a step in the cumulative
function, and then dropping a vertical line from that point to the x axis to read the
FIGURE 3.7 F(x)

Graphical 1.00
explanation of inverse
transformation method U 2 = 0.89
for discrete variates.
0.40
U 1 = 0.27
0.10
U 3 = 0.05
1 2 3
x3 = 1 x1 = 2 x2 ⫽ 3
har01307_ch03_057-086.indd 70 20/12/10 1:49 PM

value of the variate. This process is illustrated in Figure 3.7 for x1, x2, and x3 given
U1 0.27, U2 0.89, and U3 0.05. Equivalently, if 0 Ui 0.10, then xi 1;
if 0.10 < Ui 0.40, then xi 2; if 0.40 < Ui 1.00, then xi 3. Note that should
we generate 100 variates using the above process, we would expect a value of 3 to
be returned 60 times, a value of 2 to be returned 30 times, and a value of 1 to be
returned 10 times.
The inverse transformation method can be applied to any discrete distribution
by dividing the y axis into subintervals as defined by the cumulative distribution
function. In our example case, the subintervals were [0, 0.10], (0.10, 0.40], and
(0.40, 1.00]. For each subinterval, record the appropriate value for the random
variate. Next, develop an algorithm that calls the random number generator to re-
ceive a specific value for U, searches for and locates the subinterval that contains
U, and returns the appropriate value for the random variate.
Locating the subinterval that contains a specific U value is relatively straight-
forward when there are only a few possible values for the variate. However, the
number of possible values for a variate can be quite large for some discrete distri-
butions. For example, a random variable having 50 possible values could require
that 50 subintervals be searched before locating the subinterval that contains U.
In such cases, sophisticated search algorithms have been developed that quickly
locate the appropriate subinterval by exploiting certain characteristics of the dis-
crete distribution for which the search algorithm was designed. A good source of
information on the subject can be found in Law (2007).
3.4.3 Generating Random Variates from

Common Continuous Distributions
Section 3.4.2 introduced methodologies for generating random variates. In this
section, specific equations derived mostly from those methodologies are presented
for generating random variates from some common continuous distributions.
Uniform
The continuous uniform distribution is used to generate random variates from all
other distributions. By itself, it is used in simulation models when the modeler
wishes to make all observations equally likely to occur. It is also used when little
is known about the activity being simulated other than the activity’s lowest and
highest values.
The continuous uniform probability density function is given by

1
_____ for a x b
f (x) b a
0 elsewhere
with min a and max b. Using the inverse transformation method, the resulting
equation for generating variate x from the uniform(a, b) distribution is given by
x a (b a) U
where U is uniform(0, 1).
har01307_ch03_057-086.indd 71 20/12/10 1:49 PM

Triangular
Like the uniform distribution, the triangular distribution is used when little is
known about the activity being simulated other than the activity’s lowest value,
highest value, and most frequently occurring value—the mode.
The triangular probability density function is given by

2(x a)
_____________ for a x m
(b a)(m a)
f (x) _____________
2(b x)
for m x b
(b a)(b m)
0 elsewhere
with min a, max b, and mode m. Using the inverse transformation method,
the resulting equation for generating variate x from the triangular(a, m, b) distri-
bution is given by

_______________ (m a)
a (b a)(m a)U for 0 U _______
(b a)
x ____________________ (m a)
b (b a)(b m)(1 U) for _______ U 1
(b a)
where U is uniform(0, 1). Note that when the probability density function of the
random variable x is defined over separate subintervals with respect to x, then
the equation for generating a variate will be defined over separate subintervals
with respect to U. Therefore, step one is to generate U. Step two is to determine
which subinterval the value of U falls within. And step three is to plug U into the
corresponding equation to generate variate x for that subinterval.
Normal
The normal distribution is used to simulate errors such as the diameter of a pipe
with respect to its design specification. Some use it to model time such as process-
ing times. However, if this is done, negative values should be discarded.
The normal probability density function is given by
1
f (x) _______
_____
22
e(x) /(2 )
2 2
for all real x
with mean and variance 2. A method developed by Box and Muller and re-
ported in Law (2007) involves generating a variate x from the standard normal
distribution (mean 0 and variance 2 1) and then converting it into a vari-
ate x
from the normal distribution with the desired mean and variance using the
equation x
x.
The method generates two standard normally( 0, 2 1) distributed vari-
ates x1 and x2 and then converts them into normally( , 2) distributed variates x1
and x2
. The method is given by
________
x1 2 ln U1 cos 2U2
x1
x1
har01307_ch03_057-086.indd 72 20/12/10 1:49 PM

and
________
x2 2 ln U1 sin 2U2
x2
x2
where U1 and U2 are independent observations from the uniform(0, 1) distribution.
3.4.4 Generating Random Variates from

Common Discrete Distributions
Section 3.4.2 introduced methodologies for generating random variates. In this
section specific equations derived from those methodologies are presented for
generating random variates from some common discrete distributions.
Uniform
Like its continuous cousin, the discrete (consecutive integers) uniform distribu-
tion is used in simulation models when the modeler wishes to make all obser-
vations equally likely to occur. The difference is that the observations from a
discrete uniform distribution are integer values such as the values achieved by
rolling a single die. It is also used when little is known about the activity being
simulated other than the activity’s lowest and highest values.
The discrete uniform probability mass function is given by

1
_________
p ( x) b a 1 for x a, a 1, . . . , b
0 elsewhere
with min a and max b. Using the inverse transformation method, the result-
ing equation for generating discrete uniform(a, b) variate x is given by
x a Ñ(b a 1)UÅ
where U is continuous uniform(0, 1) and ÑtÅ denotes the largest integer t.
Bernoulli
The bernoulli distribution is used to model random events with two possible out-
comes (e.g., a failure or a success). It is defined by the probability of getting a suc-
cess p. If, for example, the probability that a machine produces a good part (one
that meets specifications) is p each time the machine makes a part, the bernoulli
distribution would be used to simulate if a part is good or bad.
The bernoulli probability mass function is given by

1p for x 0
p (x) p for x 1
0 elsewhere
with probability of success p. Motivated by the inverse transformation method,
the equation for generating bernoulli( p) variate x is given by
x 10 for U p
elsewhere
har01307_ch03_057-086.indd 73 20/12/10 1:49 PM

Geometric
The geometric distribution is used to model random events in which the modeler
is interested in simulating the number of independent bernoulli trials with prob-
ability of success p on each trial needed before the first success occurs. In other
words if x represents failures, how many failures will occur before the first suc-
cess? In the context of an inspection problem with a success denoted as a bad
item, how many good items will be built before the first bad item is produced?
The geometric probability mass function is given by
p (x) p (1 p)x
0
for x 0, 1, . . .
elsewhere
with x number of independent bernoulli trials each with probability of success p
occurring before the first success is realized. Motivated by the inverse transforma-
tion method, the equation for generating geometric( p) variate x is given by
x Ñln Uln (1 p)Å
3.5 Simple Spreadsheet Simulation

This chapter has covered some really useful information, and it will be construc-
tive to pull it all together for our first dynamic, stochastic simulation model. The
simulation will be implemented as a spreadsheet model because the example sys-
tem is not so complex as to require the use of a commercial simulation software
product. Furthermore, a spreadsheet will provide a nice tabulation of the random
events that produce the output from the simulation, thereby making it ideal for
demonstrating how simulations of dynamic, stochastic systems are often accom-
plished. The system to be simulated follows.
Customers arrive to use an automatic teller machine (ATM) at a mean inter-
arrival time of 3.0 minutes exponentially distributed. When customers arrive to
the system they join a queue to wait for their turn on the ATM. The queue has the
capacity to hold an infinite number of customers. Customers spend an average of
2.4 minutes exponentially distributed at the ATM to complete their transactions,
which is called the service time at the ATM. Simulate the system for the arrival
and processing of 25 customers and estimate the expected waiting time for cus-
tomers in the queue (the average time customers wait in line for the ATM) and the
expected time in the system (the average time customers wait in the queue plus
the average time it takes them to complete their transaction at the ATM).
Using the systems terminology introduced in Chapter 2 to describe the ATM
system, the entities are the customers that arrive to the ATM for processing. The
resource is the ATM that serves the customers, which has the capacity to serve
one customer at a time. The system controls that dictate how, when, and where
activities are performed for this ATM system consist of the queuing discipline,
which is first-in, first-out (FIFO). Under the FIFO queuing discipline the entities
(customers) are processed by the resource (ATM) in the order that they arrive to
har01307_ch03_057-086.indd 74 20/12/10 1:49 PM

FIGURE 3.8
Descriptive drawing of the automatic teller machine (ATM) system.
Arriving customers ATM queue ATM server Departing

(entities) (FIFO) (resource) customers
(entities)
8 7 6 5 4 3 2 1
Interarrival time
4.8 minutes
7th customer 6th customer

arrives at arrives at
21.0 min. 16.2 min.
the queue. Figure 3.8 illustrates the relationships of the elements of the system
with the customers appearing as dark circles.
Our objective in building a spreadsheet simulation model of the ATM system
is to estimate the average amount of time the 25 customers will spend waiting in
the queue and the average amount of time the customers will spend in the sys-
tem. To accomplish our objective, the spreadsheet will need to generate random
numbers and random variates, to contain the logic that controls how customers
are processed by the system, and to compute an estimate of system performance
measures. The spreadsheet simulation is shown in Table 3.2 and is divided into
three main sections: Arrivals to ATM, ATM Processing Time, and ATM Simula-
tion Logic. The Arrivals to ATM and ATM Processing Time provide the founda-
tion for the simulation, while the ATM Simulation Logic contains the spreadsheet
programming that mimics customers flowing through the ATM system. The last
row of the Time in Queue column and the Time in System column under the ATM
Simulation Logic section contains one observation of the average time customers
wait in the queue, 1.94 minutes, and one observation of their average time in the
system, 4.26 minutes. The values were computed by averaging the 25 customer
time values in the respective columns.
Do you think customer number 17 (see the Customer Number column)
became upset over having to wait in the queue for 6.11 minutes to conduct a
0.43 minute transaction at the ATM (see the Time in Queue column and the
Service Time column under ATM Simulation Logic)? Has something like this
ever happened to you in real life? Simulation can realistically mimic the behavior
of a system. More time will be spent interpreting the results of our spreadsheet
simulation after we understand how it was put together.
3.5.1 Simulating Random Variates

The interarrival time is the elapsed time between customer arrivals to the ATM.
This time changes according to the exponential distribution with a mean of
har01307_ch03_057-086.indd 75 20/12/10 1:49 PM

76
TABLE 3.2 Spreadsheet Simulation of Automatic Teller Machine (ATM)
har01307_ch03_057-086.indd 76
Arrivals to ATM ATM Processing Time ATM Simulation Logic
Random Interarrival Random Service Customer Arrival Begin Service Service Departure Time in Time in
Stream 1 Number Time Stream 2 Number Time Number Tone Time Time Time Queue System
i (Z1i) (U1i) (X1i) (Z2i) (U2i) (X2i) (1) (2) (3) (4) (5) (3) (4) (6) (3) (2) (7) (5) (2)
0 3 122
1 66 0.516 2.18 5 0.039 0.10 1 2.18 2.18 0.10 2.28 0.00 0.10
2 109 0.852 5.73 108 0.844 4.46 2 7.91 7.91 4.46 12.37 0.00 4.46
3 116 0.906 7.09 95 0.742 3.25 3 15.00 15.00 3.25 18.25 0.00 3.25
4 7 0.055 0.17 78 0.609 2.25 4 15.17 18.25 2.25 20.50 3.08 5.33
5 22 0.172 0.57 105 0.820 4.12 5 15.74 20.50 4.12 24.62 4.76 8.88
6 81 0.633 3.01 32 0.250 0.69 6 18.75 24.62 0.69 25.31 5.87 6.56
7 40 0.313 1.13 35 0.273 0.77 7 19.88 25.31 0.77 26.08 5.43 6.20
8 75 0.586 2.65 98 0.766 3.49 8 22.53 26.08 3.49 29.57 3.55 7.04
9 42 0.328 1.19 13 0.102 0.26 9 23.72 29.57 0.26 29.83 5.85 6.11
10 117 0.914 7.36 20 0.156 0.41 10 31.08 31.08 0.41 31.49 0.00 0.41
11 28 0.219 0.74 39 0.305 0.87 11 31.82 31.82 0.87 32.69 0.00 0.87
12 79 0.617 2.88 54 0.422 1.32 12 34.70 34.70 1.32 36.02 0.00 1.32
13 126 0.984 12.41 113 0.883 5.15 13 47.11 47.11 5.15 52.26 0.00 5.15
14 89 0.695 3.56 72 0.563 1.99 14 50.67 52.26 1.99 54.25 1.59 3.58
15 80 0.625 2.94 107 0.836 4.34 15 53.61 54.25 4.34 58.59 0.64 4.98
16 19 0.148 0.48 74 0.578 2.07 16 54.09 58.59 2.07 60.66 4.50 6.57
17 18 0.141 0.46 21 0.164 0.43 17 54.55 60.66 0.43 61.09 6.11 6.54
18 125 0.977 11.32 60 0.469 1.52 18 65.87 65.87 1.52 67.39 0.00 1.52
19 68 0.531 2.27 111 0.867 4.84 19 68.14 68.14 4.84 72.98 0.00 4.84
20 23 0.180 0.60 30 0.234 0.64 20 68.74 72.98 0.64 73.62 4.24 4.88
21 102 0.797 4.78 121 0.945 6.96 21 73.52 73.62 6.96 80.58 0.10 7.06
22 97 0.758 4.26 112 0.875 4.99 22 77.78 80.58 4.99 85.57 2.80 7.79
23 120 0.938 8.34 51 0.398 1.22 23 86.12 86.12 1.22 87.34 0.00 1.22
24 91 0.711 3.72 50 0.391 1.19 24 89.84 89.84 1.19 91.03 0.00 1.19
25 122 0.953 9.17 29 0.227 0.62 25 99.01 99.01 0.62 99.63 0.00 0.62
Average 1.94 4.26
20/12/10 1:49 PM
3.0 minutes. That is, the time that elapses between one customer arrival and the
next is not the same from one customer to the next but averages out to be 3.0
minutes. This is illustrated in Figure 3.8 in that the interarrival time between cus-
tomers 7 and 8 is much smaller than the interarrival time of 4.8 minutes between
customers 6 and 7. The service time at the ATM is also a random variable fol-
lowing the exponential distribution and averages 2.4 minutes. To simulate this
stochastic system, a random number generator is needed to produce observations
(random variates) from the exponential distribution. The inverse transformation
method was used in Section 3.4.2 just for this purpose.
The transformation equation is
Xi ln (1 Ui) for i 1, 2, 3, . . . , 25
where Xi represents the ith value realized from the exponential distribution with
mean , and Ui is the ith random number drawn from a uniform(0, 1) distribu-
tion. The i 1, 2, 3, . . . , 25 indicates that we will compute 25 values from the
transformation equation. However, we need to have two different versions of
this equation to generate the two sets of 25 exponentially distributed random
variates needed to simulate 25 customers because the mean interarrival time of
3.0 minutes is different than the mean service time of 2.4 minutes.
Let X1i denote the interarrival time and X2i denote the service time generated
for the ith customer simulated in the system. The equation for transforming
a random number into an interarrival time observation from the exponential
distribution with mean 3.0 minutes becomes
X1i 3.0 ln (1 U1i) for i 1, 2, 3, . . . , 25
where U1i denotes the ith value drawn from the random number generator using
Stream 1. This equation is used in the Arrivals to ATM section of Table 3.2 under
the Interarrival Time (X1i) column.
The equation for transforming a random number into an ATM service time
observation from the exponential distribution with mean 2.4 minutes becomes
X2i 2.4 ln (1 U2i) for i 1, 2, 3, . . . , 25
where U2i denotes the ith value drawn from the random number generator using
Stream 2. This equation is used in the ATM Processing Time section of Table 3.2
under the Service Time (X2i) column.
Let’s produce the sequence of U1i values that feeds the transformation equa-
tion (X1i) for interarrival times using a linear congruential generator (LCG)
similar to the one used in Table 3.1. The equations are
Z1i (21Z1i1 3) mod (128)
U1i Z1i 128 for i 1, 2, 3, . . . , 25
The authors defined Stream 1’s starting or seed value to be 3. So we will use
Z10 3 to kick off this stream of 25 random numbers. These equations are used
in the Arrivals to ATM section of Table 3.2 under the Stream 1 (Z1i) and Random
Number (U1i) columns.
har01307_ch03_057-086.indd 77 20/12/10 1:49 PM

Likewise, we will produce the sequence of U2i values that feeds the transfor-
mation equation (X2i) for service times using
Z2i (21Z2i1 3) mod (128)
U2i Z2i 128 for i 1, 2, 3, . . . , 25
and will specify a starting seed value of Z20 122, Stream 2’s seed value, to kick
off the second stream of 25 random numbers. These equations are used in the
ATM Processing Time section of Table 3.2 under the Stream 2 (Z2i) and Random
Number (U2i) columns.
The spreadsheet presented in Table 3.2 illustrates 25 random variates for both
the interarrival time, column (X1i), and service time, column (X2i). All time values
are given in minutes in Table 3.2. To be sure we pull this together correctly, let’s
compute a couple of interarrival times with mean 3.0 minutes and compare
them to the values given in Table 3.2.
Given Z10 3
Z11 (21Z10 3) mod (128) (21(3) 3) mod (128)
(66) mod (128) 66
U11 Z11128 66128 0.516
X11 ln (1 U11) 3.0 ln (1 0.516) 2.18 minutes
The value of 2.18 minutes is the first value appearing under the column,
Interarrival Time (X1i) To compute the next interarrival time value X12, we start
by using the value of Z11 to compute Z12.
Given Z11 66
Z12 (21Z11 3) mod (128) (21(66) 3) mod (128) 109
U12 Z12 128 109128 0.852
X12 3 ln (1 U12) 3.0 ln (1 0.852) 5.73 minutes
Figure 3.9 illustrates how the equations were programmed in Microsoft Excel for
the Arrivals to ATM section of the spreadsheet. Note that the U1i and X1i values
in Table 3.2 are rounded to three and two places to the right of the decimal,
respectively. The same rounding rule is used for U2i and X2i.
It would be useful for you to verify a few of the service time values with
mean 2.4 minutes appearing in Table 3.2 using
Z20 122
Z2i (21Z2i1 3) mod (128)
U2i Z2i 128
X2i 2.4 ln (1 U2i) for i 1, 2, 3, . . .
The equations started out looking a little difficult to manipulate but turned
out not to be so bad when we put some numbers in them and organized them in a
har01307_ch03_057-086.indd 78 20/12/10 1:49 PM

FIGURE 3.9
Microsoft Excel
snapshot of the ATM
spreadsheet
illustrating the
equations for the
Arrivals to ATM
section.
spreadsheet—though it was a bit tedious. The important thing to note here is that al-
though it is transparent to the user, ProModel uses a very similar method to produce
exponentially distributed random variates, and you now understand how it is done.
The LCG just given has a maximum cycle length of 128 random numbers
(you may want to verify this), which is more than enough to generate 25 interar-
rival time values and 25 service time values for this simulation. However, it is a
poor random number generator compared to the one used by ProModel. It was
chosen because it is easy to program into a spreadsheet and to compute by hand
to facilitate our understanding. The biggest difference between it and the random
number generator in ProModel is that the ProModel random number generator
manipulates much larger numbers to pump out a much longer stream of numbers
that pass all statistical tests for randomness.
Before moving on, let’s take a look at why we chose Z10 3 and Z20 122.
Our goal was to make sure that we did not use the same uniform(0, 1) random
number to generate both an interarrival time and a service time. If you look
carefully at Table 3.2, you will notice that the seed value Z20 122 is the Z125
value from random number Stream 1. Stream 2 was merely defined to start where
Stream 1 ended. Thus our spreadsheet used a unique random number to generate
each interarrival and service time. Now let’s add the necessary logic to our
spreadsheet to conduct the simulation of the ATM system.
3.5.2 Simulating Dynamic, Stochastic Systems

The heart of the simulation is the generation of the random variates that drive the
stochastic events in the simulation. However, the random variates are simply a list
of numbers at this point. The spreadsheet section labeled ATM Simulation Logic
in Table 3.2 is programmed to coordinate the execution of the events to mimic the
har01307_ch03_057-086.indd 79 20/12/10 1:49 PM

processing of customers through the ATM system. The simulation program must
keep up with the passage of time to coordinate the events. The word dynamic
appearing in the title for this section refers to the fact that the simulation tracks
the passage of time.
The first column under the ATM Simulation Logic section of Table 3.2,
labeled Customer Number, is simply to keep a record of the order in which the
25 customers are processed, which is FIFO. The numbers appearing in parentheses
under each column heading are used to illustrate how different columns are added
or subtracted to compute the values appearing in the simulation.
The Arrival Time column denotes the moment in time at which each customer
arrives to the ATM system. The first customer arrives to the system at time 2.18 min-
utes. This is the first interarrival time value (X11 2.18) appearing in the Arrival to
ATM section of the spreadsheet table. The second customer arrives to the system at
time 7.91 minutes. This is computed by taking the arrival time of the first customer
of 2.18 minutes and adding to it the next interarrival time of X12 5.73 minutes that
was generated in the Arrival to ATM section of the spreadsheet table. The arrival
time of the second customer is 2.18 5.73 7.91 minutes. The process continues
with the third customer arriving at 7.91 7.09 15.00 minutes.
The trickiest piece to program into the spreadsheet is the part that determines
the moment in time at which customers begin service at the ATM after waiting in
the queue. Therefore, we will skip over the Begin Service Time column for now
and come back to it after we understand how the other columns are computed.
The Service Time column simply records the simulated amount of time re-
quired for the customer to complete their transaction at the ATM. These values
are copies of the service time X2i values generated in the ATM Processing Time
section of the spreadsheet.
The Departure Time column records the moment in time at which a customer
departs the system after completing their transaction at the ATM. To compute the
time at which a customer departs the system, we take the time at which the cus-
tomer gained access to the ATM to begin service, column (3), and add to that the
length of time the service required, column (4). For example, the first customer
gained access to the ATM to begin service at 2.18 minutes, column (3). The ser-
vice time for the customer was determined to be 0.10 minutes in column (4). So,
the customer departs 0.10 minutes later or at time 2.18 0.10 2.28 minutes.
This customer’s short service time must be because they forgot their PIN number
and could not conduct their transaction.
The Time in Queue column records how long a customer waits in the queue
before gaining access to the ATM. To compute the time spent in the queue, we
take the time at which the ATM began serving the customer, column (3), and sub-
tract from that the time at which the customer arrived to the system, column (2).
The fourth customer arrives to the system at time 15.17 and begins getting service
from the ATM at 18.25 minutes; thus, the fourth customer’s time in the queue is
18.25 15.17 3.08 minutes.
The Time in System column records how long a customer was in the system.
To compute the time spent in the system, we subtract the customer’s departure
time, column (5), from the customer’s arrival time, column (2). The fifth customer
har01307_ch03_057-086.indd 80 20/12/10 1:49 PM

arrives to the system at 15.74 minutes and departs the system at 24.62 minutes.
Therefore, this customer spent 24.62 15.74 8.88 minutes in the system.
Now let’s go back to the Begin Service Time column, which records the time
at which a customer begins to be served by the ATM. The very first customer
to arrive to the system when it opens for service advances directly to the ATM.
There is no waiting time in the queue; thus the value recorded for the time that
the first customer begins service at the ATM is the customer’s arrival time. With
the exception of the first customer to arrive to the system, we have to capture the
logic that a customer cannot begin service at the ATM until the previous customer
using the ATM completes his or her transaction. One way to do this is with an IF
statement as follows:
IF (Current Customer’s Arrival Time < Previous Customer’s
Departure Time)
THEN (Current Customer’s Begin Service Time Previous Customer’s
Departure Time)
ELSE (Current Customer’s Begin Service Time Current Customer’s
Arrival Time)
Figure 3.10 illustrates how the IF statement was programmed in Microsoft

Excel. The format of the Excel IF statement is
IF (logical test, use this value if test is true, else use this
value if test is false)
The Excel spreadsheet cell L10 (column L, row 10) in Figure 3.10 is the Begin
Service Time for the second customer to arrive to the system and is programmed
with IF(K10<N9,N9,K10). Since the second customer’s arrival time (Excel cell
K10) is not less than the first customer’s departure time (Excel cell N9), the logical
test evaluates to “false” and the second customer’s time to begin service is set to
his or her arrival time (Excel cell K10). The fourth customer shown in Figure 3.10
provides an example of when the logical test evaluates to “true,” which results in
the fourth customer beginning service when the third customer departs the ATM.
FIGURE 3.10
Microsoft Excel
snapshot of the
ATM spreadsheet
illustrating the IF
statement for the
Begin Service Time
column.
har01307_ch03_057-086.indd 81 20/12/10 1:49 PM

3.5.3 Simulation Replications and Output Analysis

The spreadsheet model makes a nice simulation of the first 25 customers to arrive
to the ATM system. And we have a simulation output observation of 1.94 minutes
that could be used as an estimate for the average time that the 25 customers waited
in the queue (see the last value under the Time in Queue column, which represents
the average of the 25 individual customer queue times). We also have a simulation
output observation of 4.26 minutes that could be used as an estimate for the aver-
age time the 25 customers were in the system. These results represent only one
possible value of each performance measure. Why? If at least one input variable
to the simulation model is random, then the output of the simulation model will
also be random. The interarrival times of customers to the ATM system and their
service times are random variables. Thus the output of the simulation model of
the ATM system is also a random variable. Before we bet that the average time
customers spend in the system each day is 4.26 minutes, we may want to run the
simulation model again with a new sample of interarrival times and service times
to see what happens. This would be analogous to going to the real ATM on a sec-
ond day to replicate our observing the first 25 customers processed to compute a
second observation of the average time that the 25 customers spend in the system.
In simulation, we can accomplish this by changing the values of the seed numbers
used for our random number generator.
Changing the seed values Z10 and Z20 causes the spreadsheet program to re-
compute all values in the spreadsheet. When we change the seed values Z10 and
Z20 appropriately, we produce another replication of the simulation. When we run
replications of the simulation, we are driving the simulation with a set of random
numbers never used by the simulation before. This is analogous to the fact that the
arrival pattern of customers to the real ATM and their transaction times at the ATM
will not likely be the same from one day to the next. If Z10 29 and Z20 92 are
used to start the random number generator for the ATM simulation model, a new rep-
lication of the simulation will be computed that produces an average time in queue
of 0.84 minutes and an average time in system of 2.36 minutes. Review question
number eight at the end of the chapter asks you to verify this second replication.
Table 3.3 contains the results from the two replications. Obviously, the
results from this second replication are very different than those produced by
the first replication. Good thing we did not make bets on how much time cus-
tomers spend in the system based on the output of the simulation for the first
TABLE 3.3 Summary of ATM System Simulation Output
Replication Average Time in Queue Average Time in System
1 1.94 minutes 4.26 minutes

2 0.84 minutes 2.36 minutes
Average 1.39 minutes 3.31 minutes
har01307_ch03_057-086.indd 82 20/12/10 1:49 PM

replication only. Statistically speaking, we should get a better estimate of the

average time customers spend in queue and the average time they are in the
system if we combine the results from the two replications into an overall aver-
age. Doing so yields an estimate of 1.39 minutes for the average time in queue
and an estimate of 3.31 minutes for the average time in system (see Table 3.3).
However, the large variation in the output of the two simulation replications in-
dicates that more replications are needed to get reasonably accurate estimates.
How many replications are enough? You will know how to answer the question
upon completing Chapter 8.
While spreadsheet technology is effective for conducting Monte Carlo simu-
lations and simulations of simple dynamic systems, it is ineffective and inefficient
as a simulation tool for complex systems. Discrete-event simulation software
technology was designed especially for mimicking the behavior of complex sys-
tems and is the subject of Chapter 4.
3.6 Summary
Modeling random behavior begins with transforming the output produced by a
random number generator into observations (random variates) from an appropri-
ate statistical distribution. The values of the random variates are combined with
logical operators in a computer program to compute output that mimics the per-
formance behavior of stochastic systems. Performance estimates for stochastic
simulations are obtained by calculating the average value of the performance met-
ric across several replications of the simulation. Models can realistically simulate
a variety of systems.

1. What is the difference between a stochastic model and a deterministic
model in terms of the input variables and the way results are
interpreted?
2. Give an example of a discrete-change state variable and a continuous-
change state variable.
3. For each of the following simulation applications identify one discrete-
and one continuous-change state variable.
a. Inventory control of an oil storage and pumping facility.
b. Study of beverage production in a soft drink production facility.
c. Study of congestion in a busy traffic intersection.
4. Give a statistical description of the numbers produced by a random
number generator.
5. What are the two statistical properties that random numbers must
satisfy?
har01307_ch03_057-086.indd 83 20/12/10 1:49 PM

6. Given these two LCGs:

Zi (9Zi1 3) mod (32)
Zi (12Zi1 5) mod (32)
a. Which LCG will achieve its maximum cycle length? Answer the
question without computing any Zi values.
b. Compute Z1 through Z5 from a seed of 29 (Z0 29) for the
second LCG.
7. What is a random variate, and how are random variates generated?
8. Apply the inverse transformation method to generate three variates from
the following distributions using U1 0.10, U2 0.53, and U3 0.15.
a. Probability density function:

1
______
for x
f (x)
0 elsewhere
where 7 and 4.
b. Probability mass function:

x
___
for x 1, 2, 3, 4, 5
p (x) P (X x) 15
0 elsewhere
9. How would a random number generator be used to simulate a 12 percent

chance of rejecting a part because it is defective?
10. Reproduce the spreadsheet simulation of the ATM system presented
in Section 3.5. Set the random numbers seeds Z10 29 and Z20 92
to compute the average time customers spend in the queue and in the
system.
a. Verify that the average time customers spend in the queue and in the
system match the values given for the second replication in Table 3.3.
b. Verify that the resulting random number Stream 1 and random
number Stream 2 are completely different than the corresponding
streams in Table 3.2. Is this a requirement for a new replication of
the simulation?
11. Using the spreadsheet simulation from problem 10 above, replace the
exponentially distributed interarrival time with mean 3.0 minutes with
an interarrival time that is uniformly distributed (continuous case) with
mean 3.0 minutes by setting the uniform distribution’s minimum value
to 1.0 minute and the maximum value to 5.0 minutes. How did the
change influence the average time in queue and average time in system
as compared to the second replication results shown in Table 3.3, which
are based on the exponentially distributed interarrival time with mean
3.0 minutes?
har01307_ch03_057-086.indd 84 20/12/10 1:49 PM

12. Given Ut ⫽ 0.23 and Uz ⫽ 0.61 from a uniform(0, 1) distribution,

generate two random variates from each of the distributions following.
a. Uniform(12, 20) continuous case
b. Triangular(12, 16, 20)
c. Normal(16, 10)
d. Uniform(12, 20) discrete case
e. Bernoulli(0.50)
References
Banks, Jerry; John S. Carson II; Barry L. Nelson; and David M. Nicol. Discrete-Event
System Simulation. Englewood Cliffs, NJ: Prentice Hall, 2001.
Hoover, Stewart V., and Ronald F. Perry. Simulation: A Problem-Solving Approach.
Reading, MA: Addison-Wesley, 1989.
Johnson, R. A. Miller and Freund’s Probability and Statistics for Engineers. 8th ed.
Englewood Cliffs, NJ: Prentice Hall, 2010.
Law, Averill M. Simulation Modeling and Analysis 4th ed. New York: McGraw-Hill, 2007.
L’Ecuyer, P. “Random Number Generation.” In Handbook of Simulation: Principles,
Methodology, Advances, Applications, and Practice, ed. J. Banks, pp. 93–137. New
York: John Wiley & Sons, 1998.
Pooch, Udo W., and James A. Wall. Discrete Event Simulation: A Practical Approach.
Boca Raton, FL: CRC Press, 1993.
Pritsker, A. A. B. Introduction to Simulation and SLAM II. 4th ed. New York: John Wiley
& Sons, 1995.
Ross, Sheldon M.A. Simulation. 4th ed. Burlington, MA: Elsevier Academic Press, 2006.
Shannon, Robert E. System Simulation: The Art and Science. Englewood Cliffs, NJ:
Prentice Hall, 1975.
Thesen, Arne, and Laurel E. Travis. Simulation for Decision Making. Minneapolis, MN:
West Publishing, 1992.
Widman, Lawrence E.; Kenneth A. Loparo; and Norman R. Nielsen. Artificial Intelligence,
Simulation, and Modeling. New York: John Wiley & Sons, 1989.
har01307_ch03_057-086.indd 85 12/22/10 7:15 AM

har01307_ch03_057-086.indd 86 20/12/10 1:49 PM
C H A P T E R
4 DISCRETE-EVENT
SIMULATION
“When the only tool you have is a hammer, every problem begins to resemble a
nail.”
—Abraham Maslow
4.1 Introduction
Building on the foundation provided by Chapter 3 on how random numbers and
random variates are used to simulate stochastic systems, the focus of this chapter
is on discrete-event simulation, which is the main topic of this book. A discrete-
event simulation is one in which changes in the state of the simulation model
occur at discrete points in time as triggered by events. The events in the automatic
teller machine (ATM) simulation example of Chapter 3 that occur at discrete
points in time are the arrivals of customers to the ATM queue and the comple-
tion of their transactions at the ATM. However, you will learn in this chapter that
the spreadsheet simulation of the ATM system in Chapter 3 was not technically
executed as a discrete-event simulation.
This chapter summarizes the basic technical issues related to discrete-event
simulation to facilitate your understanding of how to effectively use the tool.
Questions that will be answered include these:
• How does discrete-event simulation work?
• What do commercial simulation software packages provide?
• What are the differences between simulation languages and simulators?
• What is the future of simulation technology?
A manual dynamic, stochastic, discrete-event simulation of the ATM example
system from Chapter 3 is given to further illustrate what goes on inside this type
of simulation.
87
har01307_ch04_087-114.indd 87 20/12/10 1:49 PM

4.2 How Discrete-Event Simulation Works

Most simulation software, including ProModel, presents a process-oriented world
view to the user for defining models. This is the way most humans tend to think
about systems that process entities. When describing a system, it is natural to do
so in terms of the process flow. Entities begin processing at activity A then move
on to activity B and so on. In discrete-event simulation, these process flow defi-
nitions are translated into a sequence of events for running the simulation: first
event 1 happens (an entity begins processing at activity A), then event 2 occurs
(it completes processing at activity A), and so on. Events in simulation are of two
types: scheduled and conditional, both of which create the activity delays in the
simulation to replicate the passage of time.
A scheduled event is one whose time of occurrence can be determined be-
forehand and can therefore be scheduled in advance. Assume, for example, that
an activity has just begun that will take X amount of time, where X is a normally
distributed random variable with a mean of 5 minutes and a standard deviation
of 1.2 minutes. At the start of the activity, a normal random variate is generated
based on these parameters, say 4.2 minutes, and an activity completion event is
scheduled for that time into the future. Scheduled events are inserted chrono-
logically into an event calendar to await the time of their occurrence. Events that
occur at predefined intervals theoretically all could be determined in advance and
therefore be scheduled at the beginning of the simulation. For example, entities
arriving every five minutes into the model could all be scheduled easily at the start
of the simulation. Rather than preschedule all events at once that occur at a set
frequency, however, they are scheduled only when the next occurrence must be
determined. In the case of a periodic arrival, the next arrival would not be sched-
uled until the current scheduled arrival is actually pulled from the event calendar
for processing. This postponement until the latest possible moment minimizes the
size of the event calendar and eliminates the necessity of knowing in advance how
many events to schedule when the length of the simulation may be unknown.
Conditional events are triggered by a condition being met rather than by the
passage of time. An example of a conditional event might be the capturing of
a resource that is predicated on the resource being available. Another example
would be an order waiting for all of the individual items making up the order to
be assembled. In these situations, the event time cannot be known beforehand, so
the pending event is simply placed into a waiting list until the conditions can be
satisfied. Often multiple pending events in a list are waiting for the same condi-
tion. For example, multiple entities might be waiting to use the same resource
when it becomes available. Internally, the resource would have a waiting list for
all items currently waiting to use it. While in most cases events in a waiting list
are processed first-in, first-out (FIFO), items could be inserted and removed using
a number of different criteria. For example, items may be inserted according to
item priority but be removed according to earliest due date.
Events, whether scheduled or conditional, trigger the execution of logic that
is associated with that event. For example, when an entity frees a resource, the
har01307_ch04_087-114.indd 88 20/12/10 1:49 PM

Chapter 4 Discrete-Event Simulation 89
state and statistical variables for the resource are updated, the graphical anima-
tion is updated, and the input waiting list for the resource is examined to see what
activity to respond to next. Any new events resulting from the processing of the
current event are inserted into either the event calendar or another appropriate
waiting list.
In real life, events can occur simultaneously so that multiple entities can be
doing things at the same instant in time. In computer simulation, however, espe-
cially when running on a single processor, events can be processed only one at
a time even though it is the same instant in simulated time. As a consequence, a
method or rule must be established for processing events that occur at the exact
same simulated time. For some special cases, the order in which events are pro-
cessed at the current simulation time might be significant. For example, an entity
that frees a resource and then tries to immediately get the same resource might
have an unfair advantage over other entities that might have been waiting for that
particular resource.
In ProModel, the entity, downtime, or other item that is currently being pro-
cessed is allowed to continue processing as far as it can at the current simulation
time. That means it continues processing until it reaches either a conditional event
that cannot be satisfied or a timed delay that causes a future event to be scheduled.
It is also possible that the object simply finishes all of the processing defined for
it and, in the case of an entity, exits the system. As an object is being processed,
any resources that are freed or other entities that might have been created as by-
products are placed in an action list and are processed one at a time in a similar
fashion after the current object reaches a stopping point. To deliberately suspend
the current object in order to allow items in the action list to be processed, a zero
delay time can be specified for the current object. This puts the current item into
the future events list (event calendar) for later processing, even though it is still
processed at the current simulation time.
When all scheduled and conditional events have been processed that are
possible at the current simulation time, the clock advances to the next scheduled
event and the process continues. When a termination event occurs, the simula-
tion ends and statistical reports are generated. The ongoing cycle of processing
scheduled and conditional events, updating state and statistical variables, and
creating new events constitutes the essence of discrete-event simulation (see
Figure 4.1).
4.3 A Manual Discrete-Event Simulation Example

To illustrate how discrete-event simulation works, a manual simulation is pre-
sented of the automatic teller machine (ATM) system of Chapter 3. Customers
arrive to use an automatic teller machine (ATM) at a mean interarrival time of
3.0 minutes exponentially distributed. When customers arrive to the system, they
join a queue to wait for their turn on the ATM. The queue has the capacity to hold
an infinite number of customers. The ATM itself has a capacity of one (only one
har01307_ch04_087-114.indd 89 20/12/10 1:49 PM

FIGURE 4.1
Start
Logic diagram of
how discrete-event
simulation works.
Create simulation
database and schedule
initial events.
Advance clock to
next event time.
Yes Update statistics

Termination and generate
event? output report.
No
Stop
Process event and
schedule any new
events.
Update statistics,
state variables,
and animation.
Yes Any conditional

events?
No
customer at a time can be processed at the ATM). Customers spend an average of

2.4 minutes exponentially distributed at the ATM to complete their transactions,
which is called the service time at the ATM. Assuming that the simulation starts
at time zero, simulate the ATM system for its first 22 minutes of operation and
estimate the expected waiting time for customers in the queue (the average time
har01307_ch04_087-114.indd 90 20/12/10 1:49 PM

FIGURE 4.2
Entity flow diagram for example automatic teller machine (ATM) system.
Arriving customers ATM queue ATM server Departing

(entities) (FIFO) (resource) customers
(entities)
8 7 6 5 4 3 2 1
customers wait in line for the ATM) and the expected number of customers wait-
ing in the queue (the average number of customers waiting in the queue during
the simulated time period). If you are wondering why 22 minutes was selected
as the simulation run length, it is because 22 minutes of manual simulation will
nicely fit on one page of the textbook. An entity flow diagram of the ATM system
is shown in Figure 4.2.
4.3.1 Simulation Model Assumptions

We did not list our assumptions for simulating the ATM system back in Chapter 3
although it would have been a good idea to have done so. In any simulation,
certain assumptions must be made where information is not clear or complete.
Therefore, it is important that assumptions be clearly documented. The assump-
tions we will be making for this simulation are as follows:
• There are no customers in the system initially, so the queue is empty and
the ATM is idle.
• The move time from the queue to the ATM is negligible and therefore
ignored.
• Customers are processed from the queue on a first-in, first-out (FIFO) basis.
• The ATM never experiences failures.
4.3.2 Setting Up the Simulation

A discrete-event simulation does not generate all of the events in advance as was
done in the spreadsheet simulation of Chapter 3. Instead the simulation schedules
the arrival of customer entities to the ATM queue and the departure of customer
entities from the ATM as they occur over time. That is, the simulation calls a
function to obtain an exponentially distributed interarrival time value only when a
new customer entity needs to be scheduled to arrive to the ATM queue. Likewise,
the simulation calls a function to obtain an exponentially distributed service time
value at the ATM only after a customer entity has gained access (entry) to the
ATM and the entity’s future departure time from the ATM is needed. An event
har01307_ch04_087-114.indd 91 20/12/10 1:49 PM

calendar is maintained to coordinate the processing of the customer arrival and

customer departure events. As arrival and departure event times are created, they
are placed on the event calendar in chronological order. As these events get pro-
cessed, state variables (like the contents of a queue) get changed and statistical
accumulators (like the cumulative time in the queue) are updated.
With this overview, let’s list the structured components needed to conduct the
manual simulation.
Simulation Clock
As the simulation transitions from one discrete-event to the next, the simulation
clock is fast forwarded to the time that the next event is scheduled to occur. There
is no need for the clock to tick seconds away until reaching the time at which
the next event in the list is scheduled to occur because nothing will happen that
changes the state of the system until the next event occurs. Instead, the simula-
tion clock advances through a series of time steps. Let ti denote the value of the
simulation clock at time step i, for i 0 to the number of discrete events to
process. Assuming that the simulation starts at time zero, then the initial value of
the simulation clock is denoted as t0 0. Using this nomenclature, t1 denotes the
value of the simulation clock when the first discrete event in the list is processed,
t2 denotes the value of the simulation clock when the second discrete-event in the
list is processed, and so on.
Entity Attributes
To capture some statistics about the entities being processed through the system,
a discrete-event simulation maintains an array of entity attribute values. Entity
attributes are characteristics of the entity that are maintained for that entity until
the entity exits the system. For example, to compute the amount of time an entity
waited in a queue location, an attribute is needed to remember when the entity
entered the location. For the ATM simulation, one entity attribute is used to re-
member the customer’s time of arrival to the system. This entity attribute is called
the Arrival Time attribute. The simulation program computes how long each cus-
tomer entity waited in the queue by subtracting the time that the customer entity
arrived to the queue from the value of the simulation clock when the customer
entity gained access to the ATM.
State Variables
Two discrete-change state variables are needed to track how the status (state) of
the system changes as customer entities arrive in and depart from the ATM system.
• Number of Entities in Queue at time step i, NQi.
• ATM Statusi to denote if the ATM is busy or idle at time step i.
Statistical Accumulators
The objective of the example manual simulation is to estimate the expected
amount of time customers wait in the queue and the expected number of customers
har01307_ch04_087-114.indd 92 20/12/10 1:49 PM

waiting in the queue. The average time customers wait in queue is a simple aver-
age. Computing this requires that we record how many customers passed through
the queue and the amount of time each customer waited in the queue. The average
number of customers in the queue is a time-weighted average, which is usually
called a time average in simulation. Computing this requires that we not only
observe the queue’s contents during the simulation but that we also measure the
amount of time that the queue maintained each of the observed values. We record
each observed value after it has been multiplied (weighted) by the amount of time
it was maintained.
Here’s what the simulation needs to tally at each simulation time step i to
compute the two performance measures at the end of the simulation.
Simple-Average Time in Queue.
• Record the number of customer entities processed through the queue,
Total Processed. Note that the simulation may end before all customer
entities in the queue get a turn at the ATM. This accumulator keeps track
of how many customers actually made it through the queue.
• For a customer processed through the queue, record the time that it waited
in the queue. This is computed by subtracting the value of the simulation
clock time when the entity enters the queue (stored in the entity attribute
array Arrival Time) from the value of the simulation clock time when the
entity leaves the queue, ti – Arrival Time.
Time-Average Number of Customers in the Queue.
• For the duration of the last time step, which is ti – ti1, and the number of
customer entities in the queue during the last time step, which is NQi1,
record the product of ti – ti1 and NQi1. Call the product (ti – ti1) NQi1
the Time-Weighted Number of Entities in the Queue.
Events
There are two types of recurring scheduled events that change the state of the sys-
tem: arrival events and departure events. An arrival event occurs when a customer
entity arrives to the queue. A departure event occurs when a customer entity com-
pletes its transaction at the ATM. Each processing of a customer entity’s arrival
to the queue includes scheduling the future arrival of the next customer entity to
the ATM queue. Each time an entity gains access to the ATM, its future departure
from the system is scheduled based on its expected service time at the ATM. We
actually need a third event to end the simulation. This event is usually called the
termination event.
To schedule the time at which the next entity arrives to the system, the simu-
lation needs to generate an interarrival time and add it to the current simulation
clock time, ti. The interarrival time is exponentially distributed with a mean of
3.0 minutes for our example ATM system. Assume that the function E(3.0) re-
turns an exponentially distributed random variate with a mean of 3.0 minutes.
The future arrival time of the next customer entity can then be scheduled by
using the equation ti E(3.0).
har01307_ch04_087-114.indd 93 20/12/10 1:49 PM

The customer service time at the ATM is exponentially distributed with a

mean of 2.4 minutes. The future departure time of an entity gaining access to the
ATM is scheduled by the equation ti E(2.4).
Event Calendar
The event calendar maintains the list of active events (events that have been
scheduled and are waiting to be processed) in chronological order. The simulation
progresses by removing the first event listed on the event calendar, setting the
simulation clock, ti, equal to the time at which the event is scheduled to occur, and
processing the event.
4.3.3 Running the Simulation

To run the manual simulation, we need a way to generate exponentially distributed
random variates for interarrival times [the function E(3.0)] and service times [the
function E(2.4)]. Rather than using our calculators to generate a random number
and transform it into an exponentially distributed random variate each time one
is needed for the simulation, let’s take advantage of the work we did to build the
spreadsheet simulation of the ATM system in Chapter 3. In Table 3.2 of Chapter 3,
we generated 25 exponentially distributed interarrival times with a mean of
3.0 minutes and 25 exponentially distributed service times with a mean of 2.4 min-
utes in advance of running the spreadsheet simulation. So when we need a service
time or an interarrival time for our manual simulation, let’s use the values from the
Service Time and Interarrival Time columns of Table 3.2 rather than computing
them by hand. Note, however, that we do not need to generate all random variates
in advance for our manual discrete-event simulation (that’s one of its advantages
over spreadsheet simulation). We are just using Table 3.2 to make our manual sim-
ulation a little less tedious, and it will be instructive to compare the results of the
manual simulation with those produced by the spreadsheet simulation.
Notice that Table 3.2 contains a subscript i in the leftmost column. This sub-
script denotes the customer entity number as opposed to the simulation time step.
We wanted to point this out to avoid any confusion because of the different uses of
the subscript. In fact, you can ignore the subscript in Table 3.2 as you pick values
from the Service Time and Interarrival Time columns.
A discrete-event simulation logic diagram for the ATM system is shown in
Figure 4.3 to help us carry out the manual simulation. Table 4.1 presents the re-
sults of the manual simulation after processing 12 events using the simulation
logic diagram presented in Figure 4.3. The table tracks the creation and schedul-
ing of events on the event calendar as well as how the state of the system changes
and how the values of the statistical accumulators change as events are processed
from the event calendar. Although Table 4.1 is completely filled in, it was initially
blank until the instructions presented in the simulation logic diagram were ex-
ecuted. As you work through the simulation logic diagram, you should process the
information in Table 4.1 from the first row down to the last row, one row at a time
(completely filling in a row before going down to the next row). A dash (—) in a
cell in Table 4.1 signifies that the simulation logic diagram does not require you
to update that particular cell at the current simulation time step. An arrow (↑) in a
har01307_ch04_087-114.indd 94 20/12/10 1:49 PM

Start
i=0
Initialize variables and schedule initial arrival event and termination event (Scheduled Future Events).
har01307_ch04_087-114.indd 95
i=i+1
i=i+1 • Update Event Calendar: Insert Scheduled Future Events in chronological order. i=i+1
• Advance Clock, ti, to the Time of the first event on the calendar and process the event.
Schedule arrival event

Arrive Event Depart
for next customer entity
to occur at time ti + E(3). type?
End
Update statistics and generate output report. Any

Yes Is ATM No Yes customers No
idle? in queue?
End
• Schedule departure event for • Store current customer • Update Entity Attribute Array by • Update Entity Attribute
current customer entity entering entity’s Arrival Time in last deleting departed customer entity Array by deleting departed
ATM to occur at time ti + E(2.4). position of Entity Attribute from first position in the array and customer entity from first
• Store current customer’s Array to reflect customer shifting waiting customers up. position of the array.
Arrival Time in first position of joining the queue. • Subtract 1 from NQi, Number of • Change ATM Statusi to Idle.
Entity Attribute Array. • Add 1 to NQi, Number of Entities in Queue.
• Change ATM Statusi to Busy. Entities in Queue. • Schedule departure event for
• Update Entities Processed • Update Time-Weighted customer entity entering the ATM
through Queue statistics to Number of Entities in Queue to occur at time ti + E(2.4).
reflect customer entering ATM. statistic. • Update Entities Processed
- Add 1 to Total Processed. - Compute value for through Queue statistics to reflect
- Record Time in Queue of 0 for (ti – ti – 1)NQi – 1 and customer entering ATM.
ti - Arrival Time and update update Cumulative. - Add 1 to Total Processed.
Cumulative. - Compute Time in Queue
• Update Time-Weighted Number value for ti - Arrival Time and
of Entities in Queue statistic. update Cumulative.
- Record value of 0 for • Update Time-Weighted Number of
(ti – ti – 1)NQi – 1 and update Entities in Queue statistic.
Cumulative. - Compute value for (ti – ti – 1)NQi –1
and update Cumulative.
95
FIGURE 4.3
Discrete-event simulation logic diagram for ATM system.
20/12/10 1:49 PM
96
TABLE 4.1 Manual Discrete-Event Simulation of ATM System
Event Calendar Processed Event System State Statistical Accumulators Scheduled Future Events
Entity Attribute Time-Weighted
Array (Entity Number, Entities Processed Number of
har01307_ch04_087-114.indd 96
Arrival Time) through Queue Entities in Queue
(Entity Number, Event, Time)

Clock, ti
Entity Number
Event
*Entity Using ATM,
array position 1
Entities Waiting in
Queue, array
positions 2, 3, . . .
Number of Entities
in Queue, NQi
ATM Statusi
Total Processed
Time in Queue,
ti Arrival Time
Cumulative
兺 (ti Arrival Time)
(ti ti1)NQi1
Cumulative
兺 (ti ti1)NQi1
(Entity Number, Event, Time)
i
0 — 0 — — *( ) 0 Idle 0 — 0 — 0 (1, Arrive, 2.18)
( ) (_, End, 22.00)
1 (1, Arrive, 2.18) 2.18 1 Arrive *(1, 2.18) ↑ Busy 1 0 0 0 (2, Arrival, 7.91)
(_, End, 22.00) ( ) (1, Depart, 2.28)
2 (1, Depart, 2.28) 2.28 1 Depart *( ) ↑ Idle — — 0 — 0
(2, Arrive, 7.91) ( ) No new events
(_, End, 22.00)
3 (2, Arrive, 7.91) 7.91 2 Arrive *(2, 7.91) ↑ Busy 2 0 0 0 0 (3, Arrive, 15.00)
(_, End, 22.00) ( ) (2, Depart, 12.37)
4 (2, Depart, 12.37) 12.37 2 Depart *( ) ↑ Idle — — 0 — 0
(3, Arrive, 15.00) ( ) No new events
(_, End, 22.00)
5 (3, Arrive, 15.00) 15.00 3 Arrive *(3, 15.00) ↑ Busy 3 0 0 0 0 (4, Arrive, 15.17)
(_, End, 22.00) ( ) (3, Depart, 18.25)
6 (4, Arrive, 15.17) 15.17 4 Arrive *(3, 15.00) 1 ↑ — — 0 0 0 (5, Arrive, 15.74)
(3, Depart, 18.25) (4, 15.17)
(_, End, 22.00)
7 (5, Arrive, 15.74) 15.74 5 Arrive *(3, 15.00) 2 ↑ — — 0 0.57 0.57 (6, Arrive, 18.75)
(3, Depart, 18.25) (4, 15.17)
(_, End, 22.00) (5, 15.74)
8 (3, Depart, 18.25) 18.25 3 Depart *(4, 15.17) 1 ↑ 4 3.08 3.08 5.02 5.59 (4, Depart, 20.50)
(6, Arrive, 18.75) (5, 15.74)
(_, End, 22.00)
9 (6, Arrive, 18.75) 18.75 6 Arrive *(4, 15.17) 2 ↑ — — 3.08 0.50 6.09 (7, Arrive, 19.88)
(4, Depart, 20.50) (5, 15.74)
(_, End, 22.00) (6, 18.75)
10 (7, Arrive, 19.88) 19.88 7 Arrive *(4, 15.17) 3 ↑ — — 3.08 2.26 8.35 (8, Arrive, 22.53)
(4, Depart, 20.50) (5, 15.74)
(_, End, 22.00) (6, 18.75)
(7, 19.88)
11 (4, Depart, 20.50) 20.50 4 Depart *(5, 15.74) 2 ↑ 5 4.76 7.84 1.86 10.21 (5, Depart, 24.62)
(_, End, 22.00) (6, 18.75)
(8, Arrive, 22.53) (7, 19.88)
12 (_, End, 22.00) 22.00 End — ↑ ↑ 5 — 7.84 3.00 13.21 —
(8, Arrive, 22.53)
(5, Depart, 24.62)
20/12/10 1:49 PM
cell in the table also signifies that the simulation logic diagram does not require
you to update that cell at the current time step. However, the arrows serve as a
reminder to look up one or more rows above your current position in the table to
determine the state of the ATM system. Arrows appear under the Number of Enti-
ties in Queue, NQi column, and ATM Statusi column. The only exception to the
use of dashes or arrows is that we keep a running total in the two Cumulative sub-
columns in the table for each time step. Let’s get the manual simulation started.
i ⴝ 0, t0 ⴝ 0. As shown in Figure 4.3, the first block after the start position indicates
that the model is initialized to its starting conditions. The simulation time step begins
at i 0. The initial value of the simulation clock is zero, t0 0. The system state
variables are set to ATM Status0 “Idle”; Number of Entities in Queue, NQ0 0;
and the Entity Attribute Array is cleared. This reflects the initial conditions of no
customer entities in the queue and an idle ATM. The statistical accumulator Total
Processed is set to zero. There are two different Cumulative variables in Table 4.1:
one to accumulate the time in queue values of ti Arrival Time, and the other to ac-
cumulate the values of the time-weighted number of entities in the queue, (ti ti1)
NQi1. Recall that ti Arrival Time is the amount of time that entities, which gained
access to the ATM, waited in queue. Both Cumulative variables (ti Arrival Time)
and (ti ti1)NQi1 are initialized to zero. Next, an initial arrival event and termina-
tion event are scheduled and placed under the Scheduled Future Events column. The
listing of an event is formatted as “(Entity Number, Event, and Event Time)”. Entity
Number denotes the customer number that the event pertains to (such as the first,
second, or third customer). Event is the type of event: a customer arrives, a customer
departs, or the simulation ends. Time is the future time that the event is to occur.
The event “(1, Arrive, 2.18)” under the Scheduled Future Events column prescribes
that the first customer entity is scheduled to arrive at time 2.18 minutes. The arrival
time was generated using the equation t0 E(3.0). To obtain the value returned from
the function E(3), we went to Table 3.2, read the first random variate from the Interar-
rival Time column (a value of 2.18 minutes), and added it to the current value of the
simulation clock, t0 0. The simulation is to be terminated after 22 minutes. Note the
“(__, End, 22.00)” under the Scheduled Future Events column. For the termination
event, no value is assigned to Entity Number because it is not relevant.
i ⴝ 1, t1 ⴝ 2.18. After the initialization step, the list of scheduled future events is added
to the event calendar in chronological order in preparation for the next simulation time
step i 1. The simulation clock is fast forwarded to the time that the next event is
scheduled to occur, which is t1 2.18 (the arrival time of the first customer to the ATM
queue), and then the event is processed. Following the simulation logic diagram, arrival
events are processed by first scheduling the future arrival event for the next customer
entity using the equation t1 E(3.0) 2.18 5.73 7.91 minutes. Note the value
of 5.73 returned by the function E(3.0) is the second random variate listed under the
Interarrival Time column of Table 3.2. This future event is placed under the Scheduled
Future Events column in Table 4.1 as “(2, Arrive, 7.91)”. Checking the status of the
ATM from the previous simulation time step reveals that the ATM is idle (ATM Status0
“Idle”). Therefore, the arriving customer entity immediately flows through the queue
to the ATM to conduct its transaction. The future departure event of this entity from
the ATM is scheduled using the equation t1 E(2.4) 2.18 0.10 2.28 minutes.
See “(1, Depart, 2.28)” under the Scheduled Future Events column, denoting that the
first customer entity is scheduled to depart the ATM at time 2.28 minutes. Note that the
value of 0.10 returned by the function E(2.4) is the first random variate listed under
har01307_ch04_087-114.indd 97 20/12/10 1:49 PM

the Service Time column of Table 3.2. The arriving customer entity’s arrival time is
then stored in the first position of the Entity Attribute Array to signify that it is being
served by the ATM. The ATM Status1 is set to “Busy,” and the statistical accumulators
for Entities Processed through Queue are updated. Add 1 to Total Processed and since
this entity entered the queue and immediately advanced to the idle ATM for processing,
record zero minutes in the Time in Queue, t1 Arrival Time, subcolumn and update this
statistic’s cumulative value. The statistical accumulators for Time-Weighted Number of
Entities in the Queue are updated next. Record zero for (t1 t0)NQ0 since there were
no entities in queue during the previous time step, NQ0 0, and update this statistic’s
cumulative value. Note the arrow “↑” entered under the Number of Entities in Queue,
NQ1 column. Recall that the arrow is placed there to signify that the number of entities
waiting in the queue has not changed from its previous value.
i ⴝ 2, t2 ⴝ 2.28. Following the loop back around to the top of the simulation logic
diagram, we place the two new future events onto the event calendar in chronological
order in preparation for the next simulation time step i 2. The simulation clock is
fast forwarded to t2 2.28, and the departure event for the first customer entity arriv-
ing to the system is processed. Given that there are no customers in the queue from
the previous time step, NQ1 0 (follow the arrows up to get this value), we simply
remove the departed customer from the first position of the Entity Attribute Array and
change the status of the ATM to idle, ATM Status2 “Idle”. The statistical accumu-
lators do not require updating because there are no customer entities waiting in the
queue or leaving the queue. The dashes (—) entered under the statistical accumulator
columns indicate that updates are not required. No new future events are scheduled.
As before, we follow the loop back to the top of the simulation logic diagram,
and place any new events, of which there are none at the end of time step i 2,
onto the event calendar in chronological order in preparation for the next simula-
tion time step i 3. The simulation clock is fast forwarded to t3 7.91, and the
arrival of the second customer entity to the ATM queue is processed. Complete the
processing of this event and continue the manual simulation until the termination
event “(__, End, 22.00)” reaches the top of the event calendar.
As you work through the simulation logic diagram to complete the manual simu-
lation, you will see that the fourth customer arriving to the system requires that you
use logic from a different path in the diagram. When the fourth customer entity arrives
to the ATM queue, simulation time step i 6, the ATM is busy (see ATM Status5) pro-
cessing customer entity 3’s transaction. Therefore, the fourth customer entity joins the
queue and waits to use the ATM. (Don’t forget that it invoked the creation of the fifth
customer’s arrival event.) The fourth entity’s arrival time of 15.17 minutes is stored
in the last position of the Entity Attribute Array in keeping with the first-in, first-out
(FIFO) rule. The Number of Entities in Queue, NQ6, is incremented to 1. Further, the
Time-Weighted Number of Entities in the Queue statistical accumulators are updated
by first computing (t6 t5)NQ5 (15.17 15.00)0 0 and then recording the result.
Next, this statistic’s cumulative value is updated. Customers 5, 6, and 7 also arrive to
the system finding the ATM busy and therefore take their place in the queue to wait
for the ATM.
The fourth customer waited a total of 3.08 minutes in the queue (see the
ti Arrival Time subcolumn) before it gained access to the ATM in time step
i 8 as the third customer departed. The value of 3.08 minutes in the queue for
har01307_ch04_087-114.indd 98 20/12/10 1:49 PM

the fourth customer computed in time step i 8 by t8 Arrival Time 18.25

15.17 3.08 minutes. Note that Arrival Time is the time that the fourth cus-
tomer arrived to the queue, and that the value is stored in the Entity Attribute
Array.
At time t12 22.00 minutes the simulation terminates and tallies the final
values for the statistical accumulators, indicating that a total of five customer
entities were processed through the queue. The total amount of time that these
five customers waited in the queue is 7.84 minutes. The final cumulative value
for Time-Weighted Number of Entities in the Queue is 13.21 minutes. Note that
at the end of the simulation, two customers are in the queue (customers 6 and 7)
and one is at the ATM (customer 5). A few quick observations are worth consider-
ing before we discuss how the accumulated values are used to calculate summary
statistics for a simulation output report.
This simple and brief (while tedious) manual simulation is relatively easy
to follow. But imagine a system with dozens of processes and dozens of factors
influencing behavior such as downtimes, mixed routings, resource contention,
and others. You can see how essential computers are for performing a simula-
tion of any magnitude. Computers have no difficulty tracking the many relation-
ships and updating the numerous statistics that are present in most simulations.
Equally as important, computers are not error prone and can perform millions of
instructions per second with absolute accuracy. We also want to point out that
the simulation logic diagram (Figure 4.3) and Table 4.1 were designed to convey
the essence of what happens inside a discrete-event simulation program. When
you view a trace report of a ProModel simulation in Lab Chapter 7 you will see
the simularities between the trace report and Table 4.1. Although the basic pro-
cess presented is sound, its efficiency could be improved. For example, there is
no need to keep both a “scheduled future events” list and an “event calendar.”
Instead, future events are inserted directly onto the event calendar as they are
created. We separated them to facilitate our describing the flow of information in
the discrete-event framework.
4.3.4 Calculating Results

When a simulation terminates, statistics that summarize the model’s behavior are
calculated and displayed in an output report. Many of the statistics reported in a
simulation output report consist of average values, although most software reports a
multitude of statistics including the maximum and minimum values observed during
the simulation. Average values may be either simple averages or time averages.
Simple-Average Statistic
A simple-average statistic is calculated by dividing the sum of all observation
values of a response variable by the number of observations:
兺i1 xi
n
Simple average ______
n
where n is the number of observations and xi is the value of ith observation. Exam-
ple simple-average statistics include the average time entities spent in the system
har01307_ch04_087-114.indd 99 20/12/10 1:49 PM

(from system entry to system exit), the average time entities spent at a specific
location, or the average time per use for a resource. An average of an observation-
based response variable in ProModel is computed as a simple average.
The average time that customer entities waited in the queue for their turn on
the ATM during the manual simulation reported in Table 4.1 is a simple-average
statistic. Recall that the simulation processed five customers through the queue.
Let xi denote the amount of time that the ith customer processed spent in the
queue. The average waiting time in queue based on the n 5 observations is
兺 i1 xi _____________________
5
Average time in queue _____ 0 0 0 3.08 4.76
5 5
7.84 minutes 1.57 minutes
___________
5
The values necessary for computing this average are accumulated under the Enti-
ties Processed through Queue columns of the manual simulation table (see the last
row of Table 4.1 for the cumulative value 兺(ti Arrival Time) 7.84 and Total
Processed 5).
Time-Average Statistic
A time-average statistic, sometimes called a time-weighted average, reports the
average value of a response variable weighted by the time duration for each
observed value of the variable:
兺 i1 (Tixi)
n
Time average ________
T
where xi denotes the value of the ith observation, Ti denotes the time duration of
the ith observation (the weighting factor), and T denotes the total duration over
which the observations were collected. Example time-average statistics include
the average number of entities in a system, the average number of entities at a
location, and the average utilization of a resource. An average of a time-weighted
response variable in ProModel is computed as a time average.
The average number of customer entities waiting in the queue location for
their turn on the ATM during the manual simulation is a time-average statistic.
Figure 4.4 is a plot of the number of customer entities in the queue during the
manual simulation recorded in Table 4.1. The 12 discrete-events manually simu-
lated in Table 4.1 are labeled t1, t2, t3, . . . , t11, t12, on the plot. Recall that ti denotes
the value of the simulation clock at time step i in Table 4.1, and that its initial
value is zero, t0 0.
Using the notation from the time-average equation just given, the total simu-
lation time illustrated in Figure 4.4 is T 22 minutes. The Ti denotes the dura-
tion of time step i (distance between adjacent discrete-events in Figure 4.4). That
is, Ti ti ti1 for i 1, 2, 3, . . . , 12. The xi denotes the queue’s contents (num-
ber of customer entities in the queue) during each time Ti interval. Therefore, xi
NQi1 for i 1, 2, 3, . . . , 12 (recall that in Table 4.1, NQi1 denotes the number of
har01307_ch04_087-114.indd 100 20/12/10 1:49 PM

har01307_ch04_087-114.indd 101
4
Number of customers in queue

0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
T1 2.18 T3 7.91 2.28 5.63 T4 12.37 7.91 4.46 T5 2.63 T8 2.51 …
T7 0.57 T12 1.50

T2 0.1 T6 0.17
t0 t1 t2 t3 t4 t5 t6 t7 t8 t11 t12
Simulation time, T 22
FIGURE 4.4
Number of customers in the queue during the manual simulation.
101
20/12/10 1:49 PM
customer entities in the queue from ti1 to ti). The time-average number of cus-
tomer entities in the queue (let’s call it Average NQ) is
12 12
兺i1 (Ti xi) 兺i1 (ti ti1)NQi1
Average NQ _______ _____________
T T
(2.18)(0) (0.1)(0) (5.63)(0) (4.46)(0) (2.63)(0) (0.17)(0) (0.57)(1) (2.51)(2) . . . (1.5)(2)
Average NQ _________________________________________________________________________________________
22
13.21 0.60 customers
Average NQ _____
22
You may recognize that the numerator of this equation (兺 i1(ti ti1)NQi1)
12
calculates the area under the plot of the queue’s contents during the simulation
(Figure 4.4). The values necessary for computing this area are accumulated under
the Time-Weighted Number of Entities in Queue column of Table 4.1 (see the
Cumulative value of 13.21 in the table’s last row).
4.3.5 Issues
Even though this example is a simple and somewhat crude simulation, it pro-
vides a good illustration of basic simulation issues that need to be addressed when
conducting a simulation study. First, note that the simulation start-up conditions
can bias the output statistics. Since the system started out empty, queue content
statistics are slightly less than what they might be if we began the simulation with
customers already in the system. Second, note that we ran the simulation for only
22 minutes before calculating the results. Had we ran longer, it is very likely that
the long-run average time in the queue would have been somewhat different (most
likely greater) than the time from the short run because the simulation did not
have a chance to reach a steady state.
These are the kinds of issues that should be addressed whenever running a
simulation. The modeler must carefully analyze the output and understand the
significance of the results that are given. This example also points to the need for
considering beforehand just how long a simulation should be run. These issues are
addressed in Chapters 8 and 9.
4.4 Commercial Simulation Software

While a simulation model may be programmed using any development language
such as C++ or Java, most models are built using commercial simulation software
such as ProModel. Commercial simulation software consists of several modules
for performing different functions during simulation modeling. Typical modules
are shown in Figure 4.5.
4.4.1 Modeling Interface Module

A modeler defines a model using an input or modeling interface module. This
module provides graphical tools, dialogs, and other text editing capabilities for
har01307_ch04_087-114.indd 102 20/12/10 1:49 PM

FIGURE 4.5
Typical components of
simulation software.
Modeling Simulation Output

interface interface interface
Modeling Model Simulation Output Output

processor data data data processor
Simulation Simulation
processor processor
entering and editing model information. External files used in the simulation are
specified here as well as run-time options (number of replications and so on).
4.4.2 Model Processor

When a completed model is run, the model processor takes the model database,
and any other external data files that are used as input data, and creates a simula-
tion database. This involves data translation and possibly language compilation.
This data conversion is performed because the model database and external files
are generally in a format that cannot be used efficiently by the simulation proces-
sor during the simulation. Some simulation languages are interpretive, meaning
that the simulation works with the input data the way they were entered. This al-
lows simulations to start running immediately without going through a translator,
but it slows down the speed at which the simulation executes.
In addition to translating the model input data, other data elements are created
for use during simulation, including statistical counters and state variables. Statis-
tical counters are used to log statistical information during the simulation such as
the cumulative number of entries into the system or the cumulative activity time
at a workstation.
4.4.3 Simulation Interface Module

The simulation interface module displays the animation that occurs during the
simulation run. It also permits the user to interact with the simulation to con-
trol the current animation speed, trace events, debug logic, query state variables,
har01307_ch04_087-114.indd 103 20/12/10 1:49 PM

request snapshot reports, pan or zoom the layout, and so forth. If visual interactive
capability is provided, the user is even permitted to make changes dynamically to
model variables with immediate visual feedback of the effects of such changes.
The animation speed can be adjusted and animation can even be disabled by
the user during the simulation. When unconstrained, a simulation is capable of
running as fast as the computer can process all of the events that occur within the
simulated time. The simulation clock advances instantly to each scheduled event;
the only central processing unit (CPU) time of the computer that is used is what
is necessary for processing the event logic. This is how simulation is able to run
in compressed time. It is also the reason why large models with millions of events
take so long to simulate. Ironically, in real life activities take time while events
take no time. In simulation, events take time while activities take no time. To slow
down a simulation, delay loops or system timers are used to create pauses between
events. These techniques give the appearance of elapsing time in an animation. In
some applications, it may even be desirable to run a simulation at the same rate
as a real clock. These real-time simulations are achieved by synchronizing the
simulation clock with the computer’s internal system clock. Human-in-the-loop
(such as operator training simulators) and hardware-in-the-loop (testing of new
equipment and control systems) are examples of real-time simulations.
4.4.4 Simulation Processor

The simulation processor processes the simulated events and updates the statisti-
cal accumulators and state variables. A typical simulation processor consists of
the following basic components:
• Clock variable—a variable used to keep track of the elapsed time during
the simulation.
• Event calendar—a list of scheduled events arranged chronologically
according to the time at which they are to occur.
• Event dispatch manager—the internal logic used to update the clock and
manage the execution of events as they occur.
• Event logic—algorithms that describe the logic to be executed and
statistics to be gathered when an event occurs.
• Waiting lists—one or more lists (arrays) containing events waiting for a
resource or other condition to be satisfied before continuing processing.
• Random number generator—an algorithm for generating one or more
streams of pseudo-random numbers between 0 and 1.
• Random variate generators—routines for generating random variates
from specified probability distributions.
4.4.5 Animation Processor

The animation processor interacts with the simulation database to update the
graphical data to correspond to the changing state data. Animation is usually
har01307_ch04_087-114.indd 104 20/12/10 1:49 PM

displayed during the simulation itself, although some simulation products create
an animation file that can be played back at the end of the simulation. In addition
to animated figures, dynamic graphs and history plots can be displayed during the
simulation.
Animation and dynamically updated displays and graphs provide a visual rep-
resentation of what is happening in the model while the simulation is running.
Animation comes in varying degrees of realism from three-dimensional animation
to simple animated flowcharts. Often, the only output from the simulation that is of
interest is what is displayed in the animation. This is particularly true when simula-
tion is used for facilitating conceptualization or for communication purposes.
A lot can be learned about model behavior by watching the animation (a pic-
ture is worth a thousand words, and an animation is worth a thousand pictures).
Animation can be as simple as circles moving from box to box, to detailed, real-
istic graphical representations. The strategic use of graphics should be planned in
advance to make the best use of them. While insufficient animation can weaken
the message, excessive use of graphics can distract from the central point to be
made. It is always good to dress up the simulation graphics for the final presen-
tation; however, such embellishments should be deferred at least until after the
model has been debugged.
For most simulations where statistical analysis is required, animation is no
substitute for the postsimulation summary, which gives a quantitative overview of
the entire system performance. Basing decisions on the animation alone reflects
shallow thinking and can even result in unwarranted conclusions.
4.4.6 Output Processor

The output processor summarizes the statistical data collected during the simula-
tion run and creates an output database. Statistics are reported on such perfor-
mance measures as
• Resource utilization.
• Queue sizes.
• Waiting times.
• Processing rates.
In addition to standard output reporting, most simulation software provides the
ability to write user-specified data out to external files.
4.4.7 Output Interface Module

The output interface module provides a user interface for displaying the output
results from the simulation. Output results may be displayed in the form of reports
or charts (bar charts, pie charts, or the like). Output data analysis capabilities
such as correlation analysis and confidence interval calculations also are often
provided. Some simulation products even point out potential problem areas such
as bottlenecks.
har01307_ch04_087-114.indd 105 20/12/10 1:49 PM

4.5 Simulation Using ProModel

ProModel is a powerful, yet easy-to-use, commercial simulation package that is
designed to effectively model any discrete-event processing system. It also has
continuous modeling capabilities for modeling flow in and out of tanks and other
vessels. The online tutorial in ProModel describes the building, running, and
output analysis of simulation models. A brief overview of how ProModel works is
presented here. The labs in this book provide a more detailed, hands-on approach
for actually building and running simulations using ProModel.
4.5.1 Building a Model

A model is defined in ProModel using simple graphical tools, data entry tables,
and fill-in-the-blank dialog boxes. In ProModel, a model consists of entities (the
items being processed), locations (the places where processing occurs), resources
(agents used to process and move entities), and paths (aisles and pathways along
which entities and resources traverse). Dialogs are associated with each of these
modeling elements for defining operational behavior such as entity arrivals and
processing logic. Schedules, downtimes, and other attributes can also be defined
for entities, resources, and locations.
Most of the system elements are defined graphically in ProModel (see Fig-
ure 4.6). A graphic representing a location, for example, is placed on the layout
to create a new location in the model (see Figure 4.7). Information about this
location can then be entered such as its name, capacity, and so on. Default values
are provided to help simplify this process. Defining objects graphically provides
a highly intuitive and visual approach to model building. The use of graphics is
optional, and a model can even be defined without using any graphics. In addition
to graphic objects provided by the modeling software, import capability is avail-
able to bring in graphics from other packages. This includes complete facility
layouts created using CAD software such as AutoCAD.
4.5.2 Running the Simulation

When running a model created in ProModel, the model database is translated
or compiled to create the simulation database. The animation in ProModel is
displayed concurrently with the simulation. Animation graphics are classi-
fied as either static or dynamic. Static graphics include walls, aisles, machines,
screen text, and others. Static graphics provide the background against which the
FIGURE 4.6
Sample of ProModel
graphic objects.
har01307_ch04_087-114.indd 106 20/12/10 1:49 PM

FIGURE 4.7
ProModel animation provides useful feedback.
animation takes place. This background might be a CAD layout imported into
the model. The dynamic animation objects that move around on the background
during the simulation include entities (parts, customers, and so on) and resources
(people, fork trucks, and so forth). Animation also includes dynamically updated
counters, indicators, gauges, and graphs that display count, status, and statistical
information (see Figure 4.7).
4.5.3 Output Analysis

The output processor in ProModel provides both summary and detailed statistics
on key performance measures. Simulation results are presented in the form of
reports, plots, histograms, pie charts, and others. Output data analysis capabilities
such as confidence interval estimation are provided for more precise analysis.
Outputs from multiple replications and multiple scenarios can also be summa-
rized and compared. Averaging performance across replications and showing
multiple scenario output side-by-side make the results much easier to interpret.
Summary Reports
Summary reports show totals, averages, and other overall values of interest.
Figure 4.8 shows an output report summary generated from a ProModel simula-
tion run.
har01307_ch04_087-114.indd 107 20/12/10 1:49 PM

FIGURE 4.8
Summary report of simulation activity.
--------------------------------------------------------------------------------
Mfg_cost.MOD (Model Parameters - Rep. 1)
Run Date/Time 6/15/2010 15:14
Model Title Manufacturing Costing Optimization
Model Path/File C:\Program Files\Models\Mfg_cost.MOD
Warmup Time (HR) 5
Simulation Time (HR) 10
--------------------------------------------------------------------------------
LOCATIONS
Avg Time
Scheduled Total Per Entry Avg Maximum Current % Utili-
Name Time (HR) Capacity Entries (MIN) Contents Contents Contents zation
Receive 10 2 22 54.54545 2 2 2 100
NC Lathe 1 10 1 57 10.11647 0.961065 1 1 96.1065
NC Lathe 2 10 1 57 9.891842 0.939725 1 1 93.9725
Degrease 10 2 114 10.18895 1.9359 2 2 96.795
Inspect 10 1 113 4.690053 0.883293 1 1 88.3293
Bearing Que 10 100 90 34.51747 5.17762 13 11 5.17762
Loc1 10 5 117 25.64103 5 5 5 72.0184
RESOURCES
Sche- Avg Time Avg Time %
duled Work Number Avg Time Travel Travel Blocked In %
Time Time Times Per Usage To Use To Park Travel Utili-
Name Units (HR) (HR) Used (MIN) (MIN) (MIN) zation
CellOp.1 1 10 5.78 122 2.737648 0.1038017 0 0 57.7588
CellOp.2 1 10 5.57 118 2.726542 0.1062712 0 0 55.712
CellOp.3 1 10 5.07 115 2.541652 0.102 0 0 50.67
CellOp 3 30 16.41 355 2.670465 0.1040395 0 0 54.7136
ENTITIES
Avg Avg Time Avg
Current Time In In Move Avg Time Time In Avg Time
Total Qty In System Logic Waiting Operation Blocked
Name Exits System (MIN) (MIN) (MIN) (MIN) (MIN)
Pallet 20 2 63.0418 0 30.56155 1 31.48025
Blank 0 7 0 0 0 0 0
Cog 79 3 52.5925 0.849203 3.226911 33.533266 14.98313
Reject 33 0 49.5601 0.853636 2.488545 33.065636 13.15224
Bearing 78 12 42.1855 0.05 35.58995 0 6.54559
har01307_ch04_087-114.indd 108 20/12/10 1:49 PM

FIGURE 4.9
Time-series graph
showing changes in
queue size over time.
Time Series Plots and Histograms

Sometimes a summary report is too general to capture the information being sought.
Fluctuations in model behavior over time can be viewed using a time series report.
In ProModel, time series output data are displayed graphically so that patterns can be
identified easily. Time series plots can show how inventory levels fluctuate through-
out the simulation or how changes in workload affect resource utilization. Figure 4.9
is a time-series graph showing how the length of a queue fluctuates over time.
Once collected, time series data can be grouped in different ways to show
patterns or trends. One common way of showing patterns is using a histogram.
The histogram in Figure 4.10 breaks down contents in a queue to show the per-
centage of time different quantities were in that queue.
As depicted in Figure 4.10, over 95 percent of the time, there were fewer than
12 items in the queue. This is more meaningful than simply knowing what the
average length of the queue was.
4.6 Languages versus Simulators

A distinction is frequently made between simulation languages and simulators.
Simulation languages are often considered more general purpose and have fewer
predefined constructs for modeling specific types of systems. Simulators, on the
other hand, are designed to handle specific problems—for example, a job shop
simulator or a clinic simulator. The first simulators that appeared provided little, if
any, programming capability, just as the first simulation languages provided few, if
any, special modeling constructs to facilitate modeling. Consequently, simulators
har01307_ch04_087-114.indd 109 20/12/10 1:49 PM

FIGURE 4.10
Histogram of queue
contents.
acquired the reputation of being easy to use but inflexible, while simulation lan-
guages were branded as being very flexible but difficult to use (see Figure 4.11).
Over time, the distinction between simulation languages and simulators
has become blurred as specialized modeling constructs have been added to
general-purpose simulation languages, making them easier to use. During the
same period, general programming extensions have been added to simulators,
making them more flexible. The most popular simulation tools today combine
powerful industry-specific constructs with flexible programming capabilities,
all accessible from an intuitive graphical user interface (Bowden 1998). Some
tools are even configurable, allowing the software to be adapted to specific ap-
plications yet still retaining programming capability. Rather than put languages
and simulators on opposite ends of the same spectrum as though flexibility
and ease of use were mutually exclusive, it is more appropriate to measure the
flexibility and ease of use for all simulation software along two separate axes
(Figure 4.12).
FIGURE 4.11 Simulators Languages

Old paradigm that
polarized ease of use
and flexibility. Ease of use Flexibility
har01307_ch04_087-114.indd 110 20/12/10 1:49 PM

FIGURE 4.12
Easy
New paradigm
that views ease of Early
use and flexibility Current best-of-breed
simulators products
as independent
Ease of use
characteristics.
Early
languages
Hard
Low High
Flexibility
4.7 Future of Simulation

Simulation products have evolved to provide more than simply stand-alone simu-
lation capability. Modern simulation products have open architectures based on
component technology and standard data access methods (like SQL) to provide
interfacing capability with other applications such as CAD programs and enter-
prise planning tools. Surveys reported annually in Industrial Engineering Solu-
tions show that most simulation products have the following features:
• Input data analysis for distribution fitting.
• Point-and-click graphical user interface.
• Reusable components and templates.
• Two- (2-D) and three-dimensional (3-D) animation.
• Online help and tutorials.
• Interactive debugging.
• Automatic model generation.
• Output analysis tools.
• Optimization.
• Open architecture and database connectivity.
Simulation is a technology that will continue to evolve as related technolo-
gies improve and more time is devoted to the development of the software. Prod-
ucts will become easier to use with more intelligence being incorporated into the
software itself. Evidence of this trend can already be seen by optimization and
other time-saving utilities that are appearing in simulation products. Animation
and other graphical visualization techniques will continue to play an important
har01307_ch04_087-114.indd 111 20/12/10 1:49 PM

role in simulation. As 3-D and other graphic technologies advance, these features
will also be incorporated into simulation products.
Simulation products targeted at vertical markets are on the rise. This trend
is driven by efforts to make simulation easier to use and more solution oriented.
Specific areas where dedicated simulators have been developed include call cen-
ter management, supply chain management, and high-speed processing. At the
same time many simulation applications are becoming more narrowly focused,
others are becoming more global and look at the entire enterprise or value chain
in a hierarchical fashion from top to bottom.
Perhaps the most dramatic change in simulation will be in the area of soft-
ware interoperability and technology integration. Historically, simulation has been
viewed as a stand-alone, project-based technology. Simulation models were built
to support an analysis project, to predict the performance of complex systems, and
to select the best alternative from a few well-defined alternatives. Typically these
projects were time-consuming and expensive, and relied heavily on the expertise
of a simulation analyst or consultant. The models produced were generally “single
use” models that were discarded after the project.
In recent years, the simulation industry has seen increasing interest in ex-
tending the useful life of simulation models by using them on an ongoing basis
(Harrell and Hicks 1998). Front-end spreadsheets and push-button user inter-
faces are making such models more accessible to decision makers. In these
flexible simulation models, controlled changes can be made to models through-
out the system life cycle. This trend is growing to include dynamic links to
databases and other data sources, enabling entire models actually to be built
and run in the background using data already available from other enterprise
applications.
The trend to integrate simulation as a component in enterprise applications
is part of a larger industry effort to provide software components that can be
distributed over the Internet using cloud computing and other service-oriented
technologies. These technologies promise to enable parallel and distributed model
execution and provide a mechanism for maintaining distributed model repositories
that can be shared by many modelers (Fishwick 1997). The interest in Web-based
simulation, like all other Web-based applications, continues to grow.
4.8 Summary
Most manufacturing and service systems are modeled using dynamic, stochas-
tic, discrete-event simulation. Discrete-event simulation works by converting all
activities to events and consequent reactions. Events are either time-triggered or
condition-triggered, and are therefore processed either chronologically or when a
satisfying condition has been met.
Simulation models are generally defined using commercial simulation
software that provides convenient modeling constructs and analysis tools.
Simulation software consists of several modules with which the user interacts.
har01307_ch04_087-114.indd 112 20/12/10 1:49 PM

Internally, model data are converted to simulation data, which are processed
during the simulation. At the end of the simulation, statistics are summarized
in an output database that can be tabulated or graphed in various forms. The
future of simulation is promising and will continue to incorporate exciting new
technologies.

1. In simulation, the completion of an activity time that is random must be
known at the start of the activity. Why is this necessary?
2. Give an example of an activity whose completion is a scheduled event
and one whose completion is a conditional event.
3. For the first 10 customers processed completely through the ATM
spreadsheet simulation presented in Table 3.2 of Chapter 3, construct
a table similar to Table 4.1 as you carry out a manual discrete-event
simulation of the ATM system to
a. Compute the average amount of time the first 10 customers spent in
the system. Hint: Add time in system and corresponding cumulative
columns to the table.
b. Compute the average amount of time the first 10 customers spent in
the queue.
c. Plot the number of customers in the queue over the course of the
simulation and compute the average number of customers in the
queue for the simulation.
d. Compute the utilization of the ATM. Hint: Define a utilization
variable that is equal to zero when the ATM is idle and is equal to 1
when the ATM is busy. At the end of the simulation, compute a time-
weighted average of the utilization variable.
4. Identify whether each of the following output statistics would be
computed as a simple or as a time-weighted average value.
a. Average utilization of a resource.
b. Average time entities spend in a queue.
c. Average time entities spend waiting for a resource.
d. Average number of entities waiting for a particular resource.
e. Average repair time for a machine.
5. Give an example of a situation where a time series graph would be more
useful than just seeing an average value.
6. In real life, activities take time and events take no time. During a
simulation, activities take no time and events take all of the time.
Explain this paradox.
7. What are the main components of simulation software?
8. Look up a simulation product on the Internet or in a trade journal and
describe two promotional features of the product being advertised.
har01307_ch04_087-114.indd 113 20/12/10 1:49 PM

References
Bowden, Royce. “The Spectrum of Simulation Software.” IIE Solutions, May 1998,
pp. 44–46.
Fishwick, Paul A. “Web-Based Simulation.” In Proceedings of the 1997 Winter Simulation
Conference, ed. S. Andradottir, K. J. Healy, D. H. Withers, and B. L. Nelson. Institute
of Electric and Electronics Engineers, Piscataway, NJ, 1997. pp. 100–109.
Gottfried, Byron S. Elements of Stochastic Process Simulation. Englewood Cliffs, NJ:
Prentice Hall, 1984, p. 8.
Haider, S. W., and J. Banks. “Simulation Software Products for Analyzing Manufacturing
Systems.” Industrial Engineering, July 1986, p. 98.
Harrell, Charles R., and Don Hicks. “Simulation Software Component Architecture for
Simulation-Based Enterprise Applications.” Proceedings of the 1998 Winter Simula-
tion Conference, ed. D. J. Medeiros, E. F. Watson, J. S. Carson, and M. S. Manivan-
nan. Institute of Electrical and Electronics Engineers, Piscataway, New Jersey, 1998,
pp. 1717–21.
har01307_ch04_087-114.indd 114 20/12/10 1:49 PM

C H A P T E R
5 DATA COLLECTION
AND ANALYSIS
“You can observe a lot just by watching.”

—Yogi Berra
5.1 Introduction
Simulation is all about making better decisions based on accurate information. A
decision can be no better than the information on which it is based. In this chap-
ter, we look at the data-gathering phase of a simulation project that defines the
system being modeled. The result of the data-gathering effort is a conceptual or
mental model of how the system is configured and how it operates. This concep-
tual model may take the form of a written description, a flow diagram, or even a
simple sketch on the back of an envelope. It becomes the basis for the simulation
model that will be created.
Data collection is the most challenging and time-consuming task in simula-
tion. For new systems, information is usually very sketchy and only roughly esti-
mated. For existing systems, there may be years of raw, unorganized data to sort
through. Information is seldom available in a form that is directly usable in build-
ing a simulation model. It nearly always needs to be filtered and massaged to get it
into the right format and to reflect the projected conditions under which the system
is to be analyzed. Many data-gathering efforts end up with lots of data but little
useful information. Data should be gathered purposefully to avoid wasting not only
the modeler’s time but also the time of individuals who are supplying the data.
This chapter presents guidelines and procedures for gathering data. Statistical
techniques for analyzing data and fitting probability distributions to data are also
discussed. The following questions are answered:
• What is the best procedure to follow when gathering data?
• What types of data should be gathered?
115
har01307_ch05_115-160.indd 115 20/12/10 2:01 PM

• What sources should be used when gathering data?

• What types of analyses should be performed on the data?
• How do you select the right probability distribution that represents the data?
• How should data be documented?
5.2 Guidelines for Data Gathering

Data gathering should not be performed haphazardly with the idea that you will
sort through all the information once it is collected. Data should be collected
systematically, seeking specific items of information that are needed for building
the model. If approached with the end in mind, data gathering will be much more
productive. Following are a few guidelines to keep in mind when gathering data
that will help maintain a purposeful focus.
1. Identify triggering events. When defining the activities that occur in the
system to be modeled, it is important to identify the causes or conditions
that trigger the activities. In gathering downtime data, for example, it
is useful to distinguish between downtimes due to failure, planned or
scheduled downtimes, and downtimes that are actually idle periods due
to unavailability of stock. In other words, you need to know the cause
of the downtime. Questions should be asked such as “What triggers
the movement of entities?” and “What triggers the use of a particular
resource?” To be valid, a model needs to capture the right triggering
events that initiate the activities within the system.
2. Focus only on key impact factors. Discrimination should be used when
gathering data to avoid wasting time tracking down information that
has little or no impact on system performance. For example, off-shift
activities such as preventive maintenance that do not delay the process
flow should be ignored because they are performed during off-hours.
Similarly, extremely rare downtimes, negligible move times, on-the-fly
inspections, external setups, and other activities that have little or no
impact on the overall flow may be safely ignored.
3. Isolate actual activity times. In determining activity times, it is important
to isolate only the time it takes to do the activity itself, excluding any
extraneous time waiting for material and resources so the activity can
be performed. For example, filling a customer order or performing
an assembly operation requires that items or component parts first be
available. The time to wait for items or parts should not be included
in the time to do the actual assembly. This waiting time is going to
vary based on conditions in the system and therefore can’t be known
in advance. It will be known only as the simulation is run. In some
instances, isolating actual activity times may be difficult, especially
when dealing with historical data reflecting aggregate times. Historical
har01307_ch05_115-160.indd 116 20/12/10 2:01 PM

Chapter 5 Data Collection and Analysis 117
records of repair times, for example, often lump together the time spent
waiting for repair personnel to become available and the actual time
spent performing the repair. What you would like to do is separate
the waiting time from the actual repair time because the waiting time
is a function of the availability of the repair person, which may vary
depending on the system operation.
4. Look for common groupings. When dealing with lots of variety in a
simulation such as hundreds of part types or customer profiles, it helps
to look for common groupings or patterns. If, for example, you are
modeling a process that has 300 entity types, it may be difficult to get
information on the exact mix and all of the varied routings that can
occur. Having such detailed information is usually too cumbersome to
work with even if you did have it. The solution is to reduce the data
to common behaviors and patterns. One way to group common data is
to first identify general categories into which all data can be assigned.
Then the percentage of cases that fall within each category is calculated
or estimated. It is not uncommon for beginning modelers to attempt to
use actual logged input streams such as customer arrivals or material
shipments when building a model. After struggling to define hundreds
of individual arrivals or routings, it begins to dawn on them that they
can group this information into a few categories and assign probabilities
that any given instance will fall within a particular category. This allows
dozens and sometimes hundreds of unique instances to be described in a
few brief commands. The secret to identifying common groupings is to
“think probabilistically.”
5. Focus on essence rather than substance. A system definition for
modeling purposes should capture the cause-and-effect relationships
and ignore the meaningless (to simulation) details. This is called system
abstraction and seeks to define the essence of system behavior rather
than the substance. A system should be abstracted to the highest level
possible while still preserving the essence of the system operation.
Using this “black box” approach to system definition, we are not
concerned about the nature of the activity being performed, such as
milling, grinding, or inspection. We are interested only in the impact
that the activity has on the use of resources and the delay of entity flow.
A proficient modeler constantly is thinking abstractly about the system
operation and avoids getting caught up in the mechanics of the process.
6. Separate input variables from response variables. First-time modelers
often confuse input variables that define the operation of the system
with response variables that report system performance. Input variables
define how the system works (interarrival times, activity times, routing
sequences, machine downtime, and the like) and should be the focus of
the data gathering. Response variables describe how the system responds
to a given set of input variables (amount of work in process, resource
har01307_ch05_115-160.indd 117 20/12/10 2:01 PM

utilization, throughput times, and so on). Response variables do not

“drive” model behavior. Performance or response data should be of
interest only if you are building an as-is model and want to validate the
simulation by comparing the simulation output with actual performance.
Following these guidelines, the data-gathering effort will be much more pro-
ductive. In addition to keeping these guidelines in mind, it is also helpful to fol-
low a logical sequence of steps when gathering data. This ensures that time and
efforts are well spent. The steps to gathering data should follow roughly this
sequence:
Step 1: Determine data requirements.
Step 2: Identify data sources.
Step 3: Collect the data.
Step 4: Make assumptions where necessary.
Step 5: Analyze the data.
Step 6: Document and approve the data.
Of course, most of these steps overlap, and some iteration will occur as
objectives change and assumptions are refined. The balance of this chapter is
devoted to providing recommendations and examples of how to perform each of
these steps.
5.3 Determining Data Requirements

The first step in gathering data is to determine the data required for building the
model. This should be dictated primarily by the model scope and level of detail
required for achieving the objectives of the simulation. System data can be cate-
gorized as structural data, operational data, or numerical data.
5.3.1 Structural Data

Structural data involve all of the objects in the system to be modeled. This includes
such elements as entities (products, customers, and so on), resources (operators,
machines) and locations (waiting areas, workstations). Structural information
basically describes the layout or configuration of the system as well as identifies
the items that are processed. It is important that all relevant components that affect
the behavior of the system be included.
5.3.2 Operational Data

Operational data explain how the system operates—that is, when, where, and
how events and activities take place. Operational data consist of all the logical or
behavioral information about the system such as routings, schedules, downtime
behavior, and resource allocation. If the process is structured and well controlled,
har01307_ch05_115-160.indd 118 20/12/10 2:01 PM

operational information is easy to define. If, on the other hand, the process has
evolved into an informal operation with no set rules, it can be very difficult to
define. For a system to be simulated, operating policies that are undefined and
ambiguous must be codified into defined procedures and rules. If decisions and
outcomes vary, it is important to at least define this variability statistically using
probability expressions or distributions.
5.3.3 Numerical Data

Numerical data provide quantitative information about the system. Examples of
numerical data include capacities, arrival rates, activity times, and time between
failures. Some numerical values are easily determined, such as resource capacities
and working hours. Other values are more difficult to assess, such as time between
failures or routing probabilities. This is especially true if the system being mod-
eled is new and data are unavailable. It is important when gathering numerical
data to ask exactly what a particular data value represents. For example, in col-
lecting machining times, one should ask whether the time reflects load and unload
time or tool adjustments, and whether it is for processing a batch of parts rather
than individual parts.
5.3.4 Use of a Questionnaire

To help focus data-gathering efforts on the right information and to ensure produc-
tive meetings with others on whom you depend for model information, it may be
useful to prepare a questionnaire for use in interviews. It is much easier for people
to respond to specific questions than to puzzle over and guess at what information
is needed. When conducting an interview, sometimes giving the contact person a
specific list of questions in advance reduces the length of the interview and the
number of follow-up visits that would otherwise be needed.
Below is a list of questions that might be included in a questionnaire. For any
questions requiring numeric information that could take the form of a random vari-
able, it is best if either a good sample of data or the actual defining probability dis-
tribution can be obtained. If sample data are unavailable, it is useful to get at least
estimates of the minimum, most likely, and maximum values until more precise data
can be obtained. Average values should be used only as preliminary estimates.
System Description Questionnaire

1. What types of entities are processed in the system?
2. What is the routing sequence (each stopping or decision point) for each
entity type?
3. Where, when, and in what quantities do entities enter the system?
4. What are the time and resource requirements for each operation and
move?
5. In what quantities are entities processed and moved? (Define for each
location.)
har01307_ch05_115-160.indd 119 20/12/10 2:01 PM

6. What triggers entity movement from location to location (completion

of the operation, accumulation of a batch, a signal from a downstream
location, etc.)?
7. How do locations and resources determine which job to do next (oldest
waiting, highest priority, etc.)?
8. How are alternative routing and operation decisions made (percentage,
condition, etc.)?
9. How often do interruptions occur (setups, downtimes, etc.) and what
resources and time are required when they happen?
10. What is the schedule of availability for locations and resources (define
in terms of shifts, break times, scheduled maintenance intervals, etc.)?
Answers to these questions should provide most if not all of the information
needed to build a model. Depending on the nature of the system and level of detail
required, additional questions may be added.
5.4 Identifying Data Sources

Rarely is all the information needed to build a model available from a single
source. It is usually the result of reviewing reports, conducting personal inter-
views, making personal observations, and making lots of assumptions. “It has
been my experience,” notes Carson (1986), “that for large-scale real systems,
there is seldom any one individual who understands how the system works in suf-
ficient detail to build an accurate simulation model. The modeler must be willing
to be a bit of a detective to ferret out the necessary knowledge.” This is where
persistence, careful observation, and good communication skills pay off.
The types of sources to use when gathering data will depend largely on whether
you are simulating an existing system or a new system. For existing systems, there
will be a much wider range of sources from which to draw since there are records
and individuals with firsthand knowledge of the system. The fact that the system
exists also permits the modeler to personally observe the system and therefore be
less dependent on the input of others—although collaboration in data gathering is
always recommended. For new systems, the sources of information are usually
limited to those individuals directly involved in the design of the system. Most of
the information is obtained from system specifications and personal interviews with
the design team. Depending on the circumstances, good sources of data include
• Historical records—production, sales, scrap rates, equipment reliability.
• System documentation—process plans, facility layouts, work procedures.
• Personal observation—facility walk-through, time studies, work sampling.
• Personal interviews—operators (work methods), maintenance personnel
(repair procedures), engineers (routings), managers (schedules and
forecasts).
• Comparison with similar systems—within the company, within the same
industry, within any industry.
har01307_ch05_115-160.indd 120 20/12/10 2:01 PM

• Vendor claims—process times, reliability of new machines.

• Design estimates—process times, move times, and so on for a new system.
• Research literature—published research on learning curves,
predetermined time studies.
In deciding whether to use a particular data source, it is important to consider the
reliability and accessibility of the source. Reliability of the source directly impacts
the validity of the model. A manager’s recollection, for example, may not be as reli-
able as actual production logs. Opinions are frequently biased and tend to provide
distorted information. Objectivity should be sought when using any data source.
Accessibility is also an important factor in source selection. If the source is
difficult to access, such as historical data filed away in some corporate archive or
buried somewhere among hundreds of reports, it may have to be overlooked. If
the system spans several businesses, such as a supply chain, there even may be an
unwillingness to disclose information. The principle to follow is to find out what
sources are accessible and then use the sources that provide the most reliable and
complete information. Secondary sources can always validate the primary source.
5.5 Collecting the Data

When gathering data, it is best to go from general to specific. In practice, data
gathering continues through to the end of the simulation project as objectives
change and information that was unavailable at the beginning of the project be-
gins to materialize. Consistent with this progressive refinement approach, data
should be gathered in the following sequence:
1. Define the overall entity flow.
2. Develop a description of operation.
3. Define incidental details and firm up data values.
This doesn’t necessarily mean that data will become available in this sequence. It
only means that the focus of attention should follow this order. For example, some
assumptions may need to be made initially to come up with a complete definition
of entity flow. Other incidental information obtained early on, such as downtimes,
may need to be set aside until it is needed at a later stage of model building.
5.5.1 Defining the Entity Flow

The first area of focus in data collection should be in defining the basic entity flow
through the system. This establishes a skeletal framework to which additional
data can be attached. The entity flow is defined by following the entity movement
through the system, taking the perspective of the entity itself. The overall entity
flow is best described using an entity flow diagram (see Figure 5.1) or perhaps by
superimposing the entity flow on an actual layout of the system.
An entity flow diagram is slightly different from a process flowchart. A pro-
cess flowchart shows the logical sequence of activities through which entities
har01307_ch05_115-160.indd 121 20/12/10 2:01 PM

FIGURE 5.1 Product A

Entity flow diagram.
Station 1 Station 2 Station 3A
Station 3B
FIGURE 5.2 Exam

Check-in Waiting room Checkout
Entity flow diagram
counter room (3) counter
for patient processing Patient
at Dr. Brown’s office.
go and defines what happens to the entity, not where it happens. An entity flow
diagram, on the other hand, is more of a routing chart that shows the physical
movement of entities through the system from location to location. An entity flow
diagram should depict any branching that may occur in the flow such as rout-
ings to alternative work centers or rework loops. The purpose of the entity flow
diagram is to document the overall flow of entities in the system and to provide a
visual aid for communicating the entity flow to others. A flow diagram is easy to
understand and gets everyone thinking in the same way about the system. It can
easily be expanded as additional information is gathered to show activity times,
where resources are used, and so forth.
5.5.2 Developing a Description of Operation

Once an entity flow diagram has been created, a description of operation should
be developed that explains how entities are processed through the system. A de-
scription of operation may be written in a step-by-step, brief narrative form, or it
may be represented in tabular form. For simple systems, the entity flow diagram
itself can be annotated with operational information. Whatever the form, it should
identify for each type of entity at each location in the system
• The time and resource requirements of the activity or operation.
• Where, when, and in what quantities entities get routed next.
• The time and resource requirements for moving to the next location.
Example: Patients enter Dr. Brown’s office and sign in at the check-in counter be-
fore taking a seat to wait for the nurse’s call. Patients are summoned to one of three
examination rooms when one becomes available. The nurse escorts patients to the ex-
amination room, where they wait for Dr. Brown to see them. After treatment, patients
return on their own to the checkout counter, where they make payment and possibly
reschedule another visit. This process is shown in the entity flow diagram in Figure 5.2.
Using the entity flow diagram and information provided, a summary descrip-
tion of operation is created using a table format, as shown in Table 5.1.
har01307_ch05_115-160.indd 122 20/12/10 2:01 PM

TABLE 5.1 Process Description for Dr. Brown’s Office

Activity Activity Next Move Move Move
Location Time Resource Location Trigger Time Resource
Check-in counter N(1,.2) min. Secretary Waiting room None 0.2 min. None
Waiting room None None Exam room When room is 0.8 min.* Nurse
available
Exam room N(15,4) min. Doctor Checkout counter None 0.2 min. None
Checkout counter N(3,.5) min. Secretary Exit None None None
*Stops to get weighed on the way.
Notice that the description of operation really provides the details of the en-
tity flow diagram. This detail is needed for defining the simulation model. The
times associated with activities and movements can be just estimates at this stage.
The important thing to accomplish at this point is simply to describe how entities
are processed through the system.
The entity flow diagram, together with the description of operation, provides
a good data document that can be expanded as the project progresses. At this
point, it is a good idea to conduct a structured walk-through of the operation using
the entity flow diagram as the focal point. Individuals should be involved in this
review who are familiar with the operation to ensure that the description of opera-
tion is accurate and complete.
Based on this description of operation, a first cut at building the model can
begin. Using ProModel, a model for simulating Dr. Brown’s practice can be built
in a matter of just a few minutes. Little translation is needed as both the diagram
(Figure 5.2) and the data table (Table 5.1) can be entered pretty much in the same
way they are shown. The only additional modeling information required to build
a running model is the interarrival time of patients.
Getting a basic model up and running early in a simulation project helps hold
the interest of stakeholders. It also helps identify missing information and moti-
vates team members to try to fill in these information gaps. Additional questions
about the operation of the system are usually raised once a basic model is running.
Some of the questions that begin to be asked are “Have all of the routings been
accounted for?” and “Have any entities been overlooked?” In essence, modeling
the system actually helps define and validate system data.
5.5.3 Defining Incidental Details and Refining Data Values

Once a basic model has been constructed and tested, additional details of the
process such as downtimes, setups, and work priorities can be added. This infor-
mation is not essential to getting a running model, but is necessary for a complete
and accurate model. Sometimes the decision of whether to include a particular
element of the system such as a downtime is easier to make once the basic model
is running. The potential impact of the element and the time available to imple-
ment it in the model are much clearer than at the start of the project.
har01307_ch05_115-160.indd 123 20/12/10 2:01 PM

At this point, any numerical values such as activity times, arrival rates, and
others should also be firmed up. Having a running model enables estimates and
other assumptions to be tested to see if it is necessary to spend additional time
getting more accurate information. For existing systems, obtaining more accurate
data is usually accomplished by conducting time studies of the activity or event
under investigation. A sample is gathered to represent all conditions under which
the activity or event occurs. Any biases that do not represent normal operating
conditions are eliminated. The sample size should be large enough to provide
an accurate picture yet not so large that it becomes costly without really adding
additional information.
5.6 Making Assumptions

A simulation model can run with incorrect data, but it can’t run with incomplete
data. Using an architectural metaphor, a computer model is less like an artist’s
rendition of a building and much more like a full set of blueprints—complete
with dimensions and material lists. This is one of the reasons why simulation is
so beneficial. It gets down to the operational details that actually define how the
system runs.
It doesn’t take long after data gathering has commenced to realize that certain
information is unavailable or perhaps unreliable. Complete, accurate, and up-to-
date information is rarely obtainable, especially when modeling a new system
about which little is known. Even for existing systems, it may be impractical to
obtain certain types of data, such as long time intervals between machine failures.
Because the reliability of information on new systems is often questionable,
the validity of the model itself is going to be suspect. Accurate information is
going to be lacking in at least one or more areas of almost any model—even in
models of existing systems. Because a simulation nearly always is run for some
future anticipated situation, there are always going to be assumptions made re-
garding any unknown future conditions. In simulation, we must be content to
begin with assumptions and be grateful whenever we can make assertions.
Since a model that is based on assumptions is automatically invalid, the ques-
tion might be raised, What good is a simulation that is based on assumptions? We
forget that we routinely make decisions in life based on assumptions and therefore
can never be quite sure of the outcome. Any prediction about how a system will
operate in the future is based on assumptions about the future. There is nothing
wrong with making assumptions when doing simulation as long as our confidence
in the results never exceeds our confidence in the underlying assumptions. The
reason for doing simulation is to be able to predict system performance based on
given assumptions. It is better to make decisions knowing the implications of our
assumptions than to compound the error of our decision by drawing erroneous
inferences from these assumptions. Incorporating assumptions into a model can
actually help validate our assumptions by seeing whether they make sense in the
overall operation of the model. Seeing absurd behavior may tell us that certain
assumptions just don’t make sense.
har01307_ch05_115-160.indd 124 20/12/10 2:01 PM

For comparative studies in which two design alternatives are evaluated, the
fact that assumptions are made is less significant because we are evaluating rel-
ative performance, not absolute performance. For example, if we are trying to
determine whether on-time deliveries can be improved by assigning tasks to a re-
source by due date rather than by first-come, first-served, a simulation can provide
useful information to make this decision without necessarily having completely
accurate data. Because both models use the same assumptions, it may be possible
to compare relative performance. We may not know what the absolute perfor-
mance of the best option is, but we should be able to assess fairly accurately how
much better one option is than another.
Some assumptions will naturally have a greater influence on the validity of
a model than others. For example, in a system with large processing times com-
pared to move times, a move time that is off by 20 percent may make little or no
difference in system throughput. On the other hand, an activity time that is off by
20 percent could make a 20 percent difference in throughput. One way to assess
the influence of an assumption on the validity of a model is through sensitivity
analysis. Sensitivity analysis, in which a range of values is tested for potential im-
pact on model performance, can indicate just how accurate an assumption needs
to be. A decision can then be made to firm up the assumption or to leave it as is.
If, for example, the degree of variation in a particular activity time has little or no
impact on system performance, then a constant activity time may be used. At the
other extreme, it may be found that even the type of distribution has a noticeable
impact on model behavior and therefore needs to be selected carefully.
A simple approach to sensitivity analysis for a particular assumption is to
run three different scenarios showing (1) a “best” or most optimistic case, (2) a
“worst” or most pessimistic case, and (3) a “most likely” or best-guess case. These
runs will help determine the extent to which the assumption influences model
behavior. It will also help assess the risk of relying on the particular assumption.
5.7 Statistical Analysis of Numerical Data

To be of use in a simulation model, raw data must be analyzed and interpreted so
that the system operation is correctly represented in the model. Triggering events
and other cause-and-effect relationships need to be identified. Irrelevant or insig-
nificant data need to be discarded. Actual activity times need to be isolated from
system-caused delays. The numerous instances of a particular activity need to be
generalized into a few defining patterns. Complex operation descriptions need to
be translated into simplified abstractions. Input or defining system variables need
to be separated from output or response variables. Finally, summarized historical
data should be projected forward to the time period being studied by the simula-
tion. Information to help adjust data for the future can be obtained from sales
forecasts, market trends, or business plans.
Prior to developing a representation of the data, the data should be analyzed
to ascertain their suitability for use in the simulation model. Data characteristics
such as independence (randomness), homogeneity (the data come from the same
har01307_ch05_115-160.indd 125 20/12/10 2:01 PM

TABLE 5.2 100 Observed Inspection Times (Minutes)

0.99 0.41 0.89 0.59 0.98 0.47 0.70 0.94 0.39 0.92
1.30 0.67 0.64 0.88 0.57 0.87 0.43 0.97 1.20 1.50
1.20 0.98 0.89 0.62 0.97 1.30 1.20 1.10 1.00 0.44
0.67 1.70 1.40 1.00 1.00 0.88 0.52 1.30 0.59 0.35
0.67 0.51 0.72 0.76 0.61 0.37 0.66 0.75 1.10 0.76
0.79 0.78 0.49 1.10 0.74 0.97 0.93 0.76 0.66 0.57
1.20 0.49 0.92 1.50 1.10 0.64 0.96 0.87 1.10 0.50
0.60 1.30 1.30 1.40 1.30 0.96 0.95 1.60 0.58 1.10
0.43 1.60 1.20 0.49 0.35 0.41 0.54 0.83 1.20 0.99
1.00 0.65 0.82 0.52 0.52 0.80 0.72 1.20 0.59 1.60
FIGURE 5.3
Descriptive statistics
for a sample data set
of 100 observations.
distribution), and stationarity (the distribution of the data doesn’t change with time)
should be determined. Using data analysis software such as Stat::Fit, data sets can
be automatically analyzed, tested for usefulness in a simulation, and matched to the
best-fitting underlying distribution. Stat::Fit is bundled with ProModel and can be
accessed from the Tools menu after opening ProModel. To illustrate how data are
analyzed and converted to a form for use in simulation, let’s take a look at a data set
containing 100 observations of an inspection operation time, shown in Table 5.2.
By entering these data or importing them from a file into Stat::Fit, a descrip-
tive analysis of the data set can be performed. The summary of this analysis is
displayed in Figure 5.3. These parameters describe the entire sample collection.
Because reference will be made later to some of these parameters, a brief defini-
tion of each is given below.
Mean—the average value of the data.
Median—the value of the middle observation when the data are sorted in
ascending order.
Mode—the value that occurs with the greatest frequency.
har01307_ch05_115-160.indd 126 20/12/10 2:01 PM

Standard deviation—a measure of the average deviation from the mean.

Variance—the square of the standard deviation.
Coefficient of variation—the standard deviation divided by the mean. This
value provides a relative measure of the standard deviation to the mean.
Skewness—a measure of symmetry. If the maximum data value is farther
from the mean than the minimum data value, the data will be skewed to
the right, or positively skewed. Otherwise they are skewed to the left, or
negatively skewed.
Kurtosis—a measure of the flatness or peakedness of the distribution.
Another value of interest is the range, which is simply the difference between the
maximum and minimum values.
A descriptive analysis tells us key characteristics about the data set, but it does
not tell us how suitable the data are for use in a simulation model. What we would
like to be able to do is fit a theoretical distribution, such as a normal or beta dis-
tribution, to the sample data set. This requires that the data be independent (truly
random) and identically distributed (they all come from the same distribution).
When gathering data from a dynamic and possibly time-varying system, one must
be sensitive to trends, patterns, and cycles that may occur with time. The sam-
ple data points may be somewhat codependent, or they may come from multiple
sources and therefore may not be identically distributed. Specific tests can be per-
formed to determine whether data are independent and identically distributed.
5.7.1 Tests for Independence

Data are independent if the value of one observation is not influenced by the value
of another observation. Dependence is common in sampling from a finite popula-
tion when sampling is performed without replacement. For example, in sampling
from a deck of cards to find the probability of drawing an ace, the probability of
the next card being an ace increases with each draw that is not an ace. The prob-
ability of drawing an ace in the first draw is __ 4
52
. If no ace is drawn, the probability
of drawing an ace in the second draw is __ 4
51
. If, however, the drawn card is replaced
each time, the probability remains the same (__ 4
52
) for each draw.
For very large or infinite populations such as ongoing activities or repeating
events, data dependence usually manifests itself when the value of an observation
is influenced by the value of the previous observation. In sampling the probability
that an item will be defective, for example, sometimes there is a greater prob-
ability of a defective item occurring if the previous item was defective (defectives
frequently occur in waves). Perhaps the time to perform a manual cutting opera-
tion increases each time due to the blade becoming dull. Finally, when the blade
is sharpened, the activity time is reduced back to the original starting time. When
the value of an observation is dependent on the value of a previous observation,
the data are said to be correlated. Table 5.3 shows outdoor temperature recordings
taken randomly throughout the day from 8:00 A.M. to 8:00 P.M. (read left to right).
As would be expected, sampled temperatures are not independent because they
depend largely on the time of day.
har01307_ch05_115-160.indd 127 20/12/10 2:01 PM

TABLE 5.3 100 Outdoor Temperature Readings from 8:00 A.M. to 8:00 P.M.
57 57 58 59 59 60 60 62 62 62
63 63 64 64 65 66 66 68 68 69
70 71 72 72 73 74 73 74 75 75
75 75 76 77 78 78 79 80 80 81
80 81 81 82 83 83 84 84 83 84
83 83 83 82 82 81 81 80 81 80
79 79 78 77 77 76 76 75 74 75
75 74 73 73 72 72 72 71 71 71
71 70 70 70 70 69 69 68 68 68
67 67 66 66 66 65 66 65 65 64
Several techniques are used to determine data dependence or correlation.

Three techniques discussed here include the scatter plot, autocorrelation plot, and
runs test. Any of these procedures can be performed on data sets using Stat::Fit.
They are each specific to different scales of randomness, so all must be passed to
conclude data independence. The scatter plot and autocorrelation plot are para-
metric statistical tests in which the nature of the plot depends on the underlying
distribution of the data. The runs test is nonparametric and makes no assumption
about the distribution of the data. We explain how each is used.
Scatter Plot
This is a plot of adjacent points in the sequence of observed values plotted against
each other. Thus each plotted point represents a pair of consecutive observations
(Xi, Xi1) for i 1, 2, . . . , n 1. This procedure is repeated for all adjacent data
points so that 100 observations would result in 99 plotted points. If the Xi’s are
independent, the points will be scattered randomly. If, however, the data are de-
pendent on each other, a trend line will be apparent. If the Xi’s are positively cor-
related, a positively sloped trend line will appear. Negatively correlated Xi’s will
produce a negatively sloped line. A scatter plot is a simple way to detect strongly
dependent behavior.
Figure 5.4 shows a plot of paired observations using the 100 inspection times
from Table 5.2. Notice that the points are scattered randomly with no defined pat-
tern, indicating that the observations are independent.
Figure 5.5 is a plot of the 100 temperature observations shown in Table 5.3.
Notice the strong positive correlation.
Autocorrelation Plot
If observations in a sample are independent, they are also uncorrelated. Correlated
data are dependent on each other and are said to be autocorrelated. A measure of
the autocorrelation, rho (), can be calculated using the equation
_ _
(xi x)(xij x)
nj

il
______________
2(n j)
har01307_ch05_115-160.indd 128 20/12/10 2:01 PM

FIGURE 5.4
Scatter plot showing
uncorrelated data.
FIGURE 5.5
Scatter plot showing
correlated temperature
data.
where j is the lag or distance between data points; is the standard deviation of
–
the population, approximated by the standard deviation of the sample; and X is
the sample mean. The calculation is carried out to _5 of the length of the data set,
1
where diminishing pairs start to make the calculation unreliable.

This calculation of autocorrelation assumes that the data are taken from a
stationary process; that is, the data would appear to come from the same distribu-
tion regardless of when the data were sampled (that is, the data are time invariant).
har01307_ch05_115-160.indd 129 20/12/10 2:01 PM

FIGURE 5.6
Autocorrelation
graph showing
noncorrelation.
FIGURE 5.7
Autocorrelation graph
showing correlation.
In the case of a time series, this implies that the time origin may be shifted without
affecting the statistical characteristics of the series. Thus the variance for the whole
sample can be used to represent the variance of any subset. If the process being
studied is not stationary, the calculation of autocorrelation is more complex.
The autocorrelation value varies between 1 and 1 (that is, between positive
and negative correlation). If the autocorrelation is near either extreme, the data
are autocorrelated. Figure 5.6 shows an autocorrelation plot for the 100 inspection
time observations from Table 5.2. Notice that the values are near zero, indicating
har01307_ch05_115-160.indd 130 20/12/10 2:01 PM

FIGURE 5.8
Runs test based on
points above and
below the median and
number of turning
points.
little or no correlation. The numbers in parentheses below the x axis are the maxi-
mum autocorrelation in both the positive and negative directions.
Figure 5.7 is an autocorrelation plot for the sampled temperatures in Table 5.3.
The graph shows a broad autocorrelation.
Runs Tests
The runs test looks for runs in the data that might indicate data correlation. A
run in a series of observations is the occurrence of an uninterrupted sequence of
numbers showing the same trend. For instance, a consecutive set of increasing or
decreasing numbers is said to provide runs “up” or “down” respectively. Two types
of runs tests that can be made are the median test and the turning point test. Both
of these tests can be conducted automatically on data using Stat::Fit. The runs test
for the 100 sample inspection times in Table 5.2 is summarized in Figure 5.8. The
har01307_ch05_115-160.indd 131 20/12/10 2:01 PM

result of each test is either do not reject the hypothesis that the series is random or
reject that hypothesis with the level of significance given. The level of significance
is the probability that a rejected hypothesis is actually true—that is, that the test
rejects the randomness of the series when the series is actually random.
The median test measures the number of runs—that is, sequences of num-
bers, above and below the median. The run can be a single number above or below
the median if the numbers adjacent to it are in the opposite direction. If there are
too many or too few runs, the randomness of the series is rejected. This median
runs test uses a normal approximation for acceptance or rejection that requires
that the number of data points above or below the median be greater than 10.
The turning point test measures the number of times the series changes direc-
tion (see Johnson, Kotz, and Kemp 1992). Again, if there are too many turning
points or too few, the randomness of the series is rejected. This turning point runs
test uses a normal approximation for acceptance or rejection that requires more
than 12 data points.
While there are many other runs tests for randomness, some of the most sen-
sitive require larger data sets, in excess of 4,000 numbers (see Knuth 1981). The
number of runs in a series of observations indicates the randomness of those ob-
servations. A few runs indicate strong correlation, point to point. Several runs may
indicate cyclic behavior.
5.7.2 Tests for Identically Distributed Data

When collecting sample data, often it is necessary to determine whether the data
in a single data set come from the same population (they are identically distrib-
uted) or whether they represent multiple populations. In other instances, it is de-
sirable to determine whether two or more data sets can be treated as having come
from the same population or whether they need to be kept as separate populations.
There are several ways to determine whether data come from the same popula-
tion. These are sometimes called tests for homogeneity.
Let’s look at the first case of determining whether data from a single data
set are identically distributed. The first inclination when gathering data defining
a particular activity time or event interval is to assume that they are identically
distributed. Upon closer examination, however, it is sometimes discovered that
they are actually nonhomogeneous data. Examples of data sets that tend to be
nonhomogeneous include
• Activity times that are longer or shorter depending on the type of entity
being processed.
• Interarrival times that fluctuate in length depending on the time of day or
day of the week.
• Time between failures and time to repair where the failure may result
from a number of different causes.
In each of these instances, the data collected may actually represent multiple
populations.
har01307_ch05_115-160.indd 132 20/12/10 2:01 PM

FIGURE 5.9 Part jams Mechanical failures

Bimodal distribution of
Frequency of
downtimes indicating
occurrence
multiple causes.
Repair time
Testing to see if data are identically distributed can be done in several ways.
One approach is to visually inspect the distribution to see if it has more than
one mode. You may find, for example, that a plot of observed downtime values
is bimodal, as shown in Figure 5.9. The fact that there are two clusters of data
indicates that there are at least two distinct causes of downtimes, each producing
different distributions for repair times. Perhaps after examining the cause of the
downtimes, it is discovered that some were due to part jams that were quickly
fixed while others were due to mechanical failures that took longer to repair.
One type of nonhomogeneous data occurs when the distribution changes over
time. This is different from two or more distributions manifesting themselves over
the same time period such as that caused by mixed types of downtimes. An ex-
ample of a time-changing distribution might result from an operator who works
20 percent faster during the second hour of a shift than during the first hour.
Over long periods of time, a learning curve phenomenon occurs where workers
perform at a faster rate as they gain experience on the job. Such distributions
are called nonstationary or time variant because of their time-changing nature. A
common example of a distribution that changes with time is the arrival rate of cus-
tomers to a service facility. Customer arrivals to a bank or store, for example, tend
to occur at a rate that fluctuates throughout the day. Nonstationary distributions
can be detected by plotting subgroups of data that occur within successive time
intervals. For example, sampled arrivals between 8 A.M. and 9 A.M. can be plotted
separately from arrivals between 9 A.M. and 10 A.M., and so on. If the distribution
is of a different type or is the same distribution but shifted up or down such that
the mean changes value over time, the distribution is nonstationary. This fact will
need to be taken into account when defining the model behavior. Figure 5.10 is a
plot of customer arrival rates for a department store occurring by half-hour inter-
val between 10 A.M. and 6 P.M. Note that while the type of distribution (Poisson) is
the same for each period, the rate (and hence the mean interarrival time) changes
every half hour.
The second case we look at is where two sets of data have been gathered and
we desire to know whether they come from the same population or are identically
distributed. Situations where this type of testing is useful include the following:
• Interarrival times have been gathered for different days and you want to
know if the data collected for each day come from the same distribution.
har01307_ch05_115-160.indd 133 20/12/10 2:01 PM

FIGURE 5.10
Change in rate of
Rate of arrival
customer arrivals
between 10 A.M. and
6 P.M.
10:00 A.M. 12:00 noon 2:00 P.M. 4:00 P.M. 6:00 P.M.
Time of day
• Activity times for two different operators were collected and you want to
know if the same distribution can be used for both operators.
• Time to failure has been gathered on four similar machines and you are
interested in knowing if they are all identically distributed.
One easy way to tell whether two sets of data have the same distribution is to run
Stat::Fit and see what distribution best fits each data set. If the same distribution
fits both sets of data, you can assume that they come from the same population. If
in doubt, they can simply be modeled as separate distributions.
Several formal tests exist for determining whether two or more data sets can
be assumed to come from identical populations. Some of them apply to specific
families of distributions such as analysis of variance (ANOVA) tests for normally
distributed data. Other tests are distribution independent and can be applied to
compare data sets having any distribution, such as the Kolmogorov-Smirnov two-
sample test and the chi-square multisample test (see Hoover and Perry 1990). The
Kruskal-Wallis test is another nonparametric test because no assumption is made
about the distribution of the data (see Law 2007).
5.8 Distribution Fitting

Once numerical data have been tested for independence and correlation, they can
be converted to a form suitable for use in the simulation model. Sample numerical
data that have been gathered on activity times, arrival intervals, batch quantities,
and so forth can be represented in a simulation model in one of three ways. First,
the data can be used exactly the way they were recorded. Second, an empiri-
cal distribution can be used that characterizes the data. The third and preferred
method is to select a theoretical distribution that best fits the data.
Using the data exactly as they were recorded is relatively straightforward but
problematic. During the simulation, data values are read directly from the sample
data file in the order in which they were collected. The problem with this tech-
nique is that the data set is usually not large enough to represent the population. It
also presents a problem for running multiple replications of a simulation experi-
ment because a separate data set, which usually isn’t available, is needed for each
har01307_ch05_115-160.indd 134 20/12/10 2:01 PM

replication. Sample data sets tend to be used directly in a simulation when they
are sampled data from another related simulation model from which it receives its
input. This is a model partitioning technique sometimes used in simulating large
systems. The exit times for entities in one simulation become the arrival stream
used for the model into which it feeds.
Using an empirical distribution requires that the data be converted to either
a continuous or discrete frequency distribution. A continuous frequency distribu-
tion summarizes the percentage of values that fall within given intervals. In the
case of a discrete frequency distribution, it is the percentage of times a particular
value occurs. For continuous frequency distributions the intervals need not be of
equal width. During the simulation, random variates are generated using a con-
tinuous, piecewise-linear empirical distribution function based on the grouped
data (see Law 2007). The drawbacks to using an empirical distribution as input to
a simulation are twofold. First, an insufficient sample size may create an artificial
bias or “choppiness” in the distribution that does not represent the true underlying
distribution. Second, empirical distributions based on a limited sample size often
fail to capture rare extreme values that may exist in the population from which
they were sampled. As a general rule, empirical distributions should be used only
for rough-cut modeling or when the shape is very irregular and doesn’t permit a
good distribution fit. Within ProModel, an empirical distribution is called a user
distribution because the distribution is defined by the user rather than by selecting
a distribution from a menu of predefined theoretical distributions.
Representing the data using a theoretical distribution involves fitting a theo-
retical distribution to the data. During the simulation, random variates are gener-
ated from the probability distribution to provide the simulated random values.
Fitting a theoretical distribution to sample data smooths artificial irregularities
in the data and ensures that extreme values are included. For these reasons, it is
best to use theoretical distributions if a reasonably good fit exists. Most popular
simulation software provide utilities for fitting distributions to numerical data,
thus relieving the modeler from performing this complicated procedure. A mod-
eler should be careful when using a theoretical distribution to ensure that if an
unbounded distribution is used, the extreme values that can be generated are
realistic. Techniques for controlling the range of unbounded distributions in a
simulation model are presented in Chapter 6. Appendix A provides a list of the
theoretical distributions available in ProModel.
5.8.1 Frequency Distributions

Regardless of whether an empirical or theoretical distribution is used, it is a good
idea to construct a frequency distribution to get a summary view of the data. Then
a fitness test can be conducted to evaluate whether a standard theoretical distribu-
tion, such as a normal distribution, fits the data. A frequency distribution groups or
bins the data into intervals or cells according to frequency of occurrence. Showing
the frequency with which values occur within each cell gives a clear idea of how
observations are distributed throughout the range of the data. Frequency distribu-
tions may be either discrete or continuous.
har01307_ch05_115-160.indd 135 20/12/10 2:01 PM

TABLE 5.4 Data for Customer Arrival at a Restaurant

Arrivals Arrivals
per 5-Minute Interval Frequency per 5-Minute Interval Frequency
0 15 6 7
1 11 7 4
2 19 8 5
3 16 9 3
4 8 10 3
5 8 11 1
Discrete Frequency Distributions

A discrete frequency distribution is limited to specific values. That is, only fre-
quencies of a finite set of specific values are shown. To illustrate a discrete fre-
quency distribution, suppose a sample is collected on the number of customer
entries to a fast-food restaurant per five-minute interval. Table 5.4 summarizes
these observations, which were made during lunchtime (between 11 A.M. and
1 P.M.) for five workdays over a 20-week period.
Frequency distributions can be graphically shown using a histogram. A
histogram depicting the discrete frequency distribution in Table 5.4 is shown in
Figure 5.11.
Continuous Frequency Distributions

A continuous frequency distribution defines ranges of values within which sample
values fall. Going back to our inspection time sample consisting of 100 observa-
tions, we can construct a continuous distribution for these data because values can
FIGURE 5.11
Histogram showing arrival count per five-minute interval.
20
15
Frequency
10
0
0 1 2 3 4 5 6 7 8 9 10 11
Arrival count
har01307_ch05_115-160.indd 136 20/12/10 2:01 PM

FIGURE 5.12
Frequency distribution for 100 observed inspection times.
take on any value within the interval specified. A frequency distribution for the
data has been constructed using Stat::Fit (Figure 5.12).
Note that a relative frequency or density is shown (third column) as well as
cumulative (ascending and descending) frequencies. All of the relative densities
add up to 1, which is verified by the last value in the ascending cumulative fre-
quency column.
A histogram based on this frequency distribution was also created in Stat::Fit
and is shown in Figure 5.13.
Note that the frequency distribution and histogram for our sample inspection
times are based on dividing the data into six even intervals or cells. While there
are guidelines for determining the best interval or cell size, the most important
thing is to make sure that enough cells are defined to show a gradual transition
in values, yet not so many cells that groupings become obscured. The number of
intervals should be based on the total number of observations and the variance in
the data. One rule of thumb is to set the number
1
of intervals to the cube root of
__
twice the number of samples—that is (2N)3. This is the default method used in
Stat::Fit. The goal is to use the minimum number of intervals possible without
losing information about the spread of the data.
har01307_ch05_115-160.indd 137 20/12/10 2:01 PM

FIGURE 5.13
Histogram distribution
for 100 observed
inspection times.
5.8.2 Theoretical Distributions

As mentioned earlier, it may be appropriate to use the frequency distribution in
the simulation model as it is, as an empirical distribution. Whenever possible,
however, the underlying distribution from which the sample data came should
be determined. This way the actual distribution of the population can be used to
generate random variates in the simulation rather than relying only on a sampling
of observations from the population. Defining the theoretical distribution that best
fits sample data is called distribution fitting. Before discussing how theoretical
distributions are fit to data, it is helpful to have a basic understanding of at least
the most common theoretical distributions.
There are about 12 statistical distributions that are commonly used in simu-
lation (Banks and Gibson 1997). Theoretical distributions can be defined by a
simple set of parameters usually defining dispersion and density. A normal distri-
bution, for example, is defined by a mean value and a standard deviation value.
Theoretical distributions are either discrete or continuous, depending on whether
a finite set of values within the range or an infinite continuum of possible values
within a range can occur. Discrete distributions are seldom used in manufacturing
and service system simulations because they can usually be defined by simple
probability expressions. Below is a description of a few theoretical distributions
sometimes used in simulation. These particular ones are presented here purely
because of their familiarity and ease of understanding. Beginners to simulation
usually feel most comfortable using these distributions, although the precautions
given for their use should be noted. An extensive list of theoretical distributions
and their applications is given in Appendix A.
har01307_ch05_115-160.indd 138 20/12/10 2:01 PM

Binomial Distribution
The binomial distribution is a discrete distribution that expresses the probability
( p) that a particular condition or outcome can occur in n trials. We call an occur-
rence of the outcome of interest a success and its nonoccurrence a failure. For
a binomial distribution to apply, each trial must be a Bernoulli trial: it must be
independent and have only two possible outcomes (success or failure), and the
probability of a success must remain constant from trial to trial. The mean of a
binomial distribution is given by np, where n is the number of trials and p is the
probability of success on any given trial. The variance is given by np(1 p).
A common application of the binomial distribution in simulation is to test for
the number of defective items in a lot or the number of customers of a particular
type in a group of customers. Suppose, for example, it is known that the probabil-
ity of a part being defective coming out of an operation is 0.10 and we inspect the
parts in batch sizes of 10. The number of defectives for any given sample can be
determined by generating a binomial random variate. The probability mass func-
tion for the binomial distribution in this example is shown in Figure 5.14.
Uniform Distribution
A uniform or rectangular distribution is used to describe a process in which the
outcome is equally likely to fall between the values of a and b. In a uniform dis-
tribution, the mean is (a b)2. The variance is expressed by (b a)212. The
probability density function for the uniform distribution is shown in Figure 5.15.
The uniform distribution is often used in the early stages of simulation proj-
ects because it is a convenient and well-understood source of random variation. In
FIGURE 5.14 0.6
The probability mass
function of a binomial 0.5
distribution 0.4
(n 10, p 0.10).
p(x)
0.3
0.2
0.1
0
0 1 2 3 4 5 6 7 8 9 10
x
FIGURE 5.15
The probability density
function of a uniform f (x)
distribution.
a b
x
har01307_ch05_115-160.indd 139 20/12/10 2:01 PM

FIGURE 5.16
The probability
density function of a
triangular distribution.
f (x)
a m b
x
the real world, it is extremely rare to find an activity time that is uniformly distrib-
uted because nearly all activity times have a central tendency or mode. Sometimes
a uniform distribution is used to represent a worst-case test for variation when
doing sensitivity analysis.
Triangular Distribution
A triangular distribution is a good approximation to use in the absence of data,
especially if a minimum, maximum, and most likely value (mode) can be esti-
mated. These are the three parameters of the triangular distribution. If a, m, and b
represent the minimum, mode, and maximum values respectively of a triangular
distribution, then the mean of a triangular distribution is (a m b)3. The vari-
ance is defined by (a2 m2 b2 am ab mb)18. The probability density
function for the triangular distribution is shown in Figure 5.16.
The weakness of the triangular distribution is that values in real activity times
rarely taper linearly, which means that the triangular distribution will probably
create more variation than the true distribution. Also, extreme values that may be
rare are not captured by a triangular distribution. This means that the full range
of values of the true distribution of the population may not be represented by the
triangular distribution.
Normal Distribution
The normal distribution (sometimes called the Gaussian distribution) describes
phenomena that vary symmetrically above and below the mean (hence the bell-
shaped curve). While the normal distribution is often selected for defining activity
times, in practice manual activity times are rarely ever normally distributed. They
are nearly always skewed to the right (the ending tail of the distribution is longer
than the beginning tail). This is because humans can sometimes take significantly
longer than the mean time, but usually not much less than the mean time. Examples
of normal distributions might be
• Physical measurements—height, length, diameter, weight.
• Activities involving multiple tasks (like loading a truck or filling a
customer order).
har01307_ch05_115-160.indd 140 20/12/10 2:01 PM

FIGURE 5.17
The probability density
function for a normal
distribution.
f (x)
The mean of the normal distribution is designated by the Greek letter mu (). The
variance is 2 where (sigma) is the standard deviation. The probability density
function for the normal distribution is shown in Figure 5.17.
Exponential Distribution
Sometimes referred to as the negative exponential, this distribution is used fre-
quently in simulations to represent event intervals. The exponential distribution
is defined by a single parameter, the mean (). This distribution is related to the
Poisson distribution in that if an occurrence happens at a rate that is Poisson
distributed, the time between occurrences is exponentially distributed. In other
words, the mean of the exponential distribution is the inverse of the Poisson rate.
For example, if the rate at which customers arrive at a bank is Poisson distributed
with a rate of 12 per hour, the time between arrivals is exponentially distributed
with a mean of 5 minutes. The exponential distribution has a memoryless or for-
getfulness property that makes it well suited for modeling certain phenomena that
occur independently of one another. For example, if arrival times are exponen-
tially distributed with a mean of 5 minutes, then the expected time before the next
arrival is 5 minutes regardless of how much time has elapsed since the previous
arrival. It is as though there is no memory of how much time has already elapsed
when predicting the next event—hence the term memoryless. Examples of this
distribution are
• Time between customer arrivals at a bank.
• Duration of telephone conversations.
• Time between arrivals of planes at a major airport.
• Time between failures of certain types of electronic devices.
• Time between interrupts of the CPU in a computer system.
har01307_ch05_115-160.indd 141 20/12/10 2:01 PM

FIGURE 5.18
The probability
density function
for an exponential
distribution.
f(x)
For an exponential distribution, the standard deviation is the same as the

mean. The probability density function of the exponential distribution is shown in
Figure 5.18.
5.8.3 Fitting Theoretical Distributions to Data

Fitting a theoretical distribution to data is essentially an attempt to identify the
underlying distribution from which the data were generated. Finding the best dis-
tribution that fits the sample data can be quite complicated and is not an exact
science. Fortunately, software such as Stat::Fit (accessible under the ProModel
Tools menu) is available for automatically fitting distributions. Lab 5 at the end
of the book provides exercises on how to use Stat::Fit. Distribution fitting soft-
ware follows basically the same procedures as we discuss next for manual fitting.
Many software tools provide the added benefit of being able to iterate through
all applicable distributions searching for the best fit. Figure 5.19 shows a relative
ranking in Stat::Fit of how well different distributions fit the observed inspection
times in Table 5.2.
Distribution fitting is largely a trial-and-error process. The basic procedure
consists of three steps: (1) one or more distributions are selected as candidates for
being good fits to the sample data; (2) estimates of the parameters for each distri-
bution must be calculated; and (3) goodness-of-fit tests are performed to ascertain
how well each distribution fits the data.
Choosing a distribution that appears to be a good fit to the sample data requires
a basic knowledge of the types of distributions available and their properties. It
is also helpful to have some intuition about the variable whose data are being fit.
Sometimes creating a histogram of the data can reveal important characteristics
about the distribution of the data. If, for example, a histogram of sampled as-
sembly times appears symmetric on either side of the mean, a normal distribution
har01307_ch05_115-160.indd 142 20/12/10 2:01 PM

FIGURE 5.19
Ranking distributions by goodness of fit for inspection time data set.
may be inferred. Looking at the histogram in Figure 5.13, and knowing the basic
shapes of different theoretical distributions, we might hypothesize that the data
come from a beta, lognormal, or perhaps even a triangular distribution.
After a particular type of distribution has been selected, we must estimate
the defining parameter values of the distribution based on the sample data. In the
case of a normal distribution, the parameters to be estimated are the mean and
standard deviation, which can be estimated by simply calculating the average
and standard deviation of the sample data. Parameter estimates are generally cal-
culated using moment equations or the maximum likelihood equation (see Law
2007).
Once a candidate distribution with its associated parameters has been defined,
a goodness-of-fit test can be performed to evaluate how closely the distribution fits
the data. A goodness-of-fit test measures the deviation of the sample distribution
from the inferred theoretical distribution. Three commonly used goodness-of-fit
tests are the chi-square (2), the Kolmogorov-Smirnov, and the Anderson-Darling
tests (see Breiman 1973; Banks et al. 2001; Law 2007; Stuart and Ord 1991).
Each of these goodness-of-fit tests starts with the null hypothesis that the fit
is good and calculates a test statistic for comparison to a standard. To test this
har01307_ch05_115-160.indd 143 20/12/10 2:01 PM

null hypothesis, a level of significance value is selected, which is the likelihood of

making a type I error—that is, rejecting the null hypothesis when it is true. Stated
in a different manner, it is the probability that you will make a mistake and reject
a distribution that is actually a good fit. Therefore, the smaller this value, the less
likely you are to reject when you should accept.
The oldest and one of the most common statistical tests for goodness of fit is
the chi-square test. The chi-square, goodness-of-fit test is a versatile test that can
be used to perform hypothesis tests for either discrete or continuous distributions.
It is also a useful test to determine data independence (see Chapter 3). The chi-
square test basically compares the frequency distribution of the sample data to the
way the same number of data points would be distributed if the data came from
the inferred distribution. The chi-square, goodness-of-fit test can be broken down
into nine steps:
1. Analyze the data and infer an underlying distribution.
2. Create a frequency distribution of the data with equiprobable cells based
on the inferred distribution.
3. Calculate the expected frequency (ei) for each cell.
4. Adjust cells if necessary so that all expected frequencies are at least 5.
5. Calculate the chi-square test statistic.
6. Determine the number of degrees of freedom (k 1).
7. Choose a desired level of significance ().
8. Find the critical chi-square value from the chi-square table (Appendix D).
9. Reject the distribution if the chi-square statistic exceeds the critical
value.
To illustrate how the chi-square, goodness-of-fit test is performed, let’s look at an
example problem in which 40 observations were made of activity times for a bag-
ging operation. The duration in seconds of each observation is shown in Table 5.5.
What we would like to know is what distribution best fits the data.
Step 1: Analyze the data and infer an underlying distribution. A
distribution should be chosen as a candidate for the true, underlying
distribution that has similar characteristics to those exhibited by the data.
A descriptive analysis, like the one described for the data in Table 5.2, can
help reveal some of the defining characteristics of the data. This includes
calculating parameters such as the mean, range, standard deviation, and
so on. If, for example, the mean and median are roughly the same, and the
standard deviation is roughly one-sixth of the range, we might infer that
the data are normally distributed. Constructing a histogram of the data like
that shown in Figure 5.13 can also be a good way to match a distribution
to the data.
For our example, the data appear to be evenly distributed throughout
the range, so we will hypothesize that the observations are uniformly
distributed between 7 and 17 seconds. This is our null hypothesis (H0).
har01307_ch05_115-160.indd 144 20/12/10 2:01 PM

TABLE 5.5 40 Observed Bagging Times (seconds)

11.3 8.2 13.8 10.3
7.2 8.6 15.2 9.6
12.5 7.4 13.5 11.1
14.3 11.1 9.2 11.8
12.8 12.3 16.3 9.5
10.2 16.8 14.9 16.3
7.7 15.2 12.9 12.4
11.0 8.3 14.3 16.9
13.2 14.5 7.5 13.2
14.4 10.7 15.1 10.7
TABLE 5.6 Frequency Distribution for Bagging Times

Cell Interval Observed H0 Probability H0 Expected (oi ei) 2
________
(i) (seconds) Frequency (oi) (pi) Frequency (ei) ei
1 7–9 7 .20 8 0.125

2 9–11 7 .20 8 0.125
3 11–13 10 .20 8 0.50
4 13–15 8 .20 8 0
5 15–17 8 .20 8 0
The alternate hypothesis (H1) is that the observations are not uniformly
distributed between 7 and 17 seconds.
Step 2: Create a frequency distribution of the data with equiprobable
cells based on the inferred distribution. To create a frequency
distribution, recall that a good 1rule of thumb for determining the appropriate
__
number of cells (k) is k (2n)3. Normally a frequency distribution is
divided into cells of equal intervals. When conducting a chi-square test,
however, it often works better if cell intervals have equal probabilities rather
than equal interval ranges. If, for example, you want to create a frequency
distribution with six cells for a particular data set, and you suspect that
the data come from a normal distribution, you would create equiprobable
intervals that each take in one-sixth of the population of the corresponding
normal distribution. 1 1
__ __
For this example, k (2n)3 (2 40)3 4.3 5. Because we are
assuming a uniform distribution in which equal interval widths also have equal
probabilities, we can divide the range into five equal intervals each with a
cell width of 105 2 and a probability of 15 .20. For each cell (i), the
resultant observed frequency (oi) is shown in Table 5.6 together with the
hypothesized probability (pi). That each cell has a probability of .2 based
har01307_ch05_115-160.indd 145 20/12/10 2:01 PM

on a cell width of 2 can be verified by calculating the probability of the first

interval as

9 9
p(7 x 9) f (x)dx ___ x 9 ___
1 dx ___ 9 ___
7 ___
2 .20.
7 7 10 10 7 10 10 10
For a uniform distribution the probabilities for all intervals are equal, so the
remaining intervals also have a hypothesized probability of .20.
Step 3: Calculate the expected frequency for each cell (ei). The expected
frequency (e) for each cell (i) is the expected number of observations that
would fall into each interval if the null hypothesis were true. It is calculated
by multiplying the total number of observations (n) by the probability
(p) that an observation would fall within each cell. So for each cell, the
expected frequency (ei) equals npi.
In our example, since the hypothesized probability (p) of each cell is
the same, the expected frequency for every cell is ei npi 40 .2 8.
Step 4: Adjust cells if necessary so that all expected frequencies are
at least 5. If the expected frequency of any cell is less than 5, the cells
must be adjusted. This “rule of five” is a conservative rule that provides
satisfactory validity of the chi-square test. When adjusting cells, the easiest
approach is to simply consolidate adjacent cells. After any consolidation,
the total number of cells should be at least 3; otherwise you no longer
have a meaningful differentiation of the data and, therefore, will need to
gather additional data. If you merge any cells as the result of this step, you
will need to adjust the observed frequency, hypothesized probability, and
expected frequency of those cells accordingly.
In our example, the expected frequency of each cell is 8, which meets
the minimum requirement of 5, so no adjustment is necessary.
Step 5: Calculate the chi-square statistic. The equation for calculating
the chi-square statistic is 2calc i1 (oi ei)2ei. If the fit is good the chi-
k
square statistic will be small.

For our example,
(oi ei)2
5
2calc ________
i1
e .125 .125 .50 0 0 0.75.
i
Step 6: Determine the number of degrees of freedom (k ⴚ 1). A simple

way of determining a conservative number of degrees of freedom is to take
the number of cells minus 1, or k 1.
For our example, the number of degrees of freedom is
k 1 5 1 4.
Note: The number of degrees of freedom is often computed to be
k s 1, where k is the number of cells and s is the number of parameters
estimated from the data for defining the distribution. So for a normal
har01307_ch05_115-160.indd 146 20/12/10 2:01 PM

distribution with two parameters (mean and standard deviation) that are
estimated from the data, s 2. In Stat::Fit the number of degrees of
freedom is assumed to be k 1.
Step 7: Choose a desired level of significance (␣). The significance level is
the probability of rejecting the null hypothesis when it is true—that is, when
the hypothesized distribution is correct. A typical level of significance is
.05; however, it could also be .10 or any other desired value.
For our example we will use a level of significance () of .05.
Step 8: Find the critical chi-square value from the chi-square table. A
chi-square table can be found in Appendix D. The critical values in the chi-
square table are based on the number of degrees of freedom (k 1) and the
desired level of significance ().
For our example,
2kl, 24,0.05 9.488
Step 9: Reject the distribution if the chi-square statistic exceeds the

critical value. If the calculated test statistic is larger than the critical value
from the chi-square table, then the observed and expected values are not
close and the model is a poor fit to the data. Therefore, we reject H0 if
2calc 2kl,.
For our example, 0.75 is not greater than 9.488; therefore, we fail to
reject H0, which means we can assume that a uniform distribution is a good
fit. If we are satisfied with our find we can terminate our testing; otherwise,
we can return to step 2 and try another distribution.
Note: While the test statistic for the chi-square test can be useful, the
p-value is actually more useful in determining goodness of fit. The p-value
is defined as the probability that another sample will compare the same
as the current sample given that the fit is appropriate. A small p-value
indicates that another sample would not likely compare the same, and
therefore, the fit should be rejected. Conversely, a high p-value indicates
that another sample would likely compare the same, and, therefore, the fit
should not be rejected. Thus the higher the p-value, the more likely that the
fit is appropriate. When comparing two different distributions for goodness
of fit, the distribution with the higher p-value is likely to be the better fit
regardless of the level of significance. Most distribution-fitting software
uses the p-value for ranking goodness of fit.
If very few data are available, then a goodness-of-fit test is unlikely to re-
ject any candidate distribution. However, if a lot of data are available, then a
goodness-of-fit test may reject all candidate distributions. Therefore, failing to
reject a candidate distribution should be taken as one piece of evidence in favor
of the choice, while rejecting a particular distribution is only one piece of evi-
dence against the choice. Before making a decision, it is a good idea to look at
a graphical display that compares the theoretical distribution to a histogram of
har01307_ch05_115-160.indd 147 20/12/10 2:01 PM

FIGURE 5.20
Visual comparison
between beta
distribution and a
histogram of the 100
sample inspection time
values.
the data. Figure 5.20, for example, overlays a beta distribution (the distribution
with the highest fit ranking for this particular data set) on top of a histogram of
the data.
5.9 Selecting a Distribution in the Absence of Data

There are a couple of ways to obtain reasonable estimates of system behavior
when no historical records exist or when direct measurement isn’t possible. The
best source is the people who have expertise in some aspect of the system being
modeled. Operators, maintenance personnel, vendors, and others who are experi-
enced in similar processes can often give good approximations of what to expect
in new but similar situations. In addition to getting expert opinions, sometimes the
nature of the process itself can indicate what to expect. For example, in attempting
to select an appropriate distribution to represent a manual task for which a most
likely time has been estimated, experience has shown that such tasks tend to have
a triangular or beta distribution. As another example, customer interarrival times
tend to be exponentially distributed.
For numerical data such as activity times, estimates are usually given in one
of the following forms:
• A single most likely or mean value.
• Minimum and maximum values defining a range.
• Minimum, most likely, and maximum values.
har01307_ch05_115-160.indd 148 20/12/10 2:01 PM

5.9.1 Most Likely or Mean Value

Sometimes the only estimate that can be obtained for a system variable is a mean
or most likely value. This is a single “best guess” estimate of what the value is.
Examples of single-value estimates include
• About 10 customer arrivals per hour.
• Approximately 20 minutes to assemble parts in a kit.
• Around five machine failures each day.
A simple sensitivity test for these types of estimates would be to increase and
decrease the value by 20 percent to see how much of a difference it makes in
the simulation results. One of the problems with single-value estimates is that
no consideration is given to possible variation. Fluctuations in activity times can
have a significant impact on model performance. Sensitivity analysis should be
performed looking at the potential impact of variation. For example, if the esti-
mated time for a manual activity is 1.2 minutes, a test run might be made using
a normal distribution with a mean of 1.2 minutes and a somewhat large standard
deviation value of, say, .3 minute (the coefficient of variation .25). Comparing
this run with a run using a constant value of 1.2 minutes will help determine how
sensitive the model is to variation in the activity time.
5.9.2 Minimum and Maximum Values

Sometimes only the range of values can be given as an estimate for a particular
variable. Examples of minimum and maximum values include
• 1.5 to 3 minutes to inspect items.
• 5 to 10 customer arrivals per hour.
• Four to six minutes to set up a machine.
Often a uniform distribution is used to represent a range of values because no
central tendency is specified. The problem with using a uniform distribution is
that it tends to give unrealistic weight to extreme values and therefore exagger-
ates the random behavior beyond what it really is. A normal distribution may be
more appropriate in this situation because it recognizes that a higher density of
occurrences exists near the mean. If a normal distribution is used, the standard
deviation could be reasonably set at _16 of the range. This ensures that 99.73 per-
cent of all values will fall within the specified range. So for a time range of .25 to
1.75 minutes, a normal distribution could be used with a mean of 1 minute and a
standard deviation of .25 minute (see Figure 5.21).
5.9.3 Minimum, Most Likely, and Maximum Values

The best estimate to obtain for a random variable when historical data can’t be
obtained is the minimum, most likely, and maximum values. Because most activ-
ity times are not symmetrical about a mean, this type of estimate is probably going
har01307_ch05_115-160.indd 149 20/12/10 2:01 PM

FIGURE 5.21
Normal distribution with mean 1 and standard deviation .25.
.25 .5 .75 1 1.25 1.50 1.75
FIGURE 5.22
A triangular
distribution with
minimum 2,
mode 5, and
maximum 15.
2 5 15
to be the most realistic. Minimum, most likely, and maximum values closely re-
semble a triangular or a beta distribution. Examples of minimum, most likely, and
maximum values might be
• 2 to 15 with most likely 5 minutes to repair a machine.
• 3 to 5 with most likely 2.5 minutes between customer arrivals.
• 1 to 3 with most likely 1.5 minutes to complete a purchase order.
Minimum, most likely, and maximum values can be easily set up as a triangular
distribution. Looking at the first repair time example, if we fit these times to a
triangular distribution, it would produce a minimum of 2 minutes, a mode of
5 minutes and a maximum of 15 minutes (see Figure 5.22).
har01307_ch05_115-160.indd 150 20/12/10 2:01 PM

Because these values are still based only on estimates, it is a good idea to per-
form a sensitivity analysis, especially on the mode value (assuming the minimum
and maximum values represent best- and worst-case possibilities, respectively).
5.10 Bounded versus Boundless Distributions

Continuous theoretical distributions may be bounded (doubly or singly) or un-
bounded. An unbounded distribution can have values that go to either negative or
positive infinity. The normal distribution, for example, ranges from to and is
therefore unbounded. A bounded distribution may be bounded at both ends or only
on one end. The exponential distribution, for example, is a nonnegative distribution
(it can’t take on negative values) and is therefore bounded only on the lower end of
the distribution. The uniform and triangular distributions are doubly bounded, having
both lower and upper boundaries. Empirical continuous distributions are always dou-
bly bounded by the mere fact that they are defined by a finite number of cell intervals.
Sampling during a simulation from a boundless distribution may cause val-
ues to be generated that are unrealistic in real-world situations. An operation time
that is normally distributed with a mean of three minutes and a standard deviation
of one minute would have a possibility, for example, of being negative on the
low side as well as a remote chance of being absurdly large on the high side. It
is often desirable, therefore, to draw random samples during a simulation from a
truncated continuous distribution where either the lower or upper tail is cut short.
ProModel automatically converts negative values to 0 when an unbounded con-
tinuous distribution is specified for a time value. It has no way of knowing, how-
ever, what a reasonably acceptable upper bound should be. This must be specified
by the modeler. Lower and upper bounds can be imposed on a distribution by
simply screening for values outside of the acceptable range. Either values gener-
ated outside the acceptable range can be rejected and another value generated or
the generated value can be brought within the acceptable value range.
5.11 Modeling Discrete Probabilities

Using Continuous Distributions
Discrete probability distributions are used much less in simulation than continu-
ous probability distributions. Most discrete probabilities—such as the probability
of routing to a particular location or the probability of an item being defective—
are modeled using simple probability expressions like 0.34. In simulation, most
occasions for generating values from a discrete probability distribution involve
generating integer values rather than real values. For example, the discrete ran-
dom variable might represent the number of items in a batch. One useful way to
generate discretely distributed integer values is to use a continuous distribution
where fractions are simply truncated. In a manufacturing system, for example,
suppose that when filling containers or pallets there is sometimes a remaining
amount that fills only a partial container or pallet. If a full pallet holds 20 boxes,
har01307_ch05_115-160.indd 151 20/12/10 2:01 PM

the number of boxes on a partial pallet would likely be uniformly distributed

between 1 and 19. To model the partial pallet quantity, you would simply define
a continuous uniform distribution with a range between 1 and 20. A uniform ran-
dom variable X will have values generated for X where min X max. Thus val-
ues will never be equal to the maximum value (in this case 20). Because generated
values are automatically truncated when used in a context requiring an integer,
only integer values that are evenly distributed from 1 to 19 will occur (this is ef-
fectively a discrete uniform distribution).
5.12 Data Documentation and Approval

When it is felt that all relevant data have been gathered, analyzed, and converted
into a usable form, it is advisable to document the data using tables, relational
diagrams, and assumption lists. Sources of data should also be noted. This docu-
ment should then be reviewed by those in a position to evaluate the validity of
the data and approve the assumptions made. Where more formal documentation
is required, a separate document will need to be created. This document will be
helpful later if modifications need to be made to the model or to analyze why the
actual system ends up working differently than the simulation.
In addition to identifying data used to build the model, the document should
also specify factors that were deliberately excluded from the model because they
were deemed insignificant or irrelevant. For example, if break times are not in-
cluded in the system description because of their perceived insignificance, the
document should state this. Justification for omissions should also be included if
necessary. Stating why certain factors are being excluded from the system descrip-
tion will help resolve later questions that may arise regarding the model premises.
Reviewing and approving input data can be a time-consuming and difficult
task, especially when many assumptions are made. In practice, data validation
ends up being more of a consensus-building process where agreement is reached
that the information is good enough for the purposes of the simulation. The data
document is not a static document but rather a dynamic one that often changes
as model building and experimentation get under way. Much if not all of the data
documentation can be scripted right into the model. Most software provides the
capability to write comments in the model. Where more formal documentation is
required, a separate data document will need to be created.
5.12.1 Data Documentation Example

To illustrate how system data might be documented, imagine you have just col-
lected information for an assembly operation for three different monitor sizes.
Here is an example of how the data collected for this system might be dia-
grammed and tabulated. The diagram and data should be sufficiently clear and
complete for those familiar with the system to validate the data and for a model to
be constructed. Review the data and see if they are sufficiently clear to formulate
a mental image of the system being modeled.
har01307_ch05_115-160.indd 152 20/12/10 2:01 PM

Objective
The objective of the study is to determine station utilization and throughput of
the system.
Entity Flow Diagram

Rejected monitors
19" & 21"

monitor
Station 1 Station 2 Inspection
19", 21", & 25"
monitors 21" monitor
Reworked
monitors Station 3
Entities
19" monitor
21" monitor
25" monitor
Workstation Information
Workstation Buffer Capacity Defective Rate
Station 1 5 5%
Station 2 8 8%
Station 3 5 0%
Inspection 5 0%
Processing Sequence
Operating Time in Minutes

Entity Station (min, mode, max)
19" monitor Station 1 0.8, 1, 1.5

Station 2 0.9, 1.2, 1.8
Inspection 1.8, 2.2, 3
21" monitor Station 1 0.8, 1, 1.5
Station 2 1.1, 1.3, 1.9
Inspection 1.8, 2.2, 3
25" monitor Station 1 0.9, 1.1, 1.6
Station 2 1.2, 1.4, 2
Inspection 1.8, 2.3, 3.2
Station 3 0.5, 0.7, 1
har01307_ch05_115-160.indd 153 12/20/10 9:58 PM

Handling Defective Monitors

• Defective monitors are detected at inspection and routed to whichever
station created the problem.
• Monitors waiting at a station for rework have a higher priority than first-
time monitors.
• Corrected monitors are routed back to inspection.
• A reworked monitor has only a 2 percent chance of failing again, in which
case it is removed from the system.
Arrivals
A cartload of four monitor assemblies arrives every four hours normally distrib-
uted with a standard deviation of 0.2 hour. The probability of an arriving monitor
being of a particular size is
Monitor Size Probability
19" .6
21" .3
25" .1
Move Times
All movement is on an accumulation conveyor with the following times:
From To Time (seconds)
Station 1 Station 2 12
Station 2 Inspection 15
Inspection Station 3 12
Station 1 Inspection 18
Move Triggers
Entities move from one location to the next based on available capacity of the
input buffer at the next location.
Work Schedule
Stations are scheduled to operate eight hours a day.
Assumption List
• No downtimes (downtimes occur too infrequently).
• Dedicated operators at each workstation are always available during the
scheduled work time.
• Rework times are half of the normal operation times.
har01307_ch05_115-160.indd 154 20/12/10 2:01 PM

Simulation Time and Replications

The simulation is run for 40 hours (10 hours of warm-up). There are five replications.
5.13 Summary
Data for building a model should be collected systematically with a view of how
the data are going to be used in the model. Data are of three types: structural, op-
erational, and numerical. Structural data consist of the physical objects that make
up the system. Operational data define how the elements behave. Numerical data
quantify attributes and behavioral parameters.
When gathering data, primary sources should be used first, such as historical
records or specifications. Developing a questionnaire is a good way to request in-
formation when conducting personal interviews. Data gathering should start with
structural data, then operational data, and finally numerical data. The first piece
of the puzzle to be put together is the routing sequence because everything else
hinges on the entity flow.
Numerical data for random variables should be analyzed to test for indepen-
dence and homogeneity. Also, a theoretical distribution should be fit to the data
if there is an acceptable fit. Some data are best represented using an empirical
distribution. Theoretical distributions should be used wherever possible.
Data should be documented, reviewed, and approved by concerned individu-
als. This data document becomes the basis for building the simulation model and
provides a baseline for later modification or for future studies.

1. Mr. Robert Vaughn is designing the Green Machine Manufacturing
Company. He has asked you to build a simulation model for a segment
of the facility that processes parts through each of four machines, which
operate in series. The interarrival time of parts to the first machine is
constant at 1.175 minutes. The processing time for the part at each
green machine is given in the table following.
Process Times at Green Machine Manufacturing Co.

Activity Process Time (minutes) ProModel Format
Machine 1 Exponential with mean 1.0 Wait E(1.0) min

Machine 2 Weibull with minimum 0.10, 5, 1 Wait (0.10 W(5,1)) min
A plentiful amount of storage can be provided for waiting parts in

front of each of the four green machines. Your task is to estimate the
har01307_ch05_115-160.indd 155 20/12/10 2:01 PM

time-average amount of work-in-process inventory (the expected

number of parts in the system) for the proposed configuration.
a. What are the input variables for the model?
b. What is the response variable for the model?
2. Give two examples of structural data, operational data, and numerical
data to be gathered when building a model.
3. Why is it best to begin gathering data by defining entity routings?
4. Of the distributions shown in the chapter, which theoretical distribution
most likely would be representative of time to failure for a machine?
5. Why would a normal distribution likely be a poor representation of an
activity time?
6. Assume a new system is being simulated and the only estimate
available for a manual operation is a most likely value. How would you
handle this situation?
7. Under what circumstances would you use an empirical distribution
instead of a standard theoretical distribution for an activity time?
8. Why is the distribution for interarrival times often nonstationary?
9. Assuming you had historical data on truck arrivals for the past year,
how would you arrive at an appropriate arrival distribution to model the
system for the next six months?
10. A new machine is being considered for which the company has no
reliability history. How would you obtain the best possible estimate of
reliability for the machine?
11. Suppose you are interested in looking at the impact of having workers
inspect their own work instead of having a dedicated inspection station.
If this is a new system requiring lots of assumptions to be made, how
can simulation be useful in making the comparison?
12. State whether the following are examples of a discrete probability
distribution or a continuous probability distribution.
a. Activity times.
b. Batch sizes.
c. Time between arrivals.
d. Probability of routing to one of six possible destinations.
13. Conceptually, how would you model a random variable X that
represents an activity time that is normally distributed with a mean of
10 minutes and a standard deviation of 3 minutes but is never less than
8 minutes?
14. Using Stat::Fit, generate a list of 50 random values between 10 and
100. Choose the Scatter Plot option and plot the data. Now put the data
in ascending order using the Input/Operate commands and plot the data.
Explain the correlation, if any, that you see in each scatter plot.
15. How can you check to see if a distribution in ProModel is giving the
right values?
har01307_ch05_115-160.indd 156 20/12/10 2:01 PM

16. Since many theoretical distributions are unbounded on the bottom,

what happens in ProModel if a negative value is sampled for an
activity time?
17. Go to a small convenience store or the university bookstore and
collect data on the interarrival and service times of customers. Make
histograms of the number of arrivals per time period and the number
of service completions per period. Note if these distributions vary by
the time of the day and by the day of the week. Record the number
of service channels available at all times. Make sure you secure
permission to perform the study.
18. The following are throughput time values for 30 simulation runs.
Calculate an estimate of the mean, variance, standard deviation, and
coefficient of variation for the throughput time. Construct a histogram
that has six cells of equal width.
10.7, 5.4, 7.8, 12.2, 6.4, 9.5, 6.2, 11.9, 13.1, 5.9, 9.6, 8.1, 6.3, 10.3, 11.5,
12.7, 15.4, 7.1, 10.2, 7.4, 6.5, 11.2, 12.9, 10.1, 9.9, 8.6, 7.9, 10.3, 8.3, 11.1
19. Customers calling into a service center are categorized according
to the nature of their problems. Five types of problem categories (A
through E) have been identified. One hundred observations were
made of customers calling in during a day, with a summary of the data
shown here. By inspection, you conclude that the data are most likely
uniformly distributed. Perform a chi-square goodness-of-fit test of this
hypothesis.
Type A B C D E
Observations 10 14 12 9 5
20. While doing your homework one afternoon, you notice that you are
frequently interrupted by friends. You decide to record the times
between interruptions to see if they might be exponentially distributed.
Here are 30 observed times (in minutes) that you have recorded;
conduct a goodness-of-fit test to see if the data are exponentially
distributed. (Hint: Use the data average as an estimate of the mean. For
the range, assume a range between 0 and infinity. Divide the cells based
on equal probabilities (pi) for each cell rather than equal cell intervals.)
2.08 6.86 4.86 2.55 5.94

2.96 0.91 2.13 2.20 1.40
16.17 2.11 2.38 0.83 2.81
14.57 0.29 2.73 0.73 1.76
2.79 11.69 18.29 5.25 7.42
2.15 0.96 6.28 0.94 13.76
har01307_ch05_115-160.indd 157 20/12/10 2:01 PM

5.15 Case Study
COLLECTING AND DOCUMENTING DATA FOR

HARRY’S DRIVE-THROUGH RESTAURANT
This exercise is intended to help you practice sorting through and organizing data on a
service facility. The facility is a fast-food restaurant called Harry’s Drive-Through and is
shown below.
Harry’s Drive-Through caters to two types of customers, walk-in and drive-through.
During peak times, walk-in customers arrive exponentially every 4 minutes and place their
order, which is sent to the kitchen. Nonpeak arrivals occur every 8–10 minutes. Custom-
ers wait in a pickup queue while their orders are filled in the kitchen. Two workers are
assigned to take and deliver orders. Once customers pick up their orders, 60 percent of the
customers stay and eat at a table, while the other 40 percent leave with their orders. There
is seating for 30 people, and the number of people in each customer party is one 40 percent
of the time, two 30 percent of the time, three 18 percent of the time, four 10 percent of
the time, and five 2 percent of the time. Eating time is normally distributed with a mean
Drive-through order
Kitchen
Drive-through
pickup
Pickup Order
Table area
har01307_ch05_115-160.indd 158 20/12/10 2:01 PM

of 15 minutes and a standard deviation of 2 minutes. If a walk-in customer enters and sees
that more than 15 customers are waiting to place their orders, the customer will balk (that
is, leave).
Harry’s is especially popular as a drive-through restaurant. Cars enter at a rate of 10 per
hour during peak times, place their orders, and then pull forward to pick up their orders. No
more than five cars can be in the pickup queue at a time. One person is dedicated to taking
orders. If over seven cars are at the order station, arriving cars will drive on.
The time to take orders is uniformly distributed between 0.5 minute and 1.2 minutes
including payment. Orders take an average of 8.2 minutes to fill with a standard deviation
of 1.2 minutes (normal distribution). These times are the same for both walk-in and drive-
through customers.
The objective of the simulation is to analyze performance during peak periods to see
how long customers spend waiting in line, how long lines become, how often customers
balk (pass by), and what the utilization of the table area is.
Problem
Summarize these data in table form and list any assumptions that need to be made in order
to conduct a meaningful simulation study based on the data given and objectives specified.
References
System Simulation. 5th ed. Englewood Cliffs, NJ: Prentice Hall, 2009.
Banks, Jerry, and Randall R. Gibson. “Selecting Simulation Software.” IIE Solutions, May
1997, pp. 29–32.
———. Stat::Fit. South Kent, CT: Geer Mountain Software Corporation, 1996.
Breiman, Leo. Statistics: With a View toward Applications. New York: Houghton Mifflin,
1973.
Brunk, H. D. An Introduction to Mathematical Statistics. 2nd ed. N.Y.: Blaisdell Publish-
ing Co. 1965.
Carson, John S. “Convincing Users of Model’s Validity Is Challenging Aspect of Mod-
eler’s Job.” Industrial Engineering, June 1986, p. 77.
Hoover, S. V., and R. F. Perry. Simulation: A Problem Solving Approach. Reading, MA,
Addison-Wesley, 1990.
Johnson, Norman L.; Samuel Kotz; and Adrienne W. Kemp. Univariate Discrete Distribu-
tions. New York: John Wiley & Sons, 1992, p. 425.
Knuth, Donald E. Seminumerical Algorithms. Reading, MA: Addison-Wesley, 1981.
Law, Averill M. Simulation Modeling & Analysis. 4th ed. New York: McGraw-Hill, 2007.
Stuart, Alan, and J. Keith Ord. Kendall’s Advanced Theory of Statistics. vol. 2. Cambridge:
Oxford University Press, 1991.
har01307_ch05_115-160.indd 159 20/12/10 2:01 PM

har01307_ch05_115-160.indd 160 20/12/10 2:01 PM
C H A P T E R
6 MODEL BUILDING
“Every theory [model] should be stated [built] as simply as possible, but not
simpler.”
—Albert Einstein
6.1 Introduction
In this chapter we look at how to translate a conceptual model of a system into
a simulation model. The focus is on elements common to both manufacturing
and service systems such as entity flow and resource allocation. Modeling issues
more specific to either manufacturing or service systems will be covered in later
chapters.
Modeling is more than knowing how to use a simulation software tool. Learn-
ing to use modern, easy-to-use software is one of the least difficult aspects of
modeling. Indeed, current simulation software makes poor and inaccurate models
easier to create than ever before. Unfortunately, software cannot make decisions
about how the elements of a particular system operate and how they should inter-
act with each other. This is the role of the modeler.
Modeling is considered an art or craft as much as a science. Knowing the
theory behind simulation and understanding the statistical issues are the sci-
ence part. But knowing how to effectively and efficiently represent a system
using a simulation tool is the artistic part of simulation. It takes a special knack
to be able to look at a system in the abstract and then creatively construct a rep-
resentative logical model using a simulation tool. If three different people were
to model the same system, chances are three different modeling approaches
would be taken. Modelers tend to use the techniques with which they are most
familiar. So the best way to develop good modeling skills is to look at lots of
good examples and, most of all, practice, practice, practice! Skilled simula-
tion analysts are able to quickly translate a process into a simulation model
161
har01307_ch06_161-192.indd 161 20/12/10 1:50 PM

and begin conducting experiments. Consider the following statement by Robert

Shannon (1998):
If you want to become proficient in an art you must take up the tools (palette, canvas,
paint, and brushes) and begin to paint. As you do so, you will begin to see what works
and what doesn’t. The same thing is true in simulation. You learn the art of simulation
by simulating. Having a mentor to help show you the way can shorten the time and
effort.
The purpose of this chapter is to provide a little mentoring by showing how

typical modeling issues are handled during simulation. Some of the questions
answered in this chapter include
• How do I convert a conceptual model to a simulation model?
• What is the relationship between model simplicity and model usefulness?
• How do I determine which system elements to include in a model?
• How do I determine the best way to represent system elements in a
model?
• How are common situations modeled using simulation?
Because a simulation model relies on a simulation language for its expres-
sion, (and since this textbook is titled Simulation Using ProModel), ProModel
will be used to illustrate how system elements are modeled.
For the most part, the discussion of modeling issues in this chapter is kept at a
conceptual level. Specific modeling examples in ProModel can be found in Lab 6
of Part II.
6.2 Converting a Conceptual Model to a Simulation Model

The conceptual model is the result of the data-gathering effort and is a formula-
tion in one’s mind (supplemented with notes and diagrams) of how a particular
system operates. Building a simulation model requires that this conceptual model
be converted to a simulation model. Making this translation requires two impor-
tant transitions in one’s thinking. First, the modeler must be able to think of the
system in terms of the modeling paradigm supported by the particular model-
ing software that is being used. Second, the different possible ways to model the
system should be evaluated to determine the most efficient yet effective way to
represent the system.
6.2.1 Modeling Paradigms

A simulation model is a computer representation of how elements in a particular
system behave and interact. The internal representation of a system may not be
radically different between simulation products. However, the way that model
information is defined using a particular product may be quite different. How a
har01307_ch06_161-192.indd 162 20/12/10 1:50 PM

Chapter 6 Model Building 163
user defines a model using a particular simulation product is based on the model-
ing paradigm of the product. A modeling paradigm consists of the constructs and
associated language that dictate how the modeler should think about the system
being modeled. In this regard, a modeling paradigm determines the particular
“world view” that one should have when modeling a system. Learning a simula-
tion product requires that one first learn the particular modeling paradigm used
by the product.
Modeling paradigms differ among simulation products—although many
differences are minor. Historically, simulation products required models to be
defined line by line, specifying a list of instructions like a script for processing en-
tities. Most modern products have a process orientation and support object-based
modeling. Some products view a process as an activity sequence, while others
view it from the standpoint of an entity routing.
Conceptually, object-based modeling is quite simple. An object represents a
real-world object such as a machine, an operator, a process part, or a customer.
An object is defined in terms of attributes and behaviors. Attributes are essentially
variables that are associated with an object such as its size, condition, time in
the system, and so on. Attributes are used to carry information about the object,
either for decision making or for output reporting. Attributes may be modified
during the simulation to reflect changing values. Behaviors define the operational
logic associated with the object. This logic gets executed either whenever certain
events occur or when explicitly called in the model. Examples of behavioral logic
include operations, machine setup, routing decisions, and material movement. Ex-
amples of events or conditions that may trigger a behavior might be the comple-
tion of an operation, the elapse of a specified period of time, or the drop of a tank
or inventory to a specified level.
Objects are organized into classes according to similarity of attributes and
behavior. In this way, all objects of the same class inherit the attributes and behav-
iors defined for the class. This simplifies object definitions and allows changes to
be quickly made to a set of similar objects. It also allows objects that are defined
once to be reused in other models. In ProModel, predefined object classes include
entities, resources, locations, and paths.
While object-based modeling is consistent with modern object-oriented pro-
gramming (OOP) practices, OOP itself is quite complicated and requires a signifi-
cant amount of training to become proficient. The easiest simulation languages
are object-based, but they are not pure OOP languages. Otherwise, a modeler
might as well pick up an OOP language like Java and begin programming his or
her models.
ProModel is object-based but also provides an intuitive entity-flow modeling
paradigm. The natural way to think about most systems being modeled is to think
of them from the perspective of the entity as it flows through each workstation,
storage area, or some other location. In fact, it actually helps when building the
model to put yourself in the place of the entity flowing through the system and
describe what you do and what resources you require as you move from place
to place in the system. For modeling purposes, it is useful to think of a system
har01307_ch06_161-192.indd 163 20/12/10 1:50 PM

in terms of the same structural and operational elements that were described in
Chapter 5. This is essentially how models are defined using ProModel.
6.2.2 Model Definition

A model is a simplified representation of reality, with emphasis on the word
simplified. This means that the exact way in which an operation is performed is
not so important as the way in which the operation impacts the rest of the sys-
tem. An activity should always be viewed in terms of its effect on other system
elements rather than in terms of the detailed way in which it is performed. Such
detailed mechanics are inconsequential to the overall flow of entities and utiliza-
tion of resources.
Most models, especially those built by beginners, tend to err on the side
of being overly detailed rather than being too general. The tendency is to
reproduce the precise way that the system operates (sometimes referred to as
emulation). Not only is it difficult to create extremely detailed models, but it
is also difficult to debug and maintain them. Furthermore, all of the detail may
obscure the key issues being analyzed so that it actually weakens the model.
The power of a model is more a function of its simplicity rather than its com-
plexity. A point can be made more effectively when it is reduced to its simplest
form rather than disguised in a morass of detail. Lou Keller of PROMODEL
Corporation has suggested that the Laffer curve, borrowed from economics,
effectively illustrates the relationship between model complexity and model
utility (see Figure 6.1). Notice that a certain degree of complexity is essential
to capture the key cause-and-effect relationships in the system. However, a sys-
tem has many more cause-and-effect relationships than should be included in a
model. There is an optimum amount of complexity for any given model beyond
which additional utility begins to diminish.
FIGURE 6.1
Relationship between
model complexity and
model utility (also
known as the Laffer
Complexity
curve).
Optimum level of
model complexity
Utility
har01307_ch06_161-192.indd 164 20/12/10 1:50 PM

The remaining sections in this chapter give recommendations on how to ef-

fectively yet minimally represent system elements in a simulation model using a
simulation language like ProModel.
6.3 Structural Elements

For the most part, model objects represent the structural elements in a system such
as machines, people, work items, and work areas. For purposes of our discussion,
we will be using the simple object classification employed by ProModel:
• Entities—the items processed in the system.
• Locations—places where entities are processed or held.
• Resources—agents used in the processing of entities.
• Paths—the course of travel for entities and resources in the system.
Not all simulation products provide the same set or classification of modeling
elements. Even these elements could be further subdivided to provide greater dif-
ferentiation. Locations, for example, could be subdivided into workstations, buf-
fers, queues, and storage areas. From a system dynamics perspective, however,
these are all still simply places to which entities are routed and where operations
or activities may be performed. For this reason, it is easier just to think of them
all in the generic sense as locations. The object classification used by ProModel is
simple yet broad enough to encompass virtually any object encountered in most
manufacturing and service systems.
These model elements or objects have behavior associated with them (dis-
cussed in section 6.4, “Operational Elements”) and attributes. Most common be-
haviors and attributes are selectable from menus, which reduces the time required
to build models. The user may also predefine behaviors and attributes that are
imported into a model.
6.3.1 Entities
Entities are the objects processed in the model that represent the inputs and out-
puts of the system. Entities in a system may have special characteristics such as
speed, size, condition, and so on. Entities follow one or more different routings in
a system and have processes performed on them. They may arrive from outside
the system or be created within the system. Usually, entities exit the system after
visiting a defined sequence of locations.
Simulation models often make extensive use of entity attributes. For exam-
ple, an entity may have an attribute called Condition that may have a value of 1 for
defective or 0 for nondefective. The value of this attribute may determine where
the entity gets routed in the system. Attributes are also frequently used to gather
information during the course of the simulation. For example, a modeler may de-
fine an attribute called ValueAddedTime to track the amount of value-added time
an entity spends in the system.
har01307_ch06_161-192.indd 165 20/12/10 1:50 PM

The statistics of interest that are generally collected for entities include time
in the system (flow time), quantity processed (output), value-added time, time
spent waiting to be serviced, and the average number of entities in the system.
Entities to Include
When deciding what entities to include in a model, it is best to look at every
kind of entity that has a bearing on the problem being addressed. For example,
if a component part is assembled to a base item at an assembly station, and the
station is always stocked with the component part, it is probably unnecessary to
model the component part. In this case, what is essential to simulate is just the
time delay to perform the assembly. If, however, the component part may not al-
ways be available due to delays, then it might be necessary to simulate the flow of
component parts as well as the base items. The rule is that if you can adequately
capture the dynamics of the system without including the entity, don’t include it.
Entity Aggregating
It is not uncommon for some manufacturing systems to have hundreds of part
types or for a service system to have hundreds of different customer types. Mod-
eling each one of these entity types individually would be a painstaking task that
would yield little, if any, benefit. A better approach is to treat entity types in the
aggregate whenever possible (see Figure 6.2). This works especially well when all
entities have the same processing sequence. Even if a slight difference in process-
ing exists, it often can be handled through use of attributes or by using probabili-
ties. If statistics by entity type are not required and differences in treatment can be
defined using attributes or probabilities, it makes sense to aggregate entity types
into a single generic entity and perhaps call it part or customer.
Entity Resolution
Each individual item or person in the system need not always be represented by
a corresponding model entity. Sometimes a group of items or people can be rep-
resented by a single entity (see Figure 6.3). For example, a single entity might be
used to represent a batch of parts processed as a single unit or a party of people
eating together in a restaurant. If a group of entities is processed as a group and
Type A
Type B
Type X
Type C 9 entities 1 entity
FIGURE 6.2 FIGURE 6.3
Treating different entity types Treating multiple entities
as a single type. as a single entity.
har01307_ch06_161-192.indd 166 20/12/10 1:50 PM

moved as a group, there is no need to model them individually. Activity times or

statistics that are a function of the size of the group can be handled using an at-
tribute that keeps track of the items represented by the single entity.
High-Rate Entity Processing

In some systems, especially in the food and beverage industries, it is not un-
common to process entities at a rate of hundreds per minute. In such situations,
modeling each individual entity can significantly slow down the simulation. For
high-rate processing in which individual entity tracking isn’t critical, it may be
preferable to merely track entity production at various stages of a process using
variables or attributes instead of individual entities. Simulation languages having
a tank or accumulator construct can make this easier. Having thousands of entities
in the system at one time (especially those with multiple attributes) can consume
lots of computer memory and may slow the simulation significantly.
An alternative approach to modeling high-rate entity processing is to adjust
the resolution of the entity, as discussed earlier. Using this approach, you might
model every carton of 48 bottles or containers in a beverage-processing facility
as a single entity. This approach also works for continuously flowing substances
such as liquids or granules, which can often be converted, for simulation, into
discrete units of measure such as gallons, pounds, or barrels.
6.3.2 Locations
Locations are places in the system that entities visit for processing, waiting, or de-
cision making. A location might be a treatment room, workstation, check-in point,
queue, or storage area. Locations have a holding capacity and may have certain
times that they are available. They may also have special input and output such as
input based on highest priority or output based on first-in, first out (FIFO).
In simulation, we are often interested in the average contents of a location
such as the average number of customers in a queue or the average number of
parts in a storage rack. We might also be interested in how much time entities
spend at a particular location for processing. There are also location state statistics
that are of interest such as utilization, downtime, or idle time.
Locations to Include
Deciding what to model as a route location depends largely on what happens at
the location. If an entity merely passes through a location en route to another
without spending any time, it probably isn’t necessary to include the location. For
example, a water spray station through which parts pass without pausing probably
doesn’t need to be included in a model. In considering what to define as a location,
any point in the flow of an entity where one or more of the following actions take
place may be a candidate:
• Place where an entity is detained for a specified period of time while
undergoing an activity (such as fabrication, inspection, or cleaning).
• Place where an entity waits until some condition is satisfied (like the
availability of a resource or the accumulation of multiple entities).
har01307_ch06_161-192.indd 167 20/12/10 1:50 PM

• Place or point where some action takes place or logic gets executed, even
though no time is required (splitting or destroying an entity, sending a
signal, incrementing an attribute or variable).
• Place or point where a decision is made about further routing (a branch
in a conveyor or path, an area in a store where customers decide to which
checkout counter they will go).
Location Resolution
Depending on the level of resolution needed for the model, a location may be an
entire factory or service facility at one extreme, or individual positions on a desk or
workbench at the other. The combination of locations into a single location is done
differently depending on whether the locations are parallel or serial locations.
When combining parallel locations having identical processing times, the re-
sulting location should have a capacity equal to the combined capacities of the
individual locations; however, the activity time should equal that of only one of
the locations (see Figure 6.4). A situation where multiple locations might be com-
bined in this way is a parallel workstation. An example of a parallel workstation is
a work area where three machines perform the same operation. All three machines
could be modeled as a single location with a capacity equal to three. Combining
locations can significantly reduce model size, especially when the number of par-
allel units gets very large, such as a 20-station checkout area in a large shopping
center or a 100-seat dining area in a restaurant.
FIGURE 6.4
Example of combining
three parallel stations Capacity ⫽ 1
into a single station.
Capacity ⫽ 1 Capacity ⫽ 1 Capacity ⫽ 1
Operation 10 Operation 30
Capacity ⫽ 1
Operation 20
1 min. each
Operation 10 Operation 20 Operation 30

1 min.
har01307_ch06_161-192.indd 168 20/12/10 1:50 PM

FIGURE 6.5
Example of combining three serial stations into a single station.
Capacity ⫽ 1 Capacity ⫽ 1 Capacity ⫽ 1 Capacity ⫽ 1 Capacity ⫽ 1
Station 1 Station 2 Station 3 Station 4 Station 5

1 min. 1 min. 1 min.
Station 1 Stations 2–4 Station 5

1 min. 3 min. 1 min.
When combining serial locations, the resultant location should have a capac-
ity equal to the sum of the individual capacities and an activity time equal to the
sum of the activity times. An example of a combined serial sequence of locations
might be a synchronous transfer line that has multiple serial stations. All of them
could be represented as a single location with a capacity equal to the number of
stations (see Figure 6.5). Parts enter the location, spend an amount of time equal
to the sum of all the station times, and then exit the location. The behavior may
not be exactly the same as having individual stations, such as when the location
becomes blocked and up to three parts may be finished and waiting to move to
station 5. The modeler must decide if the representation is a good enough approxi-
mation for the intended purpose of the simulation.
6.3.3 Resources
Resources are the agents used to process entities in the system. Resources may be
either static or dynamic depending on whether they are stationary (like a copy ma-
chine) or move about in the system (like an operator). Dynamic resources behave
much like entities in that they both move about in the system. Like entities, re-
sources may be either animate (living beings) or inanimate (a tool or machine). The
primary difference between entities and resources is that entities enter the system,
have a defined processing sequence, and, in most cases, finally leave the system.
Resources, however, usually don’t have a defined flow sequence and remain in the
system (except for off-duty times). Resources often respond to requests for their
use, whereas entities are usually the objects requiring the use of resources.
har01307_ch06_161-192.indd 169 20/12/10 1:50 PM

In simulation, we are interested in how resources are utilized, how many re-
sources are needed, and how entity processing is affected by resource availability.
The response time for acquiring a resource may also be of interest.
Resources to Include
The decision as to whether a resource should be included in a model depends
largely on what impact it has on the behavior of the system. If the resource is
dedicated to a particular workstation, for example, there may be little benefit in
including it in the model since entities never have to wait for the resource to
become available before using it. You simply assign the processing time to the
workstation. If, on the other hand, the resource may not always be available (it
experiences downtime) or is a shared resource (multiple activities compete for the
same resource), it should probably be included. Once again, the consideration is
how much the resource is likely to affect system behavior.
Resource Travel Time

One consideration when modeling the use of resources is the travel time associ-
ated with mobile resources. A modeler must ask whether a resource is imme-
diately accessible when available, or if there is some travel time involved. For
example, a special piece of test equipment may be transported around to several
locations in a facility as needed. If the test equipment is available when needed at
some location, but it takes 10 minutes to get the test equipment to the requesting
location, that time should be accounted for in the model. The time for the resource
to move to a location may also be a function of the distance it must travel.
Consumable Resources
Depending on the purpose of the simulation and degree of influence on system
behavior, it may be desirable to model consumable resources. Consumable re-
sources are used up during the simulation and may include
• Services such as electricity or compressed air.
• Supplies such as staples or tooling.
Consumable resources are usually modeled either as a function of time or as
a step function associated with some event such as the completion of an operation.
This can be done by defining a variable or attribute that changes value with time or
by event. A variable representing the consumption of packaging materials, for ex-
ample, might be based on the number of entities processed at a packaging station.
Transport Resources
Transport resources are resources used to move entities within the system. Exam-
ples of transport resources are lift trucks, elevators, cranes, buses, and airplanes.
These resources are dynamic and often are capable of carrying multiple entities.
Sometimes there are multiple pickup and drop-off points to deal with. The trans-
porter may even have a prescribed route it follows, similar to an entity routing.
A common example of this is a bus route.
har01307_ch06_161-192.indd 170 20/12/10 1:50 PM

In advanced manufacturing systems, the most complex element to model

is often the transport or material handling system. This is because of the com-
plex operation that is associated with these computer-controlled systems such
as conveyor systems and automated guided vehicle systems (AGVS). Modeling
advanced material handling systems can be simplified if the modeling language
provides special constructs for defining them; otherwise the logic describing their
behavior must be defined by the modeler. ProModel provides facilities for model-
ing material handling systems that are described in Chapter 12.
6.3.4 Paths
Paths define the course of travel for entities and resources. Paths may be isolated,
or they may be connected to other paths to create a path network. In ProModel
simple paths are automatically created when a routing path is defined. A routing
path connecting two locations becomes the default path of travel if no explicitly
defined path or path network connects the locations.
Paths linked together to form path networks are common in manufacturing and
service systems. In manufacturing, aisles are connected to create travel ways for lift
trucks and other material handlers. An AGVS sometimes has complex path networks
that allow controlled traffic flow of the vehicles in the system. In service systems,
office complexes have hallways connecting other hallways that connect to offices.
Transportation systems use roadways, tracks, and so on that are often interconnected.
When using path networks, there can sometimes be hundreds of routes to take
to get from one location to another. ProModel is able to automatically navigate
entities and resources along the shortest path sequence between two locations.
Optionally, you can explicitly define the path sequence to take to get from one
point to any other point in the network.
6.4 Operational Elements

Operational elements define the behavior of the different physical elements in the
system and how they interact. These include routings, operations, arrivals, entity
and resource movement, task selection rules, resource schedules, and downtimes
and repairs.
Most of the operational elements of a model can be defined using constructs
that are specifically provided for modeling such elements. The operational rules
for each can usually be selected from menus. There may be situations, however,
that require the modeler to use special logic such as “if–then” statements to
achieve the special operating behavior that is desired.
6.4.1 Routings
Routings define the sequence of flow for entities from location to location. When
entities complete their activity at a location, the routing defines where the entity goes
next and specifies the criterion for selecting from among multiple possible locations.
har01307_ch06_161-192.indd 171 20/12/10 1:50 PM

Frequently entities may be routed to more than one possible location. When
choosing from among multiple alternative locations, a rule or criterion must be
defined for making the selection. A few typical rules that might be used for select-
ing the next location in a routing decision include
• Probabilistic—entities are routed to one of several locations according to
a frequency distribution.
• First available—entities go to the first available location in the order they
are listed.
• By turn—the selection rotates through the locations in the list.
• Most available capacity—entities select the location that has the most
available capacity.
• Until full—entities continue to go to a single location until it is full and
then switch to another location, where they continue to go until it is full,
and so on.
• Random—entities choose randomly from among a list of locations.
• User condition—entities choose from among a list of locations based on a
condition defined by the user.
Recirculation
Sometimes entities revisit or pass through the same location multiple times. The
best approach to modeling this situation is to use an entity attribute to keep track
of the number of passes through the location and determine the operation or rout-
ing accordingly. When using an entity attribute, the attribute is incremented either
on entry to or on exit from a location and tested before making the particular op-
eration or routing decision to see which pass the entity is currently on. Based on
the value of the attribute, a different operation or routing may be executed.
Unordered Routings
Certain systems may not require a specific sequence for visiting a set of locations
but allow activities to be performed in any order as long as they all eventually get
performed. An example is a document requiring signatures from several depart-
ments. The sequence in which the signatures are obtained may be unimportant as
long as all signatures are obtained.
In unordered routing situations, it is important to keep track of which loca-
tions have or haven’t been visited. Entity attributes are usually the most practical
way of tracking this information. An attribute may be defined for each possible
location and then set to 1 whenever that location is visited. The routing is then
based on which of the defined attributes are still set to zero.
6.4.2 Entity Operations

An entity operation defines what happens to an entity when it enters a location.
For modeling purposes, the exact nature of the operation (machining, patient
check-in, or whatever) is unimportant. What is essential is to know what happens
har01307_ch06_161-192.indd 172 20/12/10 1:50 PM

in terms of the time required, the resources used, and any other logic that impacts
system performance. For operations requiring more than a time and resource des-
ignation, detailed logic may need to be defined using if–then statements, variable
assignment statements, or some other type of statement (see section 6.4.8, “Use
of Programming Logic”).
An entity operation is one of several different types of activities that take
place in a system. As with any other activity in the system, the decision to include
an entity operation in a model should be based on whether the operation impacts
entity flow in some way. For example, if a labeling activity is performed on enti-
ties in motion on a conveyor, the activity need not be modeled unless there are
situations where the labeler experiences frequent interruptions.
Consolidation of Entities
Entities often undergo operations where they are consolidated or become either
physically or logically connected with other entities. Examples of entity consoli-
dation include batching and stacking. In such situations, entities are allowed to
simply accumulate until a specified quantity has been gathered, and then they
are grouped together into a single unit. Entity consolidation may be temporary,
allowing them to later be separated, or permanent, in which case the consolidated
entities no longer retain their individual identities. Figure 6.6 illustrates these two
types of consolidation.
Examples of consolidating multiple entities to a single entity include
• Accumulating multiple items to fill a container.
• Gathering people together into groups of five for a ride at an amusement
park.
• Grouping items to load them into an oven for heating.
In ProModel, entities are consolidated permanently using the COMBINE com-
mand. Entities may be consolidated temporarily using the GROUP command.
Attachment of Entities
In addition to consolidating accumulated entities at a location, entities can also
be attached to a specific entity at a location. Examples of attaching entities might
be attaching a shipping document to an order ready for delivery, assembling
FIGURE 6.6
Consolidation
of entities into a before after
single entity. In (a)
(a) permanent
consolidation, batched
entities get destroyed.
In (b) temporary
consolidation, batched
entities are preserved before
(b) after
for later unbatching.
har01307_ch06_161-192.indd 173 20/12/10 1:50 PM

FIGURE 6.7
Attachment of one
or more entities to
another entity.
In (a) permanent after
attachment, the
attached entities before
get destroyed. (a)
In (b) temporary
attachment, the
attached entities are
preserved for later
detachment.
after
before
(b)
wheels to a chassis, or placing items into a container. The difference between

attaching entities and consolidating entities is that entities become attached to an
existing base or main entity that must be present at the location. It is the presence
of the main entity at the location that triggers the routing of the entities to be at-
tached. In consolidation, entities are combined in whichever order they happen
to enter the location. Like consolidation, attachment may be temporary, where
the attached entities are later detached, or permanent, where the attached entities
are destroyed and only the initial entity to which they were attached continues.
Figure 6.7 illustrates these two types of attachment.
Examples of attaching entities to another entity include
• Attaching component parts to a base assembly.
• Delivering a completed order to a waiting customer.
• Loading material into a container.
In ProModel, entities are attached to another entity using the LOAD command
for temporary attachments or the JOIN command for permanent attachments.
Corresponding LOAD and JOIN routings must also be defined for the entities to be
loaded or joined.
Dividing Entities
In some entity processes, a single entity is converted into two or more new enti-
ties. An example of entity splitting might be an item that is cut into smaller pieces
or a purchase order that has carbon copies removed for filing or sending to ac-
counting. Entities are divided in one of two ways: either the entity is split up into
two or more new entities and the original entity no longer exists; or additional
entities are merely created (cloned) from the original entity, which continues to
exist. These two methods are shown in Figure 6.8.
har01307_ch06_161-192.indd 174 20/12/10 1:50 PM

FIGURE 6.8
Multiple entities
created from a single before after
entity. Either (a) the (a)
entity splits into
multiple entities (the
original entity is
destroyed) or (b) the
entity creates one
or more entities before
(the original entity
after
continues).
(b)
Examples of entities being split or creating new entities from a single entity
include
• A container or pallet load being broken down into the individual items
comprising the load.
• Driving in and leaving a car at an automotive service center.
• Separating a form from a multiform document.
• A customer placing an order that is processed while the customer waits.
• A length of bar stock being cut into smaller pieces.
In ProModel, entities are split using a SPLIT statement. New entities are cre-
ated from an existing entity using a CREATE statement. Alternatively, entities can be
conveniently split or created using the routing options provided in ProModel.
6.4.3 Entity Arrivals

Entity arrivals define the time, quantity, frequency, and location of entities enter-
ing the system. Examples of arrivals are customers arriving at a post office or
cars arriving at an intersection. Entities may arrive to a manufacturing or service
system in one of several different ways:
• Periodic—they arrive at a periodic interval.
• Scheduled—they arrive at specified times.
• Fluctuating—the rate of arrival fluctuates with time.
• Event triggered—they arrive when triggered by some event.
In any case, entities can arrive individually or in batches.
Periodic Arrivals
Periodic arrivals occur more or less at the same interval each time. They may
occur in varying quantities, and the interval is often defined as a random variable.
Periodic arrivals are often used to model the output of an upstream process that
feeds into the system being simulated. For example, computer monitors might
har01307_ch06_161-192.indd 175 20/12/10 1:50 PM

arrive from an assembly line to be packaged at an interval that is normally distrib-

uted with a mean of 1.6 minutes and a standard deviation of 0.2 minute. Examples
of periodic arrivals include
• Parts arriving from an upstream operation that is not included in the model.
• Customers arriving to use a copy machine.
• Phone calls for customer service during a particular part of the day.
Periodic arrivals are defined in ProModel by using the arrivals table.
Scheduled Arrivals
Scheduled arrivals occur when entities arrive at specified times with possibly
some defined variation (that is, a percentage will arrive early or late). Scheduled
arrivals may occur in quantities greater than one such as a shuttle bus transporting
guests at a scheduled time. It is often desirable to be able to read in a schedule
from an external file, especially when the number of scheduled arrivals is large
and the schedule may change from run to run. Examples of scheduled arrivals
include
• Customer appointments to receive a professional service such as
counseling.
• Patients scheduled for lab work.
• Production release times created through an MRP (material requirements
planning) system.
Scheduled arrivals sometime occur at intervals, such as appointments that occur
at 15-minute intervals with some variation. This may sound like a periodic arrival;
however, periodic arrivals are autocorrelated in that the absolute time of each
arrival is dependent on the time of the previous arrival. In scheduled arrival inter-
vals, each arrival occurs independently of the previous arrival. If one appointment
arrives early or late, it will not affect when the next appointment arrives.
ProModel provides a straightforward way for defining scheduled arrivals
using the arrivals table. A variation may be assigned to a scheduled arrival to
simulate early or late arrivals for appointments.
Fluctuating Arrivals
Sometimes entities arrive at a rate that fluctuates with time. For example, the rate
at which customers arrive at a bank usually varies throughout the day with peak
and lull times. This pattern may be repeated each day (see Figure 6.9). Examples
of fluctuating arrivals include
• Customers arriving at a restaurant.
• Arriving flights at an international airport.
• Arriving phone calls for customer service.
In ProModel, fluctuating arrivals are specified by defining an arrival cycle
pattern for a time period that may be repeated as often as desired.
har01307_ch06_161-192.indd 176 20/12/10 1:50 PM

FIGURE 6.9
A daily cycle pattern 120
of arrivals.
100
Arrival quantity
80
60
40
20
9A.M. 10A.M. 11A.M.12 NOON 1P.M. 2P.M. 3P.M. 4P.M. 5P.M. 6P.M.
Hour of day
Event-Triggered Arrivals
In many situations, entities are introduced to the system by some internal trigger
such as the completion of an operation or the lowering of an inventory level to a
reorder point. Triggered arrivals might occur when
• A kanban or delivery signal is received.
• Inventory drops to the reorder-point level.
• Conditions have been met to start processing a new entity.
Sometimes the signal to initiate an arrival or to begin production of an item is
triggered by some other arrival. An example of this is the arrival of a customer order
for a product. In such instances, it isn’t the order that flows through the system but
the product the order triggers. To model this situation, you define the characteristic
of the order arrival (frequency, type, or the like) and then have the arriving order
trigger production of the product.
In ProModel, entity arrivals can be triggered using an ORDER statement. This
statement can be used anywhere in the model so that any situation can cause a new
entity to be ordered into the model.
6.4.4 Entity and Resource Movement

Entities and resources seldom remain stationary in a system. Entities move
through the system from location to location for processing. Resources also move
to different locations where they are requested for use. Additionally, resources
frequently move or escort entities in the system.
Movement can be handled in three basic ways in a simulation:
1. Ignore the movement.
2. Model the move using a simple move time (which may also be defined
by speed and distance).
har01307_ch06_161-192.indd 177 20/12/10 1:50 PM

3. Model the move using a path network that requires the moving entity or
resource to contend with traffic.
A path network essentially imitates the network of aisles or hallways found
on plant floors and in office facilities. A path network reduces the number of paths
that need to be defined if there are a lot of different routings yet all movement
shares common path segments.
The decision as to which method to use depends once again on the level of
detail needed for a valid model. In making this determination, the following rules
may be useful:
• If the move time is negligible compared to activity times, or if it makes
sense to simply include it as part of the operation time, it may be ignored.
• If move times are significant but traffic congestion is light, a simple move
time (or speed and distance) should be defined.
• If move times are significant and traffic congestion is heavy, a path
network should be defined.
6.4.5 Accessing Locations and Resources

Much of the activity in a simulation is governed by how entities are able to ac-
cess locations and resources for processing. Entities may be given priorities when
contending with other entities for a particular location or resource. The location
or resource might also be given decision-making capability for selecting from
among multiple items awaiting input.
Use of Priorities
Locations and resources may be requested with a particular priority in ProModel.
Priorities range from 0 to 999, with higher values having higher priority. If no
priority is specified, it is assumed to be 0. For simple prioritizing, you should
use priorities from 0 to 99. Priorities greater than 99 are for preempting entities
and downtimes currently in control of a location or resource. The command Get
Operator, 10 will attempt to get the resource called Operator with a priority of
10. If the Operator is available when requested, the priority has no significance. If,
however, the Operator is currently busy, this request having a priority of 10 will
get the Operator before any other waiting request with a lower priority.
Preemption
Sometimes it is desirable to have a resource or location respond immediately to
a task, interrupting the current activity it is doing. This ability to bump another
activity or entity that is using a location or resource is referred to as preemption.
For example, an emergency patient requesting a particular doctor may preempt a
patient receiving routine treatment by the doctor. Manufacturing and service orga-
nizations frequently have “hot” jobs that take priority over any routine job being
done. Preemption is achieved by specifying a preemptive priority (100 to 999) for
entering the location or acquiring the resource.
har01307_ch06_161-192.indd 178 20/12/10 1:50 PM

Priority values in ProModel are divided into 10 levels (0 to 99, 100 to 199, . . . ,
900 to 999). Levels higher than 99 are used to preempt entities or downtimes of
a lower level. Multiple preemptive levels make it possible to preempt entities or
downtimes that are themselves preemptive.
The level of preemption to specify is determined by whether an entity or a
downtime is being preempted. To preempt an entity that is currently using a
location or resource, the preempting entity or downtime must have a priority that
is at least one level higher than the current entity. For example, a priority of 100
would preempt any entity that had acquired a location or resource with a priority
of 0 to 99. To preempt a downtime currently in effect at a location or resource, the
preempting entity or downtime must have a priority that is at least two levels higher
than the downtime (the exception to this is a setup downtime, which is preempted
using the same rules as those for preempting an entity).
The primary issue associated with preemption is what to do with the pre-
empted entity. The common way of handling a preempted entity is to simply put
it on hold until the resource or location becomes available again for use. Then
any remaining time is completed. This is the default way that ProModel handles
preemption. Alternatively, it may be desirable to get a different resource instead
of waiting for the one being used by the preempting entity. Sometimes it may be
necessary to delay the preemption or possibly even ignore it, depending on certain
circumstances in the model. For example, it may be unallowable to interrupt a
machining cycle or a surgical operation. In such instances, a preemptive attempt
may just have to wait. ProModel allows you to write special preemption logic for
achieving these kinds of effects.
Task Selection Rules

Locations and resources are often discriminating with respect to which waiting
entity or activity they next select to service. While priorities determine the order
in which entities line up to use a location or resource, task selection rules deter-
mine which entity waiting for a location or resource is actually granted permis-
sion to use the location or resource when it becomes available. For example, parts
waiting to be processed at a machine may also be competing with parts waiting to
be reworked on the same machine. In such a situation, it may be desirable to first
select the part requiring rework.
Task selection plays a crucial role in simulation-based scheduling (Thompson
1994). Rules range from simple priority rules such as shortest processing time,
critical ratio, or earliest due date to combinatorial rules and user-defined decision
logic. User-defined decision logic provides the most flexible means for defining
task selection. It allows the user to define the logic for selecting the next task by
examining the nature of all of the tasks waiting to be performed, looking upstream
or looking ahead to see what entities or tasks will be coming along, and looking
downstream to see what tasks will best serve the needs of subsequent waiting
resources. By default locations and resources in ProModel respond to the highest-
priority request that has been waiting the longest. Other criteria may be defined,
however, to force a different response.
har01307_ch06_161-192.indd 179 20/12/10 1:50 PM

6.4.6 Resource Scheduling

Resources, as well as locations, frequently have scheduled times during which
they are unavailable. These include off-shift periods, breaks, and preventive
maintenance. If it is necessary to model periods of availability and unavailability
to have a valid model, some of the issues that need to be addressed when model-
ing scheduled availability include
• Deciding what to do with a task that is only half completed when the end
of a shift occurs.
• Making sure that resource statistics are based only on the scheduled
available time and not on the entire simulation time.
• Deciding what to do with arrivals that occur at a location that is off shift.
Going off Schedule in the Middle of a Task

It is not uncommon when running a simulation with work schedules to have a
resource go off schedule in the middle of a task (the current task is preempted).
For short tasks, this is generally not a problem; you simply allow the resource to
complete the task and then allow the schedule to take control. For longer tasks
that may take hours, it usually isn’t desirable to keep the resource until the task is
completed. There are at least three options to take:
1. Don’t start the task in the first place.
2. Interrupt the task and go off schedule.
3. If the task is nearly complete, go ahead and finish the task.
To determine whether to start the task in the first place, the modeler needs
to test for the time before the next schedule change and compare it with the an-
ticipated time to complete the task at hand. Interrupting a task usually requires
only that a preemptive priority be given to the schedule. To determine whether
the task is almost complete, a variable should be assigned to the completion
time, which can be checked at the time of the schedule change. A related situ-
ation is when several tasks are waiting that must be processed before going
off schedule (for example, a checkout stand in a grocery store closes but must
service current customers in line). This situation is facilitated if some logic can
be executed at the time a schedule is about to be executed. This logic needs to
be able to delay the schedule execution. ProModel allows these types of delays
to be defined.
Basing Resource Statistics on Scheduled Time

When including work schedules in a model, it is generally desirable to gather
statistics on the resource, such as utilization statistics, based on the scheduled
time available—not on total simulation time. ProModel automatically excludes
off-scheduled time when calculating statistics. This way, statistics are reported for
resources and locations only for the time during which they were scheduled to be
available.
har01307_ch06_161-192.indd 180 20/12/10 1:50 PM

Handling Arrivals during Off-Shift Times

It is usually desirable to prevent arrivals (especially human entities) from occur-
ring during off-schedule times because you don’t want them to have to wait until
a location or resource is back on schedule. To turn off arrivals during off-shift
times, one solution is to try to synchronize the arrivals with the work schedule.
This usually complicates the way arrivals are defined. Another solution, and usu-
ally an easier one, is to have the arrivals enter a preliminary location where they
test whether the facility is closed and, if so, exit the system. In ProModel, if a
location where entities are scheduled to arrive is unavailable at the time of an ar-
rival, the arriving entities are simply discarded.
6.4.7 Downtimes and Repairs

It is not uncommon for resources and even locations to unexpectedly go down or
become unavailable for one reason or another, such as a mechanical failure or a
personal interruption. Downtimes usually occur periodically as a function of total
elapsed time, time in use, or number of times used.
Downtimes Based on Total Elapsed Time

An example of a periodic downtime based on elapsed clock time might be a worker
who takes a break every two hours. Scheduled maintenance is also a type of down-
time performed at periodic intervals based on elapsed clock time. Figure 6.10 illus-
trates how a downtime based on elapsed time would be simulated. Notice that the
calculation of the interval between downtimes takes into account not only busy time,
but also idle time and downtime. In other words, it is the total elapsed time from the
start of one downtime to the start of the next. Downtimes based on total elapsed time
are often scheduled downtimes during which operational statistics on the location
or resource are suspended. ProModel allows you to designate whether a particular
downtime is to be counted as scheduled downtime or unscheduled downtime.
Sometimes it may be desirable to use elapsed time to define random equip-
ment failures. This is especially true if this is how historical data were gathered
on the downtime. When using historical data, it is important to determine if the
time between failure was based on (1) the total elapsed time from one failure to
the next, (2) the time between the repair of one failure to the time of the next fail-
ure (operational time between failures), or (3) the time that the machine was actu-
ally in operation (operating time between failures). ProModel accepts downtime
definitions based on cases 1 and 3, but requires that case 2 be converted to case 1.
This is done by adding the time until the next failure to the repair time of the last
FIGURE 6.10 Start Interrupt Interrupt

Resource downtime 6 10 4 6 14
occurring every
20 minutes based on
Idle Busy Idle Down Busy
total elapsed time.
Time (minutes)
har01307_ch06_161-192.indd 181 20/12/10 1:50 PM

FIGURE 6.11 Start Interrupt

Resource downtime 12 8
occurring every
20 minutes, based on
operating time. Idle Busy Idle Busy
Time (minutes)
failure. For example, if the operational time between failures is exponentially dis-
tributed with a mean of 10 minutes and the repair time is exponentially distributed
with a mean of 2 minutes, the time between failures should be defined as xlast ⫹
E(10) where xlast is the last repair time generated using E(2) minutes.
Downtimes Based on Time in Use

Most equipment and machine failures occur only when the resource is in use. A
mechanical or tool failure, for example, generally happens only when a machine
is running, not while a machine is idle. In this situation, the interval between
downtimes would be defined relative to actual machine operation time. A machine
that goes down every 20 minutes of operating time for a three-minute repair is il-
lustrated in Figure 6.11. Note that any idle times and downtimes are not included
in determining when the next downtime occurs. The only time counted is the
actual operating time.
Because downtimes usually occur randomly, the time to failure is most ac-
curately defined as a probability distribution. Studies have shown, for example,
that the operating time to failure is often exponentially distributed.
Downtimes Based on the Number of Times Used

The last type of downtime occurs based on the number of times a location was
used. For example, a tool on a machine may need to be replaced every 50 cycles
due to tool wear, or a copy machine may need paper added after a mean of 200
copies with a standard deviation of 25 copies. ProModel permits downtimes to be
defined in this manner by selecting ENTRY as the type of downtime and then speci-
fying the number of entity entries between downtimes.
Downtime Resolution
Unfortunately, data are rarely available on equipment downtime. When they are
available, they are often recorded as overall downtime and seldom broken down
into number of times down and time between failures. Depending on the nature
of the downtime information and degree of resolution required for the simulation,
downtimes can be treated in the following ways:
• Ignore the downtime.
• Simply increase processing times to adjust for downtime.
• Use average values for mean time between failures (MTBF) and mean
time to repair (MTTR).
• Use statistical distributions for time between failures and time to repair.
har01307_ch06_161-192.indd 182 20/12/10 1:50 PM

Ignoring Downtime. There are several situations where it might make sense
to ignore downtimes in building a simulation model. Obviously, one situation is
where absolutely no data are unavailable on downtimes. If there is no knowledge
of resource downtimes, it is appropriate to model the resource with no downtimes
and document it as such in the final write-up. When there are downtimes, but they
are extremely infrequent and not likely to affect model performance for the period
of the study, it is safe to ignore them. For example, if a machine fails only two or
three times a year and you are trying to predict processing capacity for the next
workweek, it doesn’t make sense to include the downtime. It is also safe to ignore
occasional downtimes that are very small compared to activity times. If, for ex-
ample, a downtime takes only seconds to correct (like clearing a part in a machine
or an occasional paper jam in a copy machine), it could be ignored.
Increasing Processing Times. A common way of treating downtime, due in part

to the lack of good downtime data, is to simply reduce the production capacity
of the machine by the downtime percentage. In other words, if a machine has an
effective capacity of 100 parts per hour and experiences a 10 percent downtime,
the effective capacity is reduced to 90 parts per hour. This spreads the downtime
across each machine cycle so that both the mean time between failures and the
mean time to repair are very small and both are constant. Thus no consideration
is given for the variability in both time between failures and time to repair that
typifies most production systems. Law (1986) has shown that this deterministic
adjustment for downtime can produce results that differ greatly from the results
based on actual machine downtimes.
MTBF/MTTR. Two parts to any downtime should be defined when modeling

downtime. One, time between failures, defines the interval between failures and is
measured as either the total elapsed time between failures, or better, the elapsed
time in actual use between failures. The other, time to repair, defines the time
required to bring a resource back online whenever it goes down. Often downtimes
are defined in terms of mean time between failures (MTBF) and mean time to re-
pair (MTTR). Using average times for these intervals presents the same problems
as using average times for any activity in a simulation: it fails to account for vari-
ability, which can have a significant impact on system performance.
Using Statistical Distributions. Whenever possible, time between failures and

time to repair should be represented by statistical distributions that reflect the
variation that is characteristic of these elements. Studies have shown that the
time until failure, particularly due to items (like bearings or tooling) that wear,
tends to follow a Weibull distribution. Repair times often follow a lognormal
distribution.
Elapsed Time or Usage Time?

When determining the distribution for time to failure, a distinction should be
made between downtime events that can occur anytime whether the resource
har01307_ch06_161-192.indd 183 20/12/10 1:50 PM

is operating or idle and downtime events that occur only when a resource is in
use. As explained earlier, downtimes that can occur anytime should be defined
as a function of clock time. If the resource goes down only while in operation, it
should be defined as a function of time in use.
Erroneously basing downtime on elapsed time when it should be based on
operating time artificially inflates time between failures by the inclusion of idle
time. It also implies that during periods of high equipment utilization, the same
amount of downtime occurs as during low utilization periods. Equipment failures
should generally be based on operating time and not on elapsed time because
elapsed time includes operating time, idle time, and downtime. It should be left to
the simulation to determine how idle time and downtime affect the overall elapsed
time between failures.
To illustrate the difference this can make, let’s assume that the following
times were logged for a given operation:
Status Time (Hours)
In use 20
Down 5
Idle 15
Total time 40
If it is assumed that downtime is a function of total time, then the percentage of

downtime would be calculated to be 5 hours/40 hours, or 12.5 percent. If, however,
it is assumed that the downtime is a function of usage time, then the downtime
percentage would be 5 hours/25 hours, or 20 percent. Now let’s suppose that the
system is modeled with increased input to the operation so that it is never starved
(the idle time ⫽ 0). If downtime is assumed to be 12.5 percent, the total time down
will be .125 × 40 ⫽ 5 hours. If, on the other hand, we use the assumption that it is
20 percent, then the time down will end up being .20 ⫻ 40 hours ⫽ 8 hours. This is
a difference of three hours, which means that if downtime is falsely assumed to be
a function of total time, the simulation would realize three extra hours of produc-
tion in a 40-hour period that shouldn’t have happened.
Handling Interrupted Entities

When a resource goes down, there might be entities that were in the middle of
being processed that are now left dangling (that is, they have been preempted).
For example, a machine might break down while running the seventh part of a
batch of 20 parts. The modeler must decide what to do with these entities. Several
alternatives may be chosen, and the modeler must select which alternative is the
most appropriate:
• Resume processing the entity after the downtime is over.
• Find another available resource to continue the process.
har01307_ch06_161-192.indd 184 20/12/10 1:50 PM

• Scrap the entity.

• Delay start of the downtime until the entity is processed.
The last option is the easiest way to handle downtimes and, in fact, may be
adequate in situations where either the processing times or the downtimes are
relatively short. In such circumstances, the delay in entity flow is still going to
closely approximate what would happen in the actual system.
If the entity resumes processing later using either the same or another re-
source, a decision must be made as to whether only the remaining process time is
used or if additional time must be added. By default, ProModel suspends process-
ing until the location or resource returns to operation. Alternatively, other logic
may be defined.
6.4.8 Use of Programming Logic

Sometimes the desired model behavior can’t be achieved using a “canned” con-
struct provided by the software. In these instances, it is necessary to use program-
ming logic that tests probabilities, variables, and attributes to make behavioral
decisions. A brief explanation of how these elements can be used in developing a
simulation model is given here.
Using Probabilities to Model Probabilistic Behavior

Sometimes operational elements (routings, operations, or others) exhibit ran-
dom behavior. To model a routing that occurs randomly, it is easy just to select
a probabilistic routing rule and specify the probability value. The routing is then
randomly chosen according to the probability specified. For operations that occur
randomly at a particular location, however, a probability function must be used in
connection with if–then logic. At an inspection and rework station, for example,
suppose that a product is defective 10 percent of the time and requires 3 ⫾ 1 min-
utes (uniformly distributed) to correct the defect. Assuming rand() is a function
that returns a random value between 0 and 1, the code for defining this behavior
might be as follows:
if rand() <= .10
then wait U(3,1) min
A similar situation occurs in activity times that have more than one dis-
tribution. For example, when a machine goes down, 30 percent of the time it
takes Triangular(0.2, 1.5, 3) minutes to repair and 70 percent of the time it takes
Triangular(3, 7.5, 15) minutes to repair. The logic for the downtime definition
might be
if rand() <= .30
then wait T(.2, 1.5, 3) min
else wait T(3, 7.5, 15) min
Another example of multidistribution activity is a call center that handles

three different types of calls that we will designate as A, B, and C. The time to
har01307_ch06_161-192.indd 185 20/12/10 1:50 PM

TABLE 6.1 Table of Service Times for Different Call Types

Call Type Probability of Occurrence Service Time (Minutes)
A .20 E(5)
B .50 E(8)
C .30 E(12)
handle calls is exponentially distributed, but the mean time is different depend-
ing on the type of call (see Table 6.1). To model this situation, the following code
might be entered (“//” are used at the beginning of a comment line):
// set a real variable (rValue) to a random number
real rValue = rand()
// test for call type A
if rValue <=.20
then wait E(5) min
else if rValue <=.70
then wait E(8) min
else wait E(12) min
Using Attributes to Model Special Decision Logic

Sometimes an item that has passed through the same location a second time re-
quires less time to be reworked than it took for the initial operation. In this situa-
tion, an attribute of the entity should be the basis of the decision because we need
to test how many times that particular entity has visited the location. An attribute
we will call Pass will need to be defined in the attribute module. If a normal opera-
tion takes five minutes but a rework operation takes only two minutes, the follow-
ing code would be entered in the operation logic for the location:
// increment the value of Pass by one
INC Pass
if Pass = 1
then wait 5 min
else wait 2 min
Using Global Variables to Gather Statistics

Global variables are placeholders for values that may change during the simula-
tion. What makes them global is that they are accessible from any place in the
model by any object in the model, and they exist during the entire simulation.
Global variables are defined by the user for user-defined purposes. For ex-
ample, a variable may be defined to keep track of the total number of entities in a
certain area of the system (work in process). Each time an entity enters the area,
it increments the variable. Also, each time an entity leaves the area, it decrements
the variable. Because the user has the option of collecting statistics on global
variables, either the simple or time-weighted average value of the variable can
be reported at the end of the simulation. A time series report showing all of the
changes in the variable over time can also be reported.
har01307_ch06_161-192.indd 186 20/12/10 1:50 PM

Using Local Variables for Looping

Local variables are variables that are declared within the logic itself, either right
before the point of use or at the beginning of the logic block. A local variable
exists only for the current object executing the block of logic. When the object
finishes executing the logic block in which the local variable was defined, the
variable disappears. A local variable may be thought of as a temporary attribute
of the object executing the logic because multiple objects executing the same
logic at the same time would each have their own local variables assigned to
them while executing the block of logic. Local variables are useful primarily
for executing a logic loop. Suppose, for example, that you wanted to assign
the value of 4 to the first 10 elements of an array called NumOfBins that was
defined in the Array module. To do this within operation logic (or in any other
logic), you would enter something like the following, where Count is defined
as a local variable:
int Count = 1
while Count < 11 do
{
NumOfBins[Count] = 4
Inc Count
}
The braces “{” and “}” are the ProModel notation (also used in C++ and
Java) for starting and ending a block of logic. In this case it is the block of state-
ments to be executed repeatedly by an object as long as the local variable Count is
less than 11.
6.5 Miscellaneous Modeling Issues

This section addresses special issues that may be encountered in simulation. They
don’t fit well under previous headings, so they are put in a collection here even
though they are not necessarily related.
6.5.1 Modeling Rare Occurrences

Often there exist situations in the real world that occur only rarely. For example,
a machine may break down once every two months, or only one in a thousand
parts is rejected at an inspection station. In simulation analysis, we are generally
interested in the normal behavior of the system, not extremely rare behavior.
It is advisable to ignore rare situations and exclude them from the simulation
model. This approach not only reduces modeling time but also simplifies the
model and helps maintain the focus of everyone involved in the model on key
input variables.
But some rare situations have a significant impact on the operation of the sys-
tem. For example, a plant shutdown may bring the entire operation to a stop. If the
focus of interest is to evaluate the effects of the rare occurrence, such as how long
har01307_ch06_161-192.indd 187 20/12/10 1:50 PM

it takes for inventories to be depleted if a shutdown occurs, then it makes sense to

include the rare occurrence in the model. The easiest way to model a rare event
is not to let the model run for six months or a year before the event occurs, but to
go ahead and force the event to happen. The point of modeling the rare event is to
see what impact it has on system behavior, not to predict when it actually might
occur. So it really doesn’t matter when it happens as long as the state of the model
is typical of what the system might be when it does occur.
6.5.2 Large-Scale Modeling

Occasionally it may be desirable to model a large system such as an entire
factory or the entire activity in an international airport. The tendency (espe-
cially for novice simulators) is to painstakingly piece together a huge, com-
plex model only to find that it runs nowhere near the way that was expected
and may not even run at all. The disappointment of not getting the model to
run correctly the first time is soon overshadowed by the utter despair in trying
to debug such an enormous model. When faced with building a supermodel, it
is always a good idea to partition the model into several submodels and tackle
the problem on a smaller scale first. Once each of the submodels has been
built and validated, they can be merged into a larger composite model. This
composite model can be structured either as a single monolithic model or as
a hierarchical model in which the details of each submodel are hidden unless
explicitly opened for viewing. Several ways have been described for merging
individual submodels into a composite model (Jayaraman and Agarwal 1996).
Three of the most common ways that might be considered for integrating sub-
models are
• Option 1: Integrate all of the submodels just as they have been built.
This approach preserves all of the detail and therefore accuracy of the
individual submodels. However, the resulting composite model may be
enormous and cause lengthy execution times. The composite model may
be structured as a flat model or, to reduce complexity, as a hierarchical
model.
• Option 2: Use only the recorded output from one or more of the
submodels. By simulating and recording the time at which each entity
exits the model for a single submodel, these exit times can be used in
place of the submodel for determining the arrival times for the larger
model. This eliminates the need to include the overhead of the individual
submodel in the composite model. This technique, while drastically
reducing the complexity of the composite model, may not be possible if
the interaction with the submodel is two-way. For submodels representing
subsystems that simply feed into a larger system (in which case the
subsystem operates fairly independently of downstream activities), this
technique is valid. An example is an assembly facility in which fabricated
components or even subassemblies feed into a final assembly line.
har01307_ch06_161-192.indd 188 20/12/10 1:50 PM

Basically, each feeder line is viewed as a “black box” whose output is

read from a file.
• Option 3: Represent the output of one or more of the submodels as
statistical distributions. This approach is the same as option 2, but
instead of using the recorded output times from the submodel in the
composite model, a statistical distribution is fit to the output times
and used to generate the input to the composite model. This technique
eliminates the need for using data files that, depending on the submodel,
may be quite large. Theoretically, it should also be more accurate
because the true underlying distribution is used instead of just a sample
unless there are discontinuities in the output. Multiple sample streams
can also be generated for running multiple replications.
6.5.3 Cost Modeling

Often it is desirable to include cost in a model to determine the most cost-effective
solution to a design problem. If, for example, two operators on an assembly
line result in the same performance as using three robots, the decision may end
up being based on cost rather than performance. There are two approaches to
modeling cost. One is to include cost factors in the model itself and dynami-
cally update cost collection variables during the simulation. ProModel includes
a cost module for assigning costs to different factors in the simulation such
as entity cost, waiting cost, and operation cost. The alternative approach is to
run a cost analysis after the simulation, applying cost factors to collected cost
drivers such as resource utilization or time spent in storage. The first method
is best when it is difficult to summarize cost drivers. For example, the cost per
unit of production may be based on the types of resources used and the time
for using each type. This may be a lot of information for each entity to carry
using attributes. It is much easier to simply update the entity’s cost attribute
dynamically whenever a particular resource has been used. Dynamic cost tracking
suffers, however, from requiring cost factors to be considered during the mode-
ling stage rather than the analysis stage. For some models, it may be difficult
to dynamically track costs during a simulation, especially when relationships
become very complex.
The preferred way to analyze costs, whenever possible, is to do a postsimula-
tion analysis and to treat cost modeling as a follow-on activity to system model-
ing rather than as a concurrent activity (see Lenz and Neitzel 1995). There are
several advantages to separating the logic model from the cost model. First, the
model is not encumbered with tracking information that does not directly affect
how the model operates. Second, and perhaps more importantly, post analysis of
costs gives more flexibility for doing “what-if ” scenarios with the cost model.
For example, different cost scenarios can be run based on varying labor rates in
a matter of seconds when applied to simulation output data that are immediately
available. If modeled during the simulation, a separate simulation would have to
be run applying each labor rate.
har01307_ch06_161-192.indd 189 20/12/10 1:50 PM

6.6 Summary
Model building is a process that takes a conceptual model and converts it to a
simulation model. This requires a knowledge of the modeling paradigm of the
particular simulation software being used and a familiarity with the different
modeling constructs that are provided in the software. Building a model involves
knowing what elements to include in the model and how to best express those
elements in the model. The principle of parsimony should always be followed,
which results in the most minimal model possible that achieves the simulation
objectives. Finally, to be successful at modeling you need to see lots of examples
and practice, practice, practice!

1. How is modeling an art as well as a science?
2. What is a modeling paradigm?
3. Describe the modeling paradigm used in a specific simulation product.
4. Suppose you were modeling an inspection station as an object. Identify
two attributes and one behavior you might want to define for it.
5. Identify five different things an entity might represent in a
manufacturing or service simulation.
6. For a manufacturing operation that produces 300 different part types,
how would you represent the different part types?
7. A stand and power cord are assembled to a monitor at an assembly
station. Stands are produced in-house and are not always available when
needed. Power cords are purchased finished and are always available
at the station. What entity types (monitor, cord, or stand) would you
include in the model and, conceptually, how would you model this
assembly operation?
8. Every hour, a stack of purchase orders is delivered to the purchasing
department for processing. The stack is routed sequentially to three
different people who review, approve, and send out the orders.
The orders move together as a stack, although they are processed
individually at each station. The processing time at each activity is a
function of the number of orders in the stack. The number of orders in
a stack can be determined by a probability distribution. Conceptually,
how would you model this process and still capture the essence of what
is happening?
9. What criteria would you use to identify the route stops or locations in a
flow sequence to include in a model?
10. Suppose you are modeling a system with bins in which parts are
transported. Would you model the bins as resources or entities? Justify
your answer.
har01307_ch06_161-192.indd 190 20/12/10 1:50 PM

11. How would you model a consumable resource such as energy

consumption that occurs based on the length of time a particular
machine is in operation?
12. What is the danger of using simple time values to model material
movement if the material handling system often encounters traffic
delays?
13. How would you use an entity attribute to model multiple routing
passes through the same location in which a different operation time is
required depending on the pass?
14. An operator must inspect every fifth part at a manual workstation.
Conceptually, how would you model this activity if the operation takes
five minutes and inspection takes two minutes?
15. When attaching entities to another entity, under what circumstances
would you want to preserve the identities of the entities being attached?
16. Define four types of arrivals and give an example of each.
17. A factory operates with two eight-hour shifts (changeover occurs
without interruption) five days a week. If you wanted to simulate the
system for four weeks, how would you set the run length in hours?
18. What is the problem with modeling downtimes in terms of mean time
between failures (MTBF) and mean time to repair (MTTR)?
19. Why should unplanned downtimes or failures be defined as a function
of usage time rather than total elapsed time on the clock?
20. In modeling repair times, how should the time spent waiting for a
repairperson be modeled?
21. What is preemption? What activities or events might preempt other
activities in a simulation?
22. A boring machine experiences downtimes every five hours
(exponentially distributed). It also requires routine preventive
maintenance (PM) after every eight hours (fixed) of operation. If a
downtime occurs within two hours of the next scheduled PM, the
PM is performed as part of the repair time (no added time is needed)
and, after completing the repair coupled with the PM, the next PM
is set for eight hours away. Conceptually, how would you model this
situation?
23. A real estate agent schedules six customers (potential buyers) each
day, one every 1.5 hours, starting at 8 A.M. Customers are expected to
arrive for their appointments at the scheduled times. However, past
experience shows that customer arrival times are normally distributed
with a mean equal to the scheduled time and a standard deviation
of five minutes. The time the agent spends with each customer
is normally distributed with a mean of 1.4 hours and a standard
deviation of .2 hours. Develop a simulation model to calculate the
expected waiting time for customers.
har01307_ch06_161-192.indd 191 20/12/10 1:50 PM

References
Jayaraman, Arun, and Arun Agarwal. “Simulating an Engine Plant.” Manufacturing Engi-
neering, November 1996, pp. 60–68.
Law, A. M. “Introduction to Simulation: A Powerful Tool for Analyzing Complex Manu-
facturing Systems.” Industrial Engineering, 1986, 18(5):57–58.
Lenz, John, and Ray Neitzel. “Cost Modeling: An Effective Means to Compare Alterna-
tives.” Industrial Engineering, January 1995, pp. 18–20.
Shannon, Robert E. “Introduction to the Art and Science of Simulation.” In Proceedings
of the 1998 Winter Simulation Conference, ed. D. J. Medeiros, E. F. Watson, J. S.
Carson, and M. S. Manivannan. Piscataway, NJ: Institute of Electrical and Electronics
Engineers, 1998.
Thompson, Michael B. “Expanding Simulation beyond Planning and Design.” Industrial
Engineering, October 1994, pp. 64–66.
har01307_ch06_161-192.indd 192 20/12/10 1:50 PM

C H A P T E R
7 MODEL VERIFICATION
AND VALIDATION
“Truth comes out of error more easily than out of confusion.”

—Francis Bacon
7.1 Introduction
Building a simulation model is much like developing an architectural plan for a
house. A good architect will review the plan and specifications with the client or
owner of the house to ensure that the design meets the client’s expectations. The
architect will also carefully check all dimensions and other specifications shown on
the plan for accuracy. Once the architect is reasonably satisfied that the right infor-
mation has been accurately represented, a contractor can be given the plan to begin
building the house. In a similar way the simulation analyst should examine the valid-
ity and correctness of the model before using it to make implementation decisions.
In this chapter we cover the importance and challenges associated with model
verification and validation. We also present techniques for verifying and validat-
ing models. Balci (1997) provides a taxonomy of more than 77 techniques for
model verification and validation. In this chapter we give only a few of the more
common and practical methods used. The greatest problem with verification and
validation is not one of failing to use the right techniques but failing to use any
technique. Questions addressed in this chapter include the following:
• What are model verification and validation?
• What are obstacles to model verification and validation?
• What techniques are used to verify and validate a model?
• How are model verification and validation maintained?
Two case studies are presented at the end of the chapter showing how verifi-
cation and validation techniques have been used in actual simulation projects.
193
har01307_ch07_193-210.indd 193 20/12/10 1:50 PM

7.2 Importance of Model Verification and Validation

Model building is, by nature, very error prone. The modeler must translate the
real-world system into a conceptual model that, in turn, must be translated into a
simulation model. This translation process is iterative as the modeler’s conception
of the real system changes and the modeling approach continues to be refined (see
Figure 7.1). In this translation process, there is plenty of room for making errors.
Verification and validation processes have been developed to reduce and ideally
eliminate these errors.
Model verification is the process of determining whether the simulation model
correctly reflects the conceptual model. Model validation is the process of determin-
ing whether the conceptual model correctly reflects the real system. Model verifi-
cation and validation are critical to the success of a simulation project. Important
decisions may be based on the outcome of the simulation experiment, and, therefore,
demonstrable evidence should exist for the validity of the model. For simulations
that are used on an ongoing basis for operational decisions such as production sched-
uling, model validation is even more important. The Department of Defense (DoD)
has recognized the importance of model validation and has established a defined set
of criteria for determining model accreditation. Accreditation, as defined by DoD
(http://www.dtic.mil/whs/directives/corres/pdf/500067p.pdf), is “The official certi-
fication that a model or simulation is acceptable for use for a specific purpose.”
7.2.1 Reasons for Neglect

Despite their importance, seldom is adequate attention given to verification and
validation when doing simulation. The primary reasons for neglecting this impor-
tant activity are
FIGURE 7.1 System

Two-step translation
process to convert a
real-world system to
a simulation model.
n
t io
Concept
l i da
Va
Ve
r ific
at
io
n
Model
har01307_ch07_193-210.indd 194 20/12/10 1:50 PM

Chapter 7 Model Verification and Validation 195
• Time and budget pressures.

• Laziness.
• Overconfidence.
• Ignorance.
Time and budget pressures can be overcome through better planning and in-
creased proficiency in the validation process. Because validation doesn’t have to
be performed to complete a simulation study, it is the activity that is most likely
to get shortchanged when pressure is being felt to complete the project on time or
within budget. Not only does the modeler become pressed for time and resources,
but often others on whom the modeler relies for feedback on model validity also
become too busy to get involved the way they should.
Laziness is a bit more difficult to deal with because it is characteristic of
human nature and is not easily overcome. Discipline and patience must be devel-
oped before one is willing to painstakingly go back over a model once it has been
built and is running. Model building is a creative activity that can actually be quite
fun. Model verification and validation, on the other hand, are a laborious effort
that is too often dreaded.
The problem of overconfidence can be dealt with only by developing a more
critical and even skeptical attitude toward one’s own work. Too often it is
assumed that if the simulation runs, it must be okay. A surprising number of deci-
sions are based on invalid simulations simply because the model runs and pro-
duces results. Computer output can give an aura of validity to results that may be
completely in error. This false sense of security in computer output is a bit like
saying, “It must be true, it came out of the computer.” Having such a naive men-
tality can be dangerously misleading.
The final problem is simply a lack of knowledge of verification and validation
procedures. This is a common problem particularly among newcomers to simula-
tion. This seems to be one of the last areas in which formal training is received
because it is presumed to be nonessential to model building and output analysis.
Another misconception of validation is that it is just another phase in a simulation
project rather than a continuous activity that should be performed throughout the
entire project. The intent of this chapter is to dispel common misconceptions, help
shake attitudes of indifference, and provide insight into this important activity.
7.2.2 Practices That Facilitate Verification and Validation

A major problem in verifying and validating models occurs because of poor
modeling practices. Beginners often build models with little or no thought about
being able to verify or validate the model. Often these models contain spa-
ghetti code that is difficult for anyone, including the modeler, to follow. This
problem becomes even more acute as models grow in complexity. When the
original model builder moves on and others inherit these models, the desperate
efforts to unravel the tangled mess and figure out what the model creator had
in mind become almost futile. It is especially discouraging when attempting to
use a poorly constructed model for future experimentation. Trying to figure out
har01307_ch07_193-210.indd 195 20/12/10 1:50 PM

what changes need to be made to model new scenarios becomes difficult if not
impossible.
The solution to creating models that ease the difficulty of verification and
validation is to first reduce the amount of complexity of the model. Frequently
the most complex models are built by amateurs who do not have sense enough
to know how to abstract system information. They code way too much detail
into the model. Once a model has been simplified as much as possible, it needs
to be coded so it is easily readable and understandable. Using object-oriented
techniques such as encapsulation can help organize model data. The right simu-
lation software can also help keep model data organized and readable by provid-
ing table entries and intelligent, parameterized constructs rather than requiring
lots of low-level programming. Finally, model data and logic code should be
thoroughly and clearly documented. This means that every subroutine used
should have an explanation of what it does, where it is invoked, what the param-
eters represent, and how to change the subroutine for modifications that may be
anticipated.
7.3 Model Verification

Verification is the process of determining whether the model operates as intended.
It doesn’t necessarily mean that the model is valid, only that it runs correctly.
According to Banks et al. (2001), “Verification is concerned with building the
model right. It is utilized in the comparison of the conceptual model to the com-
puter representation that implements that conception.” Model verification in-
volves the modeler more than the customer. During the verification process, the
modeler tries to detect unintended errors in the model data and logic and remove
them. In essence, verification is the process of debugging the model. A verified
model is a bug-free model.
Errors or bugs in a simulation model are of two types: syntax errors and se-
mantic errors. Syntax errors are like grammatical errors and include the uninten-
tional addition, omission, or misplacement of notation that either prevents the
model from running or causes it to run incorrectly. The omission or misplacement
of a decimal point in a number or parentheses in an expression can dramatically
impact the outcome of a simulation. Most simulation software has built-in error
detection that can identify inconsistencies and omissions. This feature is much
like spelling and grammar checking in a word processor. As with a word pro-
cessor, however, it still remains the responsibility of the user to ensure that the
intended meaning has been conveyed.
Semantic errors are associated with the meaning or intention of the modeler
and are therefore difficult to detect. Often they are logical errors that cause behav-
ior that is different than what was intended. For example, an if–then test on a vari-
able may branch to the wrong logic simply because the condition was incorrectly
specified or the branching was not set up correctly. Or numerical values may
be misentered and therefore result in the wrong intended behavior. Verification
har01307_ch07_193-210.indd 196 20/12/10 1:50 PM

techniques discussed here, such as event tracing and code debugging, can help
identify and correct semantic errors.
7.3.1 Preventive Measures

Obviously, the preferred method of dealing with model bugs is to avoid build-
ing models with bugs in the first place. In practice, this isn’t always possible as
bugs are often sneaky and difficult to prevent. However, a few basic practices can
minimize the number of bugs that find their way into models. The most important
practice is to simply be careful. This means focusing on what you are doing and
eliminating distractions that may cause you to make silly typing mistakes. You
should check and double-check the data and your entry of those data into the
model to make sure there are no transcription errors. Another practice that has
been used successfully for years in software development for ensuring good code
is called structured programming. Hoover and Perry (1990) describe five basic
principles of structured programming:
1. Top-down design. The simulation model and the code begin at a high
level, and then lower-level areas are added that are congruent with, yet
subordinate to, the top level. It is a hierarchical approach to modeling.
A routing to a six-machine cell may initially be defined to the cell
as a whole and then later decomposed to route to each individual
workstation.
2. Modularity. The model is built in modules or logical divisions to
simplify model development and debugging and to facilitate reuse.
Where possible, data modules should be separate from the logic modules
so that data can be easily accessed.
3. Compact modules. Modules should be kept as short and simple as
possible. Subroutines consisting of over 25 statements, for example,
should probably be broken into multiple subroutines.
4. Stepwise refinement. The model is built with complexity being
progressively added. In stepwise refinement, one goes from general to
specific so that generalities and assumptions are refined into precise
and accurate details. The idea is to create the model in multiple passes,
filling in additional details with each pass. At each pass, a test for
reasonableness can be made. It is easier to verify a model when the
model is built incrementally than when it is built all at once.
5. Structured control. GOTO statements or other unstructured branching
of control should be avoided wherever possible as they may lead to
unexpected results. Logic control should be based on structured control
statements such as
IF-THEN-ELSE
WHILE...DO
DO...WHILE
(etc.)
har01307_ch07_193-210.indd 197 20/12/10 1:50 PM

7.3.2 Establishing a Standard for Comparison

Verifying a model presupposes that some standard exists for comparison, provid-
ing a sense of whether the model is running correctly. Obviously, if you knew
exactly how the simulation should run and what results should be expected, you
wouldn’t need to be doing the simulation in the first place. Several reasonable
standards can be used for model comparison. One simple standard is common
sense (although someone has said that common sense isn’t all that common). If
your simulation is giving totally bizarre results, you know you have a problem
somewhere in your model. Say, for example, that you define a machine downtime
to occur every two hours for two minutes and yet the simulation results show that
the machine was down 80 percent of the time. Common sense tells you that the
results don’t make sense and indicates that you should check how you defined the
downtime behavior.
Another way to establish a sense of correctness is to construct an analytic
model of the problem if at all possible. Because of the simplified assumptions
made in most analytic techniques, it may be necessary to scale back the model to
verify only its basic operation. It may be necessary, for example, to run the simu-
lation without downtimes because failures may complicate an analytic formula-
tion. As additional examples, we may operate a queuing system without balking,
or we may use instantaneous replenishment rather than random lead times in an
inventory model. In these cases, the simplified model may be one that can be
analyzed without simulation, and the mathematical results can then be compared
to those from the simulation. In this way we can get an indication of whether the
simulation model, or at least a simplified version of it, is correct.
7.3.3 Verification Techniques

Once everything has been done at the outset to ensure that the model has been
built correctly and a standard for comparison has been established, several tech-
niques can be used to verify the model. Some of the more common ways include
• Conduct model code reviews.
• Check the output for reasonableness.
• Watch the animation for correct behavior.
• Use the trace and debug facilities provided with the software.
Reviewing Model Code

A model code review can be conducted by either the modeler or someone else
familiar with both modeling and the system being modeled. The purpose of the
model code review is to check for errors and inconsistencies. The simulation
model can be tested in either a bottom-up or top-down fashion. In the bottom-up
testing method, also called unit testing, the lowest modules are tested and verified
first. Next, interfaces between two or more modules are tested. This is continued
until the model can be tested as a single system.
har01307_ch07_193-210.indd 198 20/12/10 1:50 PM

In the top-down approach, the verification testing begins with the main mod-
ule and moves down gradually to lower modules. At the top level, you are more
interested that the outputs of modules are as expected given the inputs. If discrep-
ancies arise, lower-level code analysis is conducted.
In both approaches sample test data are used to test the program code. The
bottom-up approach typically requires a smaller set of data to begin with. After
exercising the model using test data, the model is stress tested using extreme input
values. With careful selection, the results of the simulation under extreme condi-
tions can be predicted fairly well and compared with the test results.
Checking for Reasonable Output

In any simulation model, there are operational relationships and quantitative val-
ues that are predictable during the simulation. An example of an operational re-
lationship is a routing that occurs whenever a particular operation is completed.
A predictable quantitative value might be the number of boxes on a conveyor
always being between zero and the conveyor capacity. Often the software itself
will flag values that are out of an acceptable range, such as an attempt to free more
resources than were captured.
For simple models, one way to help determine reasonableness of the output
is to replace random times and probabilistic outcomes with constant times and
deterministic outcomes in the model. This allows you to predict precisely what
the results will be because the analytically determined results should match the
results of the simulation.
Watching the Animation

Animation can be used to visually verify whether the simulation operates the way
you think it should. Errors are detected visually that can otherwise go unnoticed.
The animation can be made to run slowly enough for the analyst to follow along
visually. However, the amount of time required to observe an entire simulation
run can be extremely long. If the animation is sped up, the runtime will be smaller,
but inconsistent behavior will be more difficult to detect.
To provide the most meaningful information, it helps to have interactive
simulation capability, which allows the modeler to examine data associated with
simulation objects during runtime. For example, during the simulation, the user
can view the current status of a workstation or resource to see if state variables are
being set correctly.
Animation is usually more helpful in identifying a problem than in discover-
ing the cause of a problem. The following are some common symptoms apparent
from the animation that reveal underlying problems in the model:
• The simulation runs fine for hours and even days and then suddenly freezes.
• A terminating simulation that should have emptied the system at the end
of the simulation leaves some entities stranded at one or more locations.
• A resource sits idle when there is plenty of work waiting to be done.
• A particular routing never gets executed.
har01307_ch07_193-210.indd 199 20/12/10 1:50 PM

Using Trace and Debugging Facilities

Trace and debugging information provides detailed textual feedback of what is
happening during a simulation. This allows the modeler to look under the hood
and see exactly what is going on inside the simulation. Most simulation software
comes with some sort of trace and debugging facility. At a minimum, manual
trace or debugging information can be coded into the simulation model that is
either deleted or disabled once the model is debugged. The following logic,
for example, might be inserted somewhere in the model logic to test whether a
resource is available at a specific time.
If OperatorState = busy
Then display “Operator is busy”
Else display “Operator is NOT busy”
Trace messages describe chronologically what happens during the simula-

tion, event by event. A typical trace message might be the time that an entity
enters a particular location or the time that a specific resource is freed. Trace mes-
saging can be turned on or off, and trace messages can usually be either displayed
directly on the screen or written to a file for later analysis. In ProModel, tracing
can also be turned on and off programmatically by entering a Trace command at
appropriate places within the model logic. A segment of an example of a trace
message listing is shown in Figure 7.2. Notice that the simulation time is shown
in the left column followed by a description of what is happening at that time in
the right column.
Another way in which behavior can be tracked during a simulation run is
through the use of a debugger. Anyone who has used a modern commercial com-
piler to do programming is familiar with debuggers. A simulation debugger is a
utility that displays and steps through the actual logic entered by the user to define
the model. As logic gets executed, windows can be opened to display variable and
FIGURE 7.2
Fragment of a trace
listing.
har01307_ch07_193-210.indd 200 20/12/10 1:50 PM

FIGURE 7.3
ProModel debugger
window.
state values as they dynamically change. Like trace messaging, debugging can
be turned on either interactively or programmatically. An example of a debugger
window is shown in Figure 7.3.
Experienced modelers make extensive use of trace and debugging capabili-
ties. Animation and output reports are good for detecting problems in a simula-
tion, but trace and debug messages help uncover why problems occur.
Using trace and debugging features, event occurrences and state variables
can be examined and compared with hand calculations to see if the program is
operating as it should. For example, a trace list might show when a particular op-
eration began and ended. This can be compared against the input operation time
to see if they match.
One type of error that a trace or debugger is useful in diagnosing is gridlock.
This situation is caused when there is a circular dependency where one action
depends on another action, which, in turn, is dependent on the first action. You
may have experienced this situation at a busy traffic intersection or when trying
to leave a crowded parking lot after a football game. It leaves you with a sense
of utter helplessness and frustration. An example of gridlock in simulation some-
times occurs when a resource attempts to move a part from one station to another,
but the second station is unavailable because an entity there is waiting for the
same resource. The usual symptom of this situation is that the model appears to
freeze up and no more activity takes place. Meanwhile the simulation clock races
rapidly forward. A trace of events can help detect why the entity at the second sta-
tion is stuck waiting for the resource.
har01307_ch07_193-210.indd 201 20/12/10 1:50 PM

7.4 Model Validation

Validation is the process of determining whether the model is a meaningful and
accurate representation of the real system (Hoover and Perry 1990). Where veri-
fication is concerned with building the model right, validation is concerned with
building the right model. For this reason, stakeholders and customers should be-
come heavily involved in the validation process.
Because the process of validation can be very time-consuming and ultimately
somewhat elusive, we are not interested in achieving absolute validity but only
functional validity. As Neelamkavil (1987) explains, “True validation is a philo-
sophical impossibility and all we can do is either invalidate or ‘fail to invalidate.’ ”
From a functional standpoint, model validation is viewed as the process of estab-
lishing “that the model’s output behavior has sufficient accuracy for the model’s
intended purpose over the domain of the model’s intended applicability” (Sargent
1998). Through validation we convey to the user or the customer that the simula-
tion results can be trusted and used to make real-world decisions. A model can be
valid even when many simplifications may have been made and details omitted
while building the model. What we are seeking is not certainty but rather a com-
fort level in the results.
For existing systems or when building an “as-is” model, the model behavior
should correspond to that of the actual system. In the case of a new system that
is being designed, and therefore has no demonstrated performance, the input data
should accurately reflect the design specification of the system and the model
should be carefully verified.
A model may run correctly (it has been verified) but still not be an accurate
representation of the system (it is invalid). Validation can be thought of as a reality
check (not to be confused with a sanity check, which people who do simulation
may frequently need). If a model is based on accurate information, and the model
has been verified to work as intended, then the model can be said to be a valid
model. A model that has bugs cannot be a valid model (unless purely by accident),
thus the importance of verification in the validation process.
Validation actually begins with the data-gathering stage of a simulation proj-
ect before a model is even built. It is the garbage-in, garbage-out syndrome that
should be avoided. If you use erroneous data in the model, the model may work
perfectly fine (it has been verified) yet still give misleading results. Data valida-
tion issues were addressed in Chapter 5, “Data Collection and Analysis.” By way
of review, data validation is a matter of verifying that the sources of information
are reliable and that source data have been correctly interpreted. For example, the
process used to fit distributions to sample data may be checked. Any assumptions
should also be reviewed and approved by those in a position to give an informed
opinion about the validity of the assumptions. Having valid information on the
system being modeled means that the conceptual model of the system is valid.
At this point, creating a valid simulation model essentially consists of creating
a model that is verified or, in other words, that corresponds with the conceptual
model.
har01307_ch07_193-210.indd 202 20/12/10 1:50 PM

7.4.1 Determining Model Validity

There is no simple test to determine the validity of a model. Validation is an in-
ductive process in which the modeler draws conclusions about the accuracy of
the model based on the evidence available. Several techniques are given here that
are described by Sargent (1998). Some of them tend to be rather subjective, while
others lend themselves to statistical tests such as hypothesis tests and confidence
intervals. Many of the techniques are the same as those used to verify a model,
only now the customer and others who are knowledgeable about the system need
to be involved. As with model verification, it is common to use a combination of
techniques when validating a model.
• Watching the animation. The visual animation of the operational behavior
of the model is compared with one’s knowledge about how the actual
system behaves. This can include dynamic plots and counters that provide
dynamic visual feedback.
• Comparing with the actual system. Both the model and the system are run
under the same conditions and using the same inputs to see if the results
match.
• Comparing with other models. If other valid models have been built of
the process such as analytic models, spreadsheet models, and even other
valid simulation models, the output of the simulation can be compared to
these known results. For example, a simple simulation model might be
compared to a validated queuing model.
• Conducting degeneracy and extreme condition tests. There are known
situations for which model behavior degenerates, causing a particular
response variable to grow infinitely large. An example is the number of
items in a queue prior to a server where the service rate is less than the
arrival rate. These situations can be forced in the model (by increasing
the number of arrivals) to see if the model degenerates as expected. The
model can also be run under extreme conditions to see if it behaves as
would be expected. For example, if arrivals are cut off, the model should
eventually run dry of entities.
• Checking for face validity. Face validity is checked by asking people who
are knowledgeable about the system whether the model and its behavior
appear reasonable. This technique can be used in determining if the
logic in the conceptual model is correct and if a model’s input–output
relationships are reasonable.
• Testing against historical data. If historical information exists for both
operating and performance data, the model can be tested using the same
operating data and comparing the results with the historical performance
data. This is sort of an “as-once-was” model.
• Performing sensitivity analysis. This technique consists of changing
model input values to determine the effect on the model’s behavior and
its output. The same relationships should occur in the model as in the
har01307_ch07_193-210.indd 203 20/12/10 1:50 PM

real system. The modeler should have at least an intuitive idea of how the
model will react to a given change. It should be obvious, for example,
that doubling the number of resources for a bottleneck operation should
increase, though not necessarily double, throughput.
• Running traces. An entity or sequence of events can be traced through the
model processing logic to see if it follows the behavior that would occur
in the actual system.
• Conducting Turing tests. People who are knowledgeable about the
operations of a system are asked if they can discriminate between system
and model outputs. If they are unable to detect which outputs are the
model outputs and which are the actual system outputs, this is another
piece of evidence to use in favor of the model being valid.
A common method of validating a model of an existing system is to compare
the model performance with that of the actual system. This approach requires that
an as-is simulation be built that corresponds to the current system. This helps “cali-
brate” the model so that it can be used for simulating variations of the same model.
After running the as-is simulation, the performance data are compared to those of
the real-world system. If sufficient data are available on a performance measure
of the real system, a statistical test can be applied called the Student’s t test to
determine whether the sampled data sets from both the model and the actual system
come from the same distribution. An F test can be performed to test the equality
of variances of the real system and the simulation model. Some of the problems
associated with statistical comparisons include these:
• Simulation model performance is based on very long periods of time.
On the other hand, real system performances are often based on much
shorter periods and therefore may not be representative of the long-term
statistical average.
• The initial conditions of the real system are usually unknown and are
therefore difficult to replicate in the model.
• The performance of the real system is affected by many factors that may
be excluded from the simulation model, such as abnormal downtimes or
defective materials.
The problem of validation becomes more challenging when the real system
doesn’t exist yet. In this case, the simulation analyst works with one or more ex-
perts who have a good understanding of how the real system should operate. They
team up to see whether the simulation model behaves reasonably in a variety of
situations. Animation is frequently used to watch the behavior of the simulation
and spot any nonsensical or unrealistic behavior.
The purpose of validation is to mitigate the risk associated with making deci-
sions based on the model. As in any testing effort, there is an optimum amount
of validation beyond which the return may be less valuable than the investment
(see Figure 7.4). The optimum effort to expend on validation should be based
on minimizing the total cost of the validation effort plus the risk cost associated
har01307_ch07_193-210.indd 204 20/12/10 1:50 PM

FIGURE 7.4 Optimum effort

Optimum level of
validation looks at Total cost
the trade-off between Validation cost
validation cost and
Cost
risk cost.
Risk cost
Validation effort
with making a decision based on an invalid model. As mentioned previously, the

problem is seldom one of devoting too much time to validation, but in devoting
too little.
7.4.2 Maintaining Validation

Once a model is validated for a given system specification, it is sometimes nec-
essary to maintain validation as system specifications tend to evolve right up to,
and often even after, system implementation. When simulation is used for system
design, rarely is the system implemented the way it was originally simulated,
especially when traditional, over-the-wall handoff procedures are used. These
continuous changes make it nearly impossible to validate the simulation against
the final implemented system. Unfortunately, results of the implemented system
sometimes get compared with the early simulation model, and the simulation is
unfairly branded as being invalid. This is another good reason for continuing to
update the model to continually reflect current system design specifications.
One method that works effectively for maintaining validation of a model
designed for ongoing use is to deny users access to the model data so that modi-
fications to the model are controlled. ProModel, for example, allows a modeler
to protect the model data using the packaging option under the File menu. This
prevents anyone from changing the actual model data directly. Modifications to
the model can be made only through a custom interface that controls precisely
what parameters of the model can be changed as well as the extent to which they
can be changed.
7.4.3 Validation Examples

To illustrate how the validation process might be performed for a typical sim-
ulation, we look at two simulation case studies. The first one was conducted
at St. John Hospital and Medical Center for an obstetrical renovation project
(O’Conner 1994). The second project was conducted at HP for a surface mount
assembly line (Baxter and Johnson 1993). The St. John case study uses a more
har01307_ch07_193-210.indd 205 20/12/10 1:50 PM

informal and intuitive approach to validation, while the HP case study relies on
more formal validation techniques.
St. John Hospital and Medical Center Obstetrical Unit

At St. John Hospital and Medical Center, a simulation study was conducted by
the Management Engineering Department to plan a renovation of the obstetrical
unit to accommodate modern family demands to provide a more private, homelike
atmosphere with a family-oriented approach to patient care during a mother’s hos-
pital stay. The current layout did not support this approach and had encountered
inefficiencies when trying to accommodate such an approach (running newborns
back and forth and the like).
The renovation project included provision for labor, delivery, recovery, and
postpartum (LDRP) rooms with babies remaining with their mothers during the
postpartum stay. The new LDRPs would be used for both low- and high-risk
mothers. The hospital administration wanted to ensure that there would be enough
room for future growth while still maintaining high utilization of every bed.
The purpose of the simulation was to determine the appropriate number of
beds needed in the new LDRPs as well as in other related areas of the hospital. A
secondary objective was to determine what rules should be used to process patients
through the new system to make the best utilization of resources. Because this was
a radically new configuration, an as-is model was not built. This, of course, made
model validation more challenging. Data were gathered based on actual operating
records, and several assumptions were made such as the length of stay for different
patient classifications and the treatment sequence of patients.
The validation process was actually a responsibility given to a team that was
assembled to work with the simulation group. The team consisted of the nursing
managers responsible for each OB area and the Women’s Health program director.
The first phase of the review phase was to receive team approval for the assump-
tions document and flowcharts of the processes. This review was completed prior
to developing the model. This assumptions document was continually revised and
reviewed throughout the duration of the project.
The next phase of the review came after all data had been collected and
analyzed. These reviews not only helped ensure the data were valid but estab-
lished trust and cooperation between the modeling group and the review team.
This began instilling confidence in the simulation even before the results were
obtained.
Model verification was performed during model building and again at model
completion. To help verify the model, patients, modeled as entities, were traced
through the system to ensure that they followed the same processing sequence
shown on the flowcharts with the correct percentages of patients following each
probabilistic branch. In addition, time displays were placed in the model to verify
the length of stay at each step in the process.
Model validation was performed with both the nursing staff and physician
representatives. This included comparing actual data collected with model results
where comparisons made sense (remember, this was a new configuration being
har01307_ch07_193-210.indd 206 20/12/10 1:50 PM

modeled). The animation that showed the actual movement of patients and use of
resources was a valuable tool to use during the validation process and increased
the credibility of the model.
The use of a team approach throughout the data-gathering and model devel-
opment stages proved invaluable. It both gave the modeler immediate feedback
regarding incorrect assumptions and boosted the confidence of the team members
in the simulation results. The frequent reviews also allowed the group to bounce
ideas off one another. As the result of conducting a sound and convincing simula-
tion study, the hospital’s administration was persuaded by the simulation results
and implemented the recommendations in the new construction.
HP Surface Mount Assembly Line

At HP a simulation study was undertaken to evaluate alternative production lay-
outs and batching methods for a surface mount printed circuit assembly line based
on a projected growth in product volume/mix. As is typical in a simulation proj-
ect, only about 20 percent of the time was spent building the model. The bulk of
the time was spent gathering data and validating the model. It was recognized
that the level of confidence placed in the simulation results was dependent on the
degree to which model validity could be established.
Surface mount printed circuit board assembly involves placing electronic
components onto raw printed circuit boards. The process begins with solder paste
being stenciled onto the raw circuit board panel. Then the panel is transported
to sequential pick-and-place machines, where the components are placed on the
panel. The component feeders on the pick-and-place machines must be changed
depending on the components required for a particular board. This adds setup
time to the process. A third of the boards require manual placement of certain
components. Once all components are placed on the board and solder pasted, the
board proceeds into an oven where the solder paste is cured, bonding the compo-
nents to the panel. Finally, the panel is cleaned and sent to testing. Inspection steps
occur at several places in the line.
The HP surface mount line was designed for a high-mix, low-volume produc-
tion. More than 100 different board types are produced on the line with about 10
different types produced per day. The batch size for a product is typically less than
five panels, with the typical panel containing five boards.
A simulation expert built the model working closely with the process
owners—that is, the managers, engineers, and production workers. Like many
simulations, the model was first built as an as-is model to facilitate model valida-
tion. The model was then expanded to reflect projected requirements and proposed
configurations.
The validation process began early in the data-gathering phase with meetings
being held with the process owners to agree on objectives, review the assump-
tions, and determine data requirements. Early involvement of the process owners
gave them motivation to see that the project was successful. The first step was to
clarify the objectives of the simulation that would guide the data-gathering and
modeling efforts. Next a rough model framework was developed, and the initial
har01307_ch07_193-210.indd 207 20/12/10 1:50 PM

assumptions were defined based on process information gathered from the pro-
cess owners. As the model development progressed, further assumptions were
made, and decisions were made and agreed upon regarding the level of detail
required to capture the essence of the process.
The flow logic and move times were built into the model, but processing and
setup times, since they were so varied and numerous, were put into a common
lookup table so they could be verified and modified easily. The particular board
sequence and batch size were also defined for the model.
The next step was to verify the model to ensure that it operated as intended.
The simulation output was reviewed for reasonableness and a trace was used to
check the model logic where it was complex. In reality, much of this verification
was performed as the model was built because it was necessary for getting the
model just to run. Note that the modeler was the one primarily responsible for
verifying or debugging the model. This makes sense inasmuch as the modeler had
a clear specification and was the one responsible for fashioning the model to fit
the specification.
After the modeler was satisfied that the model was running correctly, all of
the assumptions were reviewed again by the production operators, supervisors,
and engineers. These assumptions were more meaningful as they were able to
watch their impact on the simulation performance. Animation was used to com-
municate model behavior and demonstrate the scope of the model. Those who
were familiar with the operation of the real line were able to confirm whether
the model appeared to be working reasonably correctly. When explaining how
the model was built, it was important to describe the modeling constructs using
familiar terminology because the process owners were not versed in modeling
terminology.
At this point it would be easy to conclude that the model was valid and ready
for use in conducting experiments. The challenging question with the model be-
came “Is it the right model?” A simulation model is viewed much more critically
once it is running and it is realized that important decisions will soon be based
on the results of the simulation experiments. With the model running correctly
and providing output that was consistent with the assumptions that were made,
it was felt that the assumptions should be challenged more critically. Because
the initial data that were gathered—that is, processing times and setup times—
were based on average times with no reflection of variability, decisions had to be
made about how much more data should be gathered or what modifications in as-
sumptions needed to be made before there was sufficient confidence in the model.
Component placement times were studied using classical time-study techniques
to achieve more realistic time estimates. Next the average placement times were
used to calculate the average run time for a particular panel, and these were com-
pared to the recorded run times.
Historical data and further time studies were used to validate the output. A
particular sequence of panels was studied as it was processed, recording the batch
size, the start time, the finish time, and some intermediate times. The model was
then run using this exact same scenario, and the output was compared with the
har01307_ch07_193-210.indd 208 20/12/10 1:50 PM

output of the real process. The initial model predicted 25 shifts to complete a
specific sequence of panels, while the process actually took 35 shifts. This led
to further process investigations of interruptions or other processing delays that
weren’t being accounted for in the model. It was discovered that feeder replace-
ment times were underestimated in the model. A discrepancy between the model
and the actual system was also discovered in the utilization of the pick-and-place
machines. After tracking down the causes of these discrepancies and making ap-
propriate adjustments to the model, the simulation results were closer to the real
process. The challenge at this stage was not to yield to the temptation of making
arbitrary changes to the model just to get the desired results. Then the model would
lose its integrity and become nothing more than a sham for the real system.
The final step was to conduct sensitivity analysis to determine how model per-
formance was affected in response to changes in model assumptions. By changing
the input in such a way that the impact was somewhat predictable, the change in
simulation results could be compared with intuitive guesses. Any bizarre results
such as a decrease in work in process when increasing the arrival rate would
raise an immediate flag. Input parameters were systematically changed based on
a knowledge of both the process behavior and the model operation. Knowing
just how to stress the model in the right places to test its robustness was an art in
itself. After everyone was satisfied that the model accurately reflected the actual
operation and that it seemed to respond as would be expected to specific changes
in input, the model was ready for experimental use.
7.5 Summary
For a simulation model to be of greatest value, it must be an accurate representa-
tion of the system being modeled. Verification and validation are two activities
that should be performed with simulation models to establish their credibility.
Model verification is basically the process of debugging the model to ensure
that it accurately represents the conceptual model and that the simulation runs cor-
rectly. Verification involves mainly the modeler without the need for much input
from the customer. Verification is not an exact science, although several proven
techniques can be used.
Validating a model begins at the data-gathering stage and may not end until
the system is finally implemented and the actual system can be compared to the
model. Validation involves the customer and other stakeholders who are in a posi-
tion to provide informed feedback on whether the model accurately reflects the
real system. Absolute validation is philosophically impossible, although a high
degree of face or functional validity can be established.
While formal methods may not always be used to validate a model, at least
some time devoted to careful review should be given. The secret to validation is
not so much the technique as it is the attitude. One should be as skeptical as can
be tolerated—challenging every input and questioning every output. In the end,
the decision maker should be confident in the simulation results, not because he or
har01307_ch07_193-210.indd 209 20/12/10 1:50 PM

she trusts the software or the modeler, but because he or she trusts the input data
and knows how the model was built.

1. What is the difference between model verification and model
validation?
2. Why are verification and validation so important in a simulation study?
3. Why are these steps often ignored or given only little attention?
4. How can one prevent syntax and semantic errors from creeping into a
simulation model?
5. How does one establish a standard for verifying a model?
6. What techniques are used to verify a model?
7. Why is it philosophically impossible to prove model validity?
8. What is functional validity?
9. What techniques are used to validate a model?
10. Why is it useful to build an as-is model if a change to an existing
system is being evaluated?
11. What specific things were done in the St. John Hospital simulation
study to validate the model?
12. What specific steps were taken in the HP simulation study to validate
the model?
References
Balci, Osman. “Verification, Validation and Accreditation of Simulation Models.” In Pro-
ceedings of the 1997 Winter Simulation Conference, ed. S. Andradottir, K. J. Healy,
D. H. Withers, and B. L. Nelson, 1997, pp. 135–41.
Banks, Jerry. “Simulation Evolution.” IIE Solutions, November 1998, pp. 26–29.
Baxter, Lori K., and Johnson, Eric. “Don’t Implement before You Validate.” Industrial
Engineering, February 1993, pp. 60–62.
Hoover, Stewart, and Ronald Perry. Simulation: A Problem Solving Approach. Reading,
MA: Addison-Wesley, 1990.
Neelamkavil, F. Computer Simulation and Modeling. New York: John Wiley & Sons, 1987.
O’Conner, Kathleen. “The Use of a Computer Simulation Model to Plan an Obstetrical
Renovation Project.” The Fifth Annual PROMODEL Users Conference, Park City,
UT, 1994.
Sargent, Robert G. “Verifying and Validating Simulation Models.” In Proceedings of the
1998 Winter Simulation Conference, ed. D. J. Medeiros, E. F. Watson, J. S. Carson,
and M. S. Manivannan, 1998, pp. 121–30.
har01307_ch07_193-210.indd 210 20/12/10 1:50 PM

C H A P T E R
8 SIMULATION OUTPUT
ANALYSIS
“Nothing is more terrible than activity without insight.”

—Thomas Carlyle
8.1 Introduction
In analyzing the output from a simulation model, there is room for both rough
analysis using judgmental procedures as well as statistically based procedures for
more detailed analysis. The appropriateness of using judgmental procedures or
statistically based procedures to analyze a simulation’s output depends largely on
the nature of the problem, the importance of the decision, and the validity of the
input data. If you are doing a go/no-go type of analysis in which you are trying to
find out whether a system is capable of meeting a minimum performance level,
then a simple judgmental approach may be adequate. Finding out whether a single
machine or a single service agent is adequate to handle a given workload may be
easily determined by a few runs unless it looks like a close call. Even if it is a
close call, if the decision is not that important (perhaps there is a backup worker
who can easily fill in during peak periods), then more detailed analysis may not
be needed. In cases where the model relies heavily on assumptions, it is of little
value to be extremely precise in estimating model performance. It does little good
to get six decimal places of precision for an output response if the input data war-
rant only precision to the nearest tens. Suppose, for example, that the arrival rate
of customers to a bank is roughly estimated to be 30 plus or minus 10 per hour. In
this situation, it is probably meaningless to try to obtain a precise estimate of teller
utilization in the facility. The precision of the input data simply doesn’t warrant
any more than a rough estimate for the output.
These examples of rough estimates are in no way intended to minimize the
importance of conducting statistically responsible experiments, but rather to em-
phasize the fact that the average analyst or manager can gainfully use simulation
211
har01307_ch08_211-242.indd 211 20/12/10 1:52 PM

without having to be a professional statistician. Anyone who understands the

cause-and-effect relationships of a system will be able to draw meaningful con-
clusions about it that will help in design and management decisions. Where pre-
cision is important and warranted, however, a statistical background or at least
consulting with a statistical advisor will help avoid drawing invalid inferences
from simulation.
In this chapter, we present basic statistical issues related to the analysis and
interpretation of simulation output. The intent is not to make you an “expert” stat-
istician, but to give you sound guidelines to follow when dealing with the output
generated by simulation models. For more comprehensive coverage of statistical
methods, see Hines et al. (2003).
8.2 Statistical Analysis of Simulation Output

Because random input variables (such as activity times) are used to drive the
model, the output measures of the simulation (throughput rate, average waiting
times) are also going to be random. This means that performance metrics based
on a model’s output are only estimates and not precise measures. Suppose, for
example, we are interested in determining the throughput for a model of a manu-
facturing system (like the number of widgets produced per hour). Suppose further
that the model has a number of random input variables such as machining times,
time between machine failures, and repair times. Running the simulation once
provides a single random sample of the model’s fluctuating throughput that may
or may not be representative of the expected or average model throughput. This
would be like rolling a die, getting a 1, and concluding that the expected average
value of the die is 1 instead of the true average of 3.5.
Statistical analysis of simulation output is based on inferential or descrip-
tive statistics. In descriptive statistics, we deal with a population, samples from
the population, and a sample size. The population is the entire set of items or
outcomes (such as all possible values that can be obtained when rolling a die). A
sample would be one unbiased observation or instance of the population (a single
roll of a die). The sample size is the number of samples we are working with (a
particular number of rolls of a die). The collected samples are independent (each
roll of the die is unaffected by previous rolls of the die). The idea behind inferen-
tial statistics is to gather a large enough sample size to draw valid inferences about
the population (such as the average value of the die) while keeping the sample-
gathering time and cost at a minimum. In some instances, collecting samples can
be very time-consuming and costly.
The same statistical concepts apply to conducting an experiment with a simu-
lation model of a system that apply to sampling from a population such as rolls of
a die or a group of people:
• Samples are generated by running the experiment.
• The population size is often very large or infinite (there are an infinite
number of possible outcomes from the experiment).
har01307_ch08_211-242.indd 212 20/12/10 1:52 PM

Chapter 8 Simulation Output Analysis 213
• A random variable is used to express each possible outcome of an

experiment as a continuous or discrete number.
• Experimental samples, called replications, are independent.
Multiple independent runs or replications of a simulation provide multiple ob-
servations that can be used to estimate the expected value of the model’s output
response (like rolling the die multiple times before estimating an expected or mean
value). With random simulation output results, we can’t determine for certain what
the true expected output will be from the model unless in some cases we run an
infinite number of replications, which is not possible or even desirable.
8.2.1 Simulation Replications

In conducting a simulation experiment, one run of the simulation constitutes one
replication of the experiment. The outcome of a replication represents a single
sample. To obtain a sample size n, we need to run n independent replications of
the experiment. A representative sample size will be a good indicator of what can
be expected to happen for any subsequent replication. Not every replication will
produce exactly the same behavior due to random variability. To understand how
independent replications are produced in simulation, it is essential to understand
the concepts of random number generation and initial seed values. These subjects
were elaborated on in Chapter 3 but will be reviewed here.
Random Number Generation

At the heart of a stochastic simulation is the random number generator. The random
number generator produces a cycling sequence of pseudo-random numbers that are
used to generate observations from various distributions (normal, exponential, or oth-
ers) that drive the stochastic processes represented in the simulation model such as
activity times, machine failure times, and travel times. As pseudo-random numbers,
the random number sequence, or stream, can be repeated if desired. After the random
number generator sequences through all random numbers in its cycle, the cycle starts
over again through the same sequence (see Figure 8.1). The length of the cycle before
it repeats is called the cycle period, which is usually extremely long (the cycle length
for the random number generator used in ProModel is over 2.1 billion).
Initial Seed Values

The random number generator requires an initial seed value before it can begin
producing random numbers. Each random number produced by the generator is
fed back to the generator to produce the next random number, and so on. Where
the sequence begins in the cycle depends on the initial seed value given to the
generator. In Figure 8.1, the seed value of Z0 17 marks the beginning of this
sequence of numbers.
You might wonder how replications are produced if the exact same sequence
of random numbers drives the simulation. Obviously, identical random number se-
quences produce identical results. This may prove useful for repeating an exact
experiment but does not give us an independent replication. On the other hand,
har01307_ch08_211-242.indd 213 20/12/10 1:52 PM

FIGURE 8.1 .52

Example of a cycling
pseudo-random
.80
number stream .31
produced by a random
number generator
with a very short cycle
Random number
length. .07 .95 Seed Z0 17
generator
.60
.25
.66
if a different seed value is appropriately selected to initialize the random number

generator, the simulation will produce different results because it will be driven by
a different segment of numbers from the random number stream. This is how the
simulation experiment is replicated to collect statistically independent observations
of the simulation model’s output response. Recall that the random number genera-
tor in ProModel can produce over 2.1 billion different values before it cycles.
To replicate a simulation experiment, then, the simulation model is initialized
to its starting conditions, all statistical variables for the output measures are reset,
and a new seed value is appropriately selected to start the random number genera-
tor. Each time an appropriate seed value is used to start the random number gen-
erator, the simulation produces a unique output response. Repeating the process
with several appropriate seed values produces a set of statistically independent
observations of a model’s output response. With most simulation software, users
need only specify how many times they wish to replicate the experiment; the soft-
ware handles the details of initializing variables and ensuring that each simulation
run is driven by nonoverlapping segments of the random number cycle.
8.2.2 Performance Estimation

Assuming we have an adequate collection of independent observations, we can
apply statistical methods to estimate the expected or mean value of the model’s
output response. There are two types of estimates: (1) point estimates (mean
and standard deviation, for example) and (2) interval estimates (also called con-
fidence intervals). This is not to say that mean and standard deviation are the
only parameter estimates in which we may be interested. For example, we may
compute point estimates for parameters such as the proportion of time a queue
had more than 10 entities waiting in it or the maximum number of entities to
simultaneously wait in a queue location. Most simulation software automatically
calculates these and other estimates for you; however, it is useful to know the
basic procedures used to better understand their meaning.
har01307_ch08_211-242.indd 214 20/12/10 1:52 PM

Point Estimates
A point estimate is a single value estimate of a parameter of interest. Point esti-
mates are calculated for the mean and standard deviation of the population. To
estimate the mean of the population (denoted as ␮), we simply calculate the aver-
age of the sample values (denoted as x̄):
i 1 xi
n
x̄ _ n
where n is the sample size (number of observations) and xi is the value of ith
observation. The sample mean x̄ estimates the population mean ␮.
The standard deviation for the population (denoted as ␴), which is a measure
of the spread of data values in the population, is similarly estimated by calculating
a standard deviation of the sample of values (denoted as s):
___

i 1 [xi x̄]
n 2
s __
n1
The sample variance s , used to estimate the variance of the population ␴2, is ob-
2
tained by squaring the sample standard deviation.

Let’s suppose, for example, that we are interested in determining the mean or
average number of customers getting a haircut at Buddy’s Style Shop on Saturday
morning. Buddy opens his barbershop at 8:00 A.M. and closes at noon on Saturday.
In order to determine the exact value for the true average number of customers getting
a haircut on Saturday morning (␮), we would have to compute the average based on
the number of haircuts given on all Saturday mornings that Buddy’s Style Shop has
been and will be open (that is, the complete population of observations). Not want-
ing to work that hard, we decide to get an estimate of the true mean ␮ by spending
the next 12 Saturday mornings watching TV at Buddy’s and recording the number of
customers that get a haircut between 8:00 A.M. and 12:00 noon (Table 8.1).
TABLE 8.1 Number of Haircuts Given on 12 Saturdays
Replication (i) Number of Haircuts (xi)
1 21
2 16
3 8
4 11
5 17
6 16
7 6
8 14
9 15
10 16
11 14
12 10
Sample mean x̄ 13.67

Sample standard deviation s 4.21
har01307_ch08_211-242.indd 215 20/12/10 1:52 PM

We have replicated the experiment 12 times and have 12 sample observations

of the number of customers getting haircuts on Saturday morning (x1, x2, . . . , x12).
Note that there are many different values for the random variable (number of
haircuts) in the sample of 12 observations. So if only one replication had been
conducted and the single observation was used to estimate the true but unknown
mean number of haircuts given, the estimate could potentially be way off. Using
the observations in Table 8.1, we calculate the sample mean as follows:
12
x̄ _i 1 xi 21 16 8 · · · 10 13.67 haircuts
____
12 12
The sample mean of 13.67 haircuts estimates the unknown true mean value ␮
for the number of haircuts given on Saturday at the barbershop. Taking additional
samples generally gives a more accurate estimate of the unknown ␮. However,
the experiment would need to be replicated on each Saturday that Buddy’s has
been and will be open to determine the true mean or expected number of haircuts
given.
In addition to estimating the population mean, we can also calculate an es-
timate for the population standard deviation based on the sample size of 12 ob-
servations and a sample mean of 13.67 haircuts. This is done by calculating the
standard deviation s of the observations in Table 8.1 as follows:
___
12 [x 13.67]2 4.21 haircuts
s
i 1
___ i
12 1
The sample standard deviation s provides only an estimate of the true but
unknown population standard deviation ␴. x̄ and s are single values; thus they are
referred to as point estimators of the population parameters ␮ and ␴. Also, note
that x̄ and s are random variables. As such, they will have different values if based
on another set of 12 independent observations from the barbershop.
Interval Estimates
A point estimate, by itself, gives little information about how accurately it esti-
mates the true value of the unknown parameter. Interval estimates constructed
using x̄ and s, on the other hand, provide information about how far off the point
estimate x̄ might be from the true mean ␮. The method used to determine this is
referred to as confidence interval estimation.
A confidence interval is a range within which we can have a certain level of
confidence that the true mean falls. The interval is symmetric about x̄, and the
distance that each endpoint is from x̄ is called the half-width (hw). A confidence
interval, then, is expressed as the probability P that the unknown true mean ␮ lies
within the interval x̄ hw. The probability P is called the confidence level.
If the sample observations used to compute x̄ and s are independent and nor-
mally distributed, the following equation can be used to calculate the half-width
of a confidence interval for a given level of confidence:
(tn1, ␣Ⲑ2)s
hw __ _
n
har01307_ch08_211-242.indd 216 20/12/10 1:52 PM

where tn1,␣2 is a factor that can be obtained from the Student’s t table in Appen-
dix B. The values are identified in the Student’s t table according to the value of
␣2 and the degrees of freedom (n 1). The term ␣ is the complement of P.
That is, ␣ 1 P and is referred to as the significance level. The significance
level may be thought of as the “risk level” or probability that ␮ will fall out-
side the confidence interval. Therefore, the probability that ␮ will fall within
the confidence interval is 1 ␣. Thus confidence intervals are often stated as
P( x̄ hw ␮ x̄ hw) 1 ␣ and are read as the probability that the true
but unknown mean ␮ falls between the interval ( x̄ hw) to ( x̄ hw) is equal to
1 ␣. The confidence interval is traditionally referred to as a 100(1 ␣) percent
confidence interval.
Assuming the data from the barbershop example are independent and nor-
mally distributed (this assumption is discussed in section 8.3), a 95 percent confi-
dence interval is constructed as follows:
Given: P confidence level 0.95
␣ significance level 1 P 1 0.95 0.05
n sample size 12
x̄ 13.67 haircuts
s 4.21 haircuts
From the Student’s t table in Appendix B, we find tn1,␣2 t11,0.025 2.201. The
half-width is computed as follows:
(t11,0.025 )s _(2.201)4.21
hw _ _ _ 2.67 haircuts
n 12
The lower and upper limits of the 95 percent confidence interval are calculated
as follows:
Lower limit x̄ hw 13.67 2.67 11.00 haircuts
Upper limit x̄ hw 13.67 2.67 16.34 haircuts
It can now be asserted with 95 percent confidence that the true but unknown mean
falls between 11.00 haircuts and 16.34 haircuts (11.00 ␮haircuts 16.34). A risk
level of 0.05(␣ 0.05) means that if we went through this process for estimating
the number of haircuts given on a Saturday morning 100 times and computed 100
confidence intervals, we would expect 5 of the confidence intervals (5 percent
of 100) to exclude the true but unknown mean number of haircuts given on a
Saturday morning.
The width of the interval indicates the accuracy of the point estimate. It is
desirable to have a small interval with high confidence (usually 90 percent or
greater). The width of the confidence interval is affected by the variability in the
output of the system and the number of observations collected (sample size). It
can be seen from the half-width equation that, for a given confidence level, the
half-width will shrink if (1) the sample size n is increased or (2) the variability
in the output of the system (standard deviation s) is reduced. Given that we have
har01307_ch08_211-242.indd 217 20/12/10 1:52 PM

little control over the variability of the system, we resign ourselves to increasing
the sample size (run more replications) to increase the accuracy of our estimates.
Actually, when dealing with the output from a simulation model, techniques
can sometimes reduce the variability of the output from the model without chang-
ing the expected value of the output. These are called variance reduction tech-
niques and are covered in Chapter 9.
8.2.3 Number of Replications (Sample Size)

Often we are interested in knowing the sample size or number of replications
needed to establish a particular confidence interval for a specified amount of
absolute error (denoted by e) between the point estimate of the mean x̄ and the
unknown true mean ␮. In our barbershop example, suppose we want to estimate the
number of replications n needed to be able to assert with 95 percent (1 ␣ .95)
confidence that the sample mean x̄ we compute is off by at most 2.00 haircuts
(e 2.00) from the true but unknown mean ␮. In other words, we would like to
reduce the half-width of 2.67 haircuts for the confidence interval we previously
computed to 2.00 haircuts.
Note that ␣ represents the probability that the difference between x̄ and ␮
will exceed a specified error amount e. In fact, e is equal to the hw of a confidence
interval (x̄ e). Setting e hw and solving for n would give an equation that
could be used to estimate the number of replications necessary to meet a specified
absolute error amount e and significance level ␣:
hw e
(tn1,␣2)s
_ _ e
n
n _ (
(tn1,␣2)s
e )
2
However, this cannot be completed because n appears on each side of the equa-
tion. So, to estimate the number of replications needed, we replace tn1,␣2 with
Z␣2, which depends only on ␣. The Z␣2 is from the standard normal distribution,
and its value can be found in the last row of the Student’s t table in Appendix B.
Note that Z␣2 t,␣2, where denotes infinity. The revised equation is
(
(Z␣2)s
n _ e ) 2
where n is a rough approximation for the number of replications that will provide
an adequate sample size for meeting the desired absolute error amount e and sig-
nificance level ␣.
Before this equation can be used, an initial sample of observations must be
collected to compute the sample standard deviation s of the output from the system.
Actually, when using the Z␣2 value, an assumption is made that the true standard
deviation for the complete population of observations ␴ is known. Of course, ␴ is
not known. However, it is reasonable to approximate ␴ with the sample standard
deviation s to get a rough approximation for the required number of replications.
har01307_ch08_211-242.indd 218 20/12/10 1:52 PM

Going back to the barbershop example, we want to estimate the number of

replications n needed to be able to assert with 95 percent (1 ␣ .95) confidence
that the sample mean x̄ we compute is off by at most 2.00 haircuts (e 2.00) from
the true but unknown mean ␮. Based on the initial sample of 12 observations, we
know that s 4.21 haircuts. We compute n as follows:
e 2.00
s 4.21 haircuts
From the last row of the Student’s t table in Appendix B, we find Z␣2 Z0.025
t,0.025 1.96. Using the previous equation for n we obtain
n _
(Z0.025)s 2
e
(1.96)4.21 2
_
2.00
17.02 observations
A fractional number of replications cannot be conducted, so the result is

typically rounded up. Thus n 18 observations. Given that we already have
12 observations, the experiment needs to be replicated an additional 6 (n n
18 12 6) times to collect the necessary observations. After the additional
6 replications are conducted, the 6 observations are combined with the original
12 observations (Table 8.2), and a new x̄, s, and confidence interval are computed
using all 18 observations.
TABLE 8.2 Number of Haircuts Given on 18 Saturdays
Replication (i) Number of Haircuts xi
1 21
2 16
3 8
4 11
5 17
6 16
7 6
8 14
9 15
10 16
11 14
12 10
13 7
14 9
15 18
16 13
17 16
18 8
Sample mean x̄ 13.06

Sample standard deviation s 4.28
har01307_ch08_211-242.indd 219 20/12/10 1:52 PM


n sample size 18
x̄ 13.06 haircuts
s 4.28 haircuts
From the Student’s t table in Appendix B, we find tn1, ␣2 t17,0.025 2.11. The
half-width is computed as follows:
(t17,0.025)s (2.11)_
4.28
hw _ _ _ 2.13 haircuts
n 18
The lower and upper limits for the new 95 percent confidence interval are calcu-
lated as follows:
Lower limit x̄ hw 13.06 2.13 10.93 haircuts
Upper limit x̄ hw 13.06 2.13 15.19 haircuts
It can now be asserted with 95 percent confidence that the true but unknown mean
falls between 10.93 haircuts and 15.19 haircuts (10.93 ␮haircuts 15.19).
Note that with the additional observations, the half-width of the confidence
interval has indeed decreased. However, the half-width is larger than the absolute
error value planned for (e 2.0). This is just luck, or bad luck, because the new
half-width could just as easily have been smaller. Why? First, 18 replications was
only a rough estimate of the number of observations needed. Second, the number
of haircuts given on a Saturday morning at the barbershop is a random variable.
Each collection of observations of the random variable will likely differ from pre-
vious collections. Therefore, the values computed for x̄, s, and hw will also differ.
This is the nature of statistics and why we deal only in estimates.
We have been expressing our target amount of error e in our point estimate x̄
as an absolute value (hw e). In the barbershop example, we selected an abso-
lute value of e 2.00 haircuts as our target value. However, it is sometimes more
convenient to work in terms of a relative error (re) value (hw reⱍ␮ⱍ). This allows
us to talk about the percentage error in our point estimate in place of the absolute
error. Percentage error is the relative error multiplied by 100 (that is, 100re per-
cent). To approximate the number of replications needed to obtain a point estimate
x̄ with a certain percentage error, we need only change the denominator of the n
equation used earlier. The relative error version of the equation becomes

2
(z␣2)s
n _ _____
re
( (1re) )x̄
where re denotes the relative error. The re(1 re) part of the denominator is an
adjustment needed to realize the desired re value because we use x̄ to estimate ␮
(see Chapter 9 of Law 2007 for details). The appeal of this approach is that we
can select a desired percentage error without prior knowledge of the magnitude
of the value of ␮.
har01307_ch08_211-242.indd 220 20/12/10 1:52 PM

As an example, say that after recording the number of haircuts given at the
barbershop on 12 Saturdays (n 12 replications of the experiment), we wish to
determine the approximate number of replications needed to estimate the mean
number of haircuts given per day with an error percentage of 17.14 percent and a
confidence level of 95 percent. We apply our equation using the sample mean and
sample standard deviation from Table 8.1.
Z␣2 Z0.025 1.96 from Appendix B
re 0.1714
x̄ 13.67
s 4.21

n _ re
_____ (
(z0.025)s 2
( (1re) )x̄
__
(1.96)4.21 2
(0.1714
)
_______ 13.67
(10.1714)

17.02 observations
Thus n 18 observations. This is the same result computed earlier on page 219
for an absolute error of e 2.00 haircuts. This occurred because we purposely
selected re 0.1714 to produce a value of 2.00 for the equation’s denominator in
order to demonstrate the equivalency of the different methods for approximating
the number of replications needed to achieve a desired level of precision in the
point estimate x̄. The SimRunner software uses the relative error methodology to
provide estimates for the number of replications needed to satisfy the specified
level of precision for a given significance level ␣.
In general, the accuracy of the estimates improves as the number of replica-
tions increases. However, after a point, only modest improvements in accuracy
(reductions in the half-width of the confidence interval) are made through con-
ducting additional replications of the experiment. Therefore, it is sometimes nec-
essary to compromise on a desired level of accuracy because of the time required
to run a large number of replications of the model.
The ProModel simulation software automatically computes confidence inter-
vals. Therefore, there is really no need to estimate the sample size required for a
desired half-width using the method given in this section. Instead, the experiment
could be replicated, say, 10 times, and the half-width of the resulting confidence
interval checked. If the desired half-width is not achieved, additional replications
of the experiment are made until it is. The only real advantage of estimating the
number of replications in advance is that it may save time over the trial-and-error
approach of repeatedly checking the half-width and running additional replica-
tions until the desired confidence interval is achieved.
8.2.4 Real-World Experiments versus

Simulation Experiments
You may have noticed that we have been working with the observations as though
they came from the real barbershop and not from a simulation model of the shop.
har01307_ch08_211-242.indd 221 20/12/10 1:52 PM

Actually, the observations were generated from a simulation of the shop by mak-
ing independent runs (replications) of the model using the method described in
section 8.2.1. Note, however, that had the observations been collected from the
real barbershop, we would have processed them in the same manner. Experimen-
tal results are generally handled in the same manner, regardless of their origin, if
the observations are unbiased, independent, and normally distributed.
Note that very different results could be obtained from the real barbershop
if the simulation model does not accurately mimic the behavior of the real shop.
Statistical statements, no matter how precise, derived from the output of a simula-
tion model provide estimates for the true but unknown population parameters of
the model and not of the real system. Whether or not the model behavior actually
corresponds to real-world behavior is a matter of model validity, which was dis-
cussed in Chapter 7.
8.3 Statistical Issues with Simulation Output

A single simulation run (replication) represents only a single sample of possible
outcomes (values of a random variable) from the simulation. Averaging the ob-
servations from a large number of runs comes closer to the true expected per-
formance of the model, but it is still only a statistical estimate. Constructing a
confidence interval is a good idea because it provides a measure of how far off the
estimate is from the model’s true expected performance. However, the following
three statistical assumptions must be met regarding the sample of observations
used to construct the confidence interval:
• Observations are independent so that no correlation exists between
consecutive observations.
• Observations are identically distributed throughout the entire duration of
the process (that is, they are time invariant).
• Observations are normally distributed.
Table 8.3 presents a framework for thinking about the output produced from
a series of simulation runs (independent replications) that will aid in the discus-
sion of these three important statistical assumptions. Suppose a system has been
TABLE 8.3 Simulation Output from a Series of Replications
Replication (i) Within Run Observations yij Average xi
1 y11, y12, y13, y14, . . . , y1m x1

2 y21, y22, y23, y24, . . . , y2m x2
3 y31, y32, y33, y34, . . . , y3m x3
4 y41, y42, y43, y44, . . . , y4m x4
n yn1, yn2, yn3, yn4, . . . , ynm xn
har01307_ch08_211-242.indd 222 20/12/10 1:52 PM

simulated to estimate the mean time that entities wait in queues in that system.
The experiment is replicated several times to collect n independent sample obser-
vations, as was done in the barbershop example. As an entity exits the system, the
time that the entity spent waiting in queues is recorded. The waiting time is de-
noted by yij in Table 8.3, where the subscript i denotes the replication from which
the observation came and the subscript j denotes the value of the counter used
to count entities as they exit the system. For example, y32 is the waiting time for
the second entity that was processed through the system in the third replication.
These values are recorded during a particular run (replication) of the model and
are listed under the column labeled “Within Run Observations” in Table 8.3.
Note that the within run observations for a particular replication, say the
ith replication, are not usually independent because of the correlation between
consecutive observations. For example, when the simulation starts and the first
entity begins processing, there is no waiting time in the queue. Obviously, the
more congested the system becomes at various times throughout the simulation,
the longer entities will wait in queues. If the waiting time observed for one entity
is long, it is highly likely that the waiting time for the next entity observed is
going to be long and vice versa. Observations exhibiting this correlation between
consecutive observations are said to be autocorrelated. Furthermore, the within
run observations for a particular replication are often nonstationary in that they
do not follow an identical distribution throughout the simulation run. Therefore,
they cannot be directly used as observations for statistical methods that require
independent and identically distributed observations such as those used in this
chapter.
At this point, it seems that we are a long way from getting a usable set of
observations. However, let’s focus our attention on the last column in Table 8.3,
labeled “Average,” which contains the x1 through xn values. xi denotes the average
waiting time of the entities processed during the ith simulation run (replication)
and is computed as follows:
j 1 yij
m
xi _ m
where m is the number of entities processed through the system and yij is the time
that the jth entity processed through the system waited in queues during the ith
replication. Although not as formally stated, you used this equation in the last
row of the ATM spreadsheet simulation of Chapter 3 to compute the average
waiting time of the m 25 customers (Table 3.2). An xi value for a particular
replication represents only one possible value for the mean time an entity waits
in queues, and it would be risky to make a statement about the true waiting time
from a single observation. However, Table 8.3 contains an xi value for each of
the n independent replications of the simulation. These xi values are statistically
independent if different seed values are used to initialize the random number
generator for each replication (as discussed in section 8.2.1). The xi values are
often identically distributed as well. Therefore, the sample of xi values can be
used for statistical methods requiring independent and identically distributed
har01307_ch08_211-242.indd 223 20/12/10 1:52 PM

observations. Thus we can use the xi values to estimate the true but unknown
average time entities wait in queues by computing the sample mean as follows:
i 1 xi
n
x̄ _ n
You used this equation in Table 3.3 to average the results from the two replica-
tions of the ATM spreadsheet simulation of Chapter 3.
The standard deviation of the average time entities wait in queues is esti-
mated by the sample standard deviation and is computed by
___

i 1 [xi x̄]
n 2
s __
n1
The third assumption for many standard statistical methods is that the sample
observations (x1 through xn) are normally distributed. For example, the normality
assumption is implied at the point for which a tn1, 2 value from the Student’s
t distribution is used in computing a confidence interval’s half-width. The cen-
tral limit theorem of statistics provides a basis for making this assumption. If a
variable is defined as the sum of several independent and identically distributed
random values, the central limit theorem states that the variable representing the
sum tends to be normally distributed. This is helpful because in computing xi for
a particular replication, the (yij) waiting times for the m entities processed in the
ith replication are summed together. However, the yij observations are autocorre-
lated and, therefore, not independent. But there are also central limit theorems for
certain types of correlated data that suggest that xi will be approximately normally
distributed if the sample size used to compute it (the number of entities processed
through the system, m from Table 8.3) is not too small (Law 2007). Therefore, we
can compute a confidence interval using
(tn1, 2)s
x̄ _ _
n
where tn1,␣2 is a value from the Student’s t table in Appendix B for an ␣ level of
significance.
ProModel automatically computes point estimates and confidence intervals
using this method. Figure 8.2 presents the output produced from running the
ProModel version of the ATM simulation of Lab Chapter 1 for five replications.
The column under the Locations section of the output report labeled “Average
Minutes per Entry” displays the average amount of time that customers waited in
the queue during each of the five replications. These values correspond to the xi
values in Table 8.3. Note that in addition to the sample mean and standard devia-
tion, ProModel also provides a 95 percent confidence interval.
Sometimes the output measure being evaluated is not based on the mean, or
sum, of a collection of random values. For example, the output measure may be
the maximum number of entities that simultaneously wait in a queue. (See the
“Maximum Contents” column in Figure 8.2.) In such cases, the output measure
may not be normally distributed, and because it is not a sum, the central limit
har01307_ch08_211-242.indd 224 20/12/10 1:52 PM

FIGURE 8.2
Replication technique used on ATM simulation of Lab Chapter 1.
-------------------------------------------------------------------
General Report
Output from C:\Bowden\3rd Edition\ATM System Ch8.MOD [ATM System]
Date: Nov/26/2010 Time: 04:24:27 PM
-------------------------------------------------------------------
Scenario : Normal Run
Replication : All
Period : Final Report (1000 hr to 1500 hr Elapsed: 500 hr)
Warmup Time : 1000
Simulation Time : 1500 hr
-------------------------------------------------------------------
LOCATIONS
Average
Location Scheduled Total Minutes Average Maximum Current
Name Hours Capacity Entries Per Entry Contents Contents Contents % Util
--------- ----- -------- ------- --------- -------- -------- -------- ------
ATM Queue 500 999999 9903 9.457 3.122 26 18 0.0 (Rep 1)
ATM Queue 500 999999 9866 8.912 2.930 24 0 0.0 (Rep 2)
ATM Queue 500 999999 9977 11.195 3.723 39 4 0.0 (Rep 3)
ATM Queue 500 999999 10006 8.697 2.900 26 3 0.0 (Rep 4)
ATM Queue 500 999999 10187 10.841 3.681 32 0 0.0 (Rep 5)
ATM Queue 500 999999 9987.8 9.820 3.271 29.4 5 0.0 (Average)
ATM Queue 0 0 124.654 1.134 0.402 6.148 7.483 0.0 (Std. Dev.)
ATM Queue 500 999999 9833.05 8.412 2.772 21.767 -4.290 0.0 (95% C.I. Low)
ATM Queue 500 999999 10142.6 11.229 3.771 37.032 14.290 0.0 (95% C.I. High)
theorem cannot be used to imply that it is normally distributed. Some refuge may
be taken in knowing that the method given in the chapter for constructing a confi-
dence interval does not completely break down when the normality assumption is
violated. So if the normality assumption is questionable, use the confidence inter-
val with the knowledge that it is only approximate. However, you must not violate
the independence assumption. This is not an obstacle because it is easy to get a set
of observations that are statistically independent by replicating the simulation.
8.4 Terminating and Nonterminating Simulations

Determining experimental procedures, such as how many replications and how
to gather statistics, depends on whether we should conduct a terminating or
nonterminating simulation. A terminating simulation is conducted when we are
interested in the behavior of the system over a particular period—for example,
the movement of cars on a roadway system during rush hour. A nonterminating
simulation is conducted when we are interested in the steady-state behavior of a
system—for example, the long-term average behavior of a production system that
steadily produces spark plugs for automobile engines. The purpose of this section
is to give guidelines for deciding which is most appropriate for a given analysis,
terminating simulations or nonterminating simulations.
har01307_ch08_211-242.indd 225 20/12/10 1:52 PM

8.4.1 Terminating Simulations

A terminating simulation is one in which the simulation starts at a defined state
or time and ends when it reaches some other defined state or time. An initial
state might be the number of parts in the system at the beginning of a workday.
A terminating state or event might occur when a particular number of parts have
been completed. Consider, for example, an aerospace manufacturer that receives
an order to manufacture 200 airplanes of a particular model. The company might
be interested in knowing how long it will take to produce the 200 airplanes. The
simulation run starts with the system empty and is terminated when the 200th
plane is completed. A point in time that would bring a terminating simulation to
an end might be the closing of shop at the end of a business day or the completion
of a weekly or monthly production period. It may be known, for example, that
a production schedule for a particular item changes weekly. At the end of each
40-hour cycle, the system is “emptied” and a new production cycle begins. In
this situation, a terminating simulation would be run in which the simulation run
length would be 40 hours.
A common type of terminating system is one that starts empty, runs for a
while, and then empties out again before shutting down. Examples of this type of
system are job shops, department stores, restaurants, banks, hair salons, and post
offices. All these systems open up for business in the morning with no customers
in the system and with all resources initially idle. At the end of the day the system
gradually closes down. Usually, at a predetermined time, the customer arrival
process is terminated. (The front door is locked to prevent new arrivals.) Service
for all existing customers is completed, and then the system finally shuts down
for the day. Most service systems tend to operate in this fashion and are therefore
modeled as terminating systems.
Terminating simulations are not intended to measure the steady-state behav-
ior of a system, although there may be periods during a terminating simulation
during which the system is in a steady state. In a terminating simulation, average
measures of performance based on the entire duration of the simulation are of
little meaning. Because a terminating simulation always contains transient peri-
ods that are part of the analysis, utilization figures, for example, have the most
meaning if reported for successive time intervals during the simulation.
8.4.2 Nonterminating Simulations

A nonterminating or steady-state simulation is one in which the steady-state
(long-term average) behavior of the system is being analyzed. A nonterminating
simulation does not mean that the simulation never ends, nor does it mean that the
system being simulated has no eventual termination. It only means that the simu-
lation could theoretically go on indefinitely with no statistical change in behavior.
This does not mean that a steady-state condition is one in which the observations
are all the same, or even one for which the variation in observations is any less
than during a transient condition. It means only that all observations throughout
har01307_ch08_211-242.indd 226 20/12/10 1:52 PM

the steady-state period will have approximately the same distribution. For such
systems, a suitable length of time to run the model in order to collect statistics on
the steady-state behavior of the system must be determined.
Nonterminating systems begin with a warm-up (or transient) state and gradu-
ally move to a steady state. Once the initial transient phase has diminished to the
point where the impact of the initial condition on the system’s response is neg-
ligible, we consider it to have reached steady state. Once in a steady state, if the
operating rules change or an input variable like the rate at which entities arrive
changes, the system reverts again to a transient state until the system has had time
to start reflecting the long-term behavior of the new operating circumstances.
An example of a nonterminating simulation is a model of a manufacturing
operation in which automotive oil filters are produced continually. The operation
runs two shifts with an hour break during each shift in which everything momen-
tarily stops. Break and third-shift times are excluded from the model because
work always continues exactly as it left off before the break or end of shift. The
length of the simulation is determined by how long it takes to get a representative
steady-state reading of the model behavior.
Nonterminating simulations can, and often do, change operating character-
istics after a period of time, but usually only after enough time has elapsed to es-
tablish a steady-state condition. Take, for example, a production system that runs
10,000 units per week for 5 weeks and then increases to 15,000 units per week
for the next 10 weeks. The system would have two different steady-state periods.
Thus, you would want to segment observations from the two different steady-state
periods and compute statistics separately for each of the two steady-state periods.
Oil and gas refineries and distribution centers are additional examples of nonter-
minating systems.
8.5 Experimenting with Terminating Simulations

Experiments involving terminating simulations are usually conducted by making
several simulation runs or replications of the period of interest using a different
seed value for each run. This procedure enables statistically independent and un-
biased observations to be made on the response variables of interest in the system
(such as the number of customers in the system or the utilization of resources)
over the period simulated. Statistics are also often gathered on performance mea-
sures for successive intervals during the period.
For terminating simulations, we are usually interested in final production
counts and changing patterns of behavior over time (store opening until closing,
for example) rather than the over all, or long-term, average behavior. It would be
absurd, for example, to conclude that because two waitresses are busy an average
of only 40 percent during the day that only one waitress is needed. This average
measure reveals little about the utilization of the waitresses during peak periods.
A more detailed report of how long customers wait for a waitress to serve them
during successive intervals during the workday may reveal that three waitresses
har01307_ch08_211-242.indd 227 20/12/10 1:52 PM

are needed to handle peak periods whereas only one waitress is necessary during
off-peak hours. Therefore, statistics are also often gathered on performance mea-
sures for successive intervals during the period.
For terminating simulations, there are three important questions to answer in
running the experiment:
1. What should be the initial state of the model?
2. What is the terminating event or time of the simulation (i.e., run length)?
3. How many replications should be made?
8.5.1 Selecting the Initial Model State

The initial state of the model represents how the system looks (or is initialized)
at the beginning of the simulation. The barbershop example presented in sec-
tion 8.2.2 is a terminating system. The barbershop opens at 8:00 A.M. and closes
around 12:00 noon. The initial state of the shop at 8:00 A.M. is that there are no
customers in the shop and all barbers are idle and reading the newspaper. There-
fore, the initial state of the model at the beginning of the simulation is that no
customers (entities) are in the barbershop (system) and all barbers (resources) are
idle. The system is empty and idle until entities begin to arrive in the system as
described by the logic in the model.
All terminating simulations do not necessarily begin with the system being
empty and idle. For example, to measure the time that it takes for everyone to exit
an arena after a sporting event, the simulation would be initialized with fans in
their seats just before the sporting event ends. Therefore, the initial state of the
model would reflect a large number of entities (fans) in the system.
8.5.2 Selecting a Terminating Event to Control Run Length

A terminating event is an event that occurs during the simulation that causes the
simulation run to end. This terminating event might occur when a certain time
of day is reached (closing time of an airport’s newspaper stand) or when certain
conditions are met (the last passenger boards an aircraft). If the termination event
is based on the satisfaction of conditions other than time, then the time that the
simulation will end is not known in advance. Using the barbershop as an example,
the simulation terminates sometime after 12:00 noon when the haircut for the last
customer allowed in the shop is completed. Therefore, the termination condition
would be when the simulation clock time is greater than or equal to 12:00 noon
and when the last entity (customer) exits the system (barbershop).
8.5.3 Determining the Number of Replications

Given the initial or starting state of the simulation model and the terminating
condition(s) for the simulation run, the analyst is ready to begin conducting exper-
iments with the model. Experiments involving terminating simulations are usu-
ally conducted by making several simulation runs or replications of the period of
interest using a different stream of random numbers for each run. This procedure
har01307_ch08_211-242.indd 228 20/12/10 1:52 PM

enables statistically independent and unbiased observations to be made on re-

sponse variables of interest in the system (the number of customers in the system,
the utilization of resources) over the period simulated, as discussed in section 8.3.
The answer to the question of how many replications are necessary is usually
based on the analyst’s desired half-width of a confidence interval. As a general
guideline, begin by making 10 independent replications of the simulation and
add more replications until the desired confidence interval half-width is reached.
For the barbershop example, 18 independent replications of the simulation were
required to achieve the desired confidence interval for the expected number of
haircuts given on Saturday mornings.
8.6 Experimenting with Nonterminating Simulations

The issues associated with generating meaningful output for terminating simula-
tions are somewhat different from those associated with nonterminating systems.
However, the same statistical methods are used to analyze the output from both
types of simulations, provided that the output observations are unbiased, indepen-
dent, and normally distributed.
Nonterminating simulations are conducted when we are interested in the
steady-state behavior of the simulation model. As such, the following topics must
be addressed:
1. Determining and eliminating the initial warm-up bias.
2. Obtaining sample observations.
3. Determining run length.
8.6.1 Determining the Warm-up Period

In a steady-state simulation, we are interested in the steady-state behavior of the
model. If a model starts out empty of entities, it usually takes some time before it
reaches steady state. In a steady-state condition, the response variables in the sys-
tem (like mean waiting time or throughput rate) exhibit statistical regularity (the
distribution of a response variable is approximately the same from one time pe-
riod to the next). Figure 8.3 illustrates the typical behavior of an output response
measure as the simulation progresses through time. Note that the distribution of
the output response changes very little once steady state is reached.
The time that it takes to reach steady state is a function of the activity times
and the amount of activity taking place. For some models, steady state might be
reached in a few hours of simulation time. For other models, it may take several
hundred hours to reach steady state. In modeling steady-state behavior, we have
the problem of determining when a model reaches steady state. This start-up pe-
riod is usually referred to as the warm-up period (transient period). We want to
wait until after the warm-up period before we start gathering any statistics. This
way we eliminate any bias due to observations taken during the transient state
of the model. For example, in Figure 8.3 any observations recorded before the
har01307_ch08_211-242.indd 229 20/12/10 1:52 PM

FIGURE 8.3
Behavior of model’s output response as it reaches steady state.
Warm-up ends
Transient state Steady state
Averaged
output
response
measure
(y )
Distribution of output response
1 2 3 4 5 6
Simulation time
warm-up ends would be biased to the low side. Therefore, we delete these ob-
servations from the beginning of the run and use the remaining observations to
estimate the true mean response of the model.
While several methods have been developed for estimating the warm-up
time, the easiest and most straightforward approach is to run a preliminary simu-
lation of the system, preferably with several (5 to 10) replications, average the
output values at each time step across replications, and observe at what time the
system reaches statistical stability. This usually occurs when the averaged output
response begins to flatten out or a repeating pattern emerges. Plotting each data
point (averaged output response) and connecting them with a line usually helps
to identify when the averaged output response begins to flatten out or begins a
repeating pattern. Sometimes, however, the variation of the output response is
so large that it makes the plot erratic, and it becomes difficult to visually identify
the end of the warm-up period. Such a case is illustrated in Figure 8.4. The raw
data plot in Figure 8.4 was produced by recording a model’s output response for
a queue’s average contents during 50-hour time periods (time slices) and averag-
ing the output values from each time period across five replications. In this case,
the model was initialized with several entities in the queue. Therefore, we need to
eliminate this apparent upward bias before recording observations of the queue’s
average contents. Table 8.4 shows this model’s output response for the 20 time
periods (50 hours each) for each of the five replications. The raw data plot in Fig-
ure 8.4 was constructed using the 20 values under the ȳ i column in Table 8.4.
When the model’s output response is erratic, as in Figure 8.4, it is use-
ful to “smooth” it with a moving average. A moving average is constructed by
har01307_ch08_211-242.indd 230 20/12/10 1:52 PM

FIGURE 8.4
SimRunner uses the Welch moving-average method to help identify the end of the warm-up period
that occurs around the third or fourth period (150 to 200 hours) for this model.
calculating the arithmetic average of the w most recent data points (averaged out-
put responses) in the data set. You have to select the value of w, which is called
the moving-average window. As you increase the value of w, you increase the
“smoothness” of the moving average plot. An indicator for the end of the warm-
up time is when the moving average plot appears to flatten out. Thus the routine is
to begin with a small value of w and increase it until the resulting moving average
plot begins to flatten out.
The moving-average plot in Figure 8.4 with a window of six (w 6) helps to
identify the end of the warm-up period—when the moving average plot appears
to flatten out at around the third or fourth period (150 to 200 hours). Therefore,
we ignore the observations up to the 200th hour and record only those after the
200th hour when we run the simulation. The moving-average plot in Figure 8.4
was constructed using the 14 values under the yi (6) column in Table 8.4 that were
computed using

sw ȳ is
w
_ if i w 1, . . . , m w
2w 1
ȳ i (w)
s(i1) ȳis
i1
__ if i 1, . . . , w
2i 1
where m denotes the total number of periods and w denotes the window of the
moving average (m 20 and w 6 in this example).
har01307_ch08_211-242.indd 231 20/12/10 1:52 PM

TABLE 8.4 Welch Moving Average Based on Five Replications and 20 Periods
Average
Total Contents Welch
Observed Mean Queue Contents (Replications) Contents for per Period, Moving
Period, Average,
yi_Totali
Period (i) Time 1 2 3 4 5 Totali yi(6)
5
1 50 4.41 4.06 6.37 11.72 1.71 28.27 5.65 5.65
2 100 3.15 2.95 2.94 3.52 2.86 15.42 3.08 3.95
3 150 2.58 2.50 3.07 3.14 4.32 15.61 3.12 3.73
4 200 2.92 3.07 4.48 4.79 3.47 18.73 3.75 3.84
5 250 3.13 3.28 2.34 3.32 3.19 15.26 3.05 3.94
6 300 2.51 3.07 5.45 5.15 5.34 21.52 4.30 3.82
7 350 5.09 3.77 4.44 3.58 2.63 19.51 3.90 3.86
8 400 3.15 2.89 3.63 4.43 4.15 18.25 3.65 3.83
9 450 2.79 3.40 9.78 4.13 4.48 24.58 4.92 3.84
10 500 4.07 3.62 4.50 2.38 3.42 17.99 3.60 3.93
11 550 3.42 3.74 2.46 3.08 2.49 15.19 3.04 3.89
12 600 3.90 1.47 5.75 5.34 4.19 20.65 4.13 3.99
13 650 3.63 3.77 2.14 4.24 6.36 20.14 4.03 3.99
14 700 9.12 4.25 3.83 3.19 5.91 26.30 5.26 3.96
15 750 3.88 3.54 3.08 2.94 2.51 15.95 3.19
16 800 3.16 6.91 2.70 5.64 2.57 20.98 4.20
17 850 5.34 3.17 2.47 2.73 2.68 16.39 3.28
18 900 2.84 3.54 2.33 10.14 3.20 22.05 4.41
19 950 2.65 4.64 5.18 4.97 3.93 21.37 4.27
20 1000 3.27 4.68 3.39 2.95 3.46 17.75 3.55
This is a somewhat busy-looking equation that turns out not to be so bad

when you put some numbers into it, which we will now do to compute the Welch
moving average, yi(6), column in Table 8.4. We use the bottom part of the equa-
tion to compute the moving averages up to the wth moving average. Our w 6,
so we will compute ȳ 1 (6), ȳ 2 (6), ȳ 3 (6), ȳ 4 (6), ȳ 5 (6), and ȳ 6 (6) using the bottom part
of the yi (w) equation as follows:
y 5.65 5.65
ȳ 1(6) _1 _
1 1
ȳ 1 ȳ 2 ȳ 3 5.65 3.08 3.12 3.95
ȳ 2(6) __ __
3 3
ȳ__ ȳ ȳ ȳ ȳ 5.65 3.08 3.12 3.75 3.05 3.73
ȳ 3(6) 1 2 3 4 5
___
5 5
ȳ 1 ȳ 2 ȳ 3 ȳ 4 ȳ 5 ȳ 6 ȳ 7 ȳ 8 ȳ 9 ȳ 10 ȳ 11
ȳ 6(6) _____ 3.82
11
har01307_ch08_211-242.indd 232 20/12/10 1:52 PM

Notice the pattern that as i increases, we average more data points together,
with the ith data point appearing in the middle of the sum in the numerator (an
equal number of data points are on each side of the center data point). This con-
tinues until we reach the (w 1)th moving average, when we switch to the top
part of the ȳ i(w) equation. For our example, the switch occurs at the 7th moving
average because w 6. From this point forward, we average the 2w 1 closest
data points. For our example, we average the 2(6) 1 13 closest data points
(the ith data point plus the w 6 closest data points above it and the w 6 closest
data point below it in Table 8.4 as follows:
y1 y2 y3 y4 y5 y6 y7 y8 y9 y10 y11 y12 y13
y7(6) ______
13
y7(6) 3.86
y2 y3 y4 y5 y6 y7 y8 y9 y10 y11 y12 y13 y14

y8(6) ______
13
y8(6) 3.83
y3 y4 y5 y6 y7 y8 y9 y10 y11 y12 y13 y14 y15

y9(6) ______
13
y9(6) 3.84
y8 y9 y10 y11 y12 y13 y14 y15 y16 y17 y18 y19 y20
y14(6) ______
13
y14(6) 3.96
Eventually we run out of data and have to stop. The stopping point occurs when
i m w. In our case with m 20 periods and w 6, we stopped when i
20 6 14.
The development of this graphical method for estimating the end of the warm-
up time is attributed to Welch (Law 2007). This method is sometimes referred to
as the Welch moving-average method and is implemented in SimRunner. Note that
when applying the method that the length of each replication should be relatively
long and the replications should allow even rarely occurring events such as infre-
quent downtimes to occur many times. Law (2007) recommend that w not exceed
a value greater than about m4. To determine a satisfactory warm-up time using
Welch’s method, one or more key output response variables, such as the average
number of entities in a queue or the average utilization of a resource, should be
monitored for successive periods. Once these variables begin to exhibit steady
state, a good practice to follow would be to extend the warm-up period by 20 to
30 percent. This approach is simple, conservative, and usually satisfactory. The
danger is in underestimating the warm-up period, not overestimating it.
har01307_ch08_211-242.indd 233 20/12/10 1:52 PM

8.6.2 Obtaining Sample Observations

It seems that for nonterminating simulations, you could simply run the simulation
long enough for statistical output to stabilize and then analyze the results. In other
words, you might use the results from a single long run after removing the bias
created by the warm-up period. For computing confidence intervals, however,
independent sample observations must be created. In a terminating simulation,
simply running multiple replications creates independent sample observations.
For steady-state simulations, we have several options for obtaining independent
sample observations. Two approaches that are widely used are running multiple
replications and interval batching.
Independent Observations via Replications

Running multiple replications for nonterminating simulations is similar to the
way it is done for terminating simulations. The only differences are that (1) the
initial warm-up bias must be determined and eliminated and (2) an appropriate
run length must be determined. The warm-up issue was addressed in section 8.6.1,
and the run length issue is addressed in section 8.6.3. Once the replications are
made, confidence intervals can be computed as described earlier in section 8.3.
This was done for the ATM simulation of Lab Chapter 1 to produce Figure 8.2.
Note the warm-up time of 1,000 hours and total simulation time of 1,500 hours
at the top of the ProModel output report. For this simulation, statistics were col-
lected from time 1,000 hours to 1,500 hours.
One advantage of running independent replications is that it ensures that
samples are independent. On the negative side, running through the warm-up
phase for each replication extends the length of time to perform the replications.
Furthermore, there is a possibility that the length of the warm-up period is un-
derestimated, causing biased results from each replication. However, because the
replication method ensures that samples are independent, we strongly encourage
its use when the warm-up period is properly detected and removed.
Independent Observations via Batch Means

Interval batching (also referred to as the batch means technique) is a method in
which a single, long run is made and statistics are collected during separate pe-
riods. Although 5 to 10 replications are typically used to determine the warm-up
period in preliminary experiments (see section 8.6.1), the simulation is run only
once to collect observations for statistical analysis purposes. Placing the output
values from the simulation run into groups or batches forms the set of observa-
tions. For example, Figure 8.5 illustrates the output response from a simulation
during a single run. Beyond the warm-up period, the run is divided into nonover-
lapping intervals. Note that the intervals are of an equal amount of time (as in
Figure 8.5) or an equal number of observations (every 1,000 entities observed).
Next, the output responses that were generated during a common time interval are
“batched” together. If the output measure of interest is a time-weighted statistic
(such as the average number of entities in the system), then the batch interval
har01307_ch08_211-242.indd 234 20/12/10 1:52 PM

FIGURE 8.5 Warm-up ends

Individual statistics
on the output measure
Output
are computed for each
batch interval. response
measure
(y)
Batch Batch Batch Batch Batch

interval 1 interval 2 interval 3 interval 4 interval 5
Simulation time
TABLE 8.5 A Single Replication Divided into Batch Intervals
Batch Interval (i) Within Batch Interval Observations yij Average xi
1 y11, y12, y13, y14, . . . , y1m x1

2 y21, y22, y23, y24, . . . , y2m x2
3 y31, y32, y33, y34, . . . , y3m x3
4 y41, y42, y43, y44, . . . , y4m x4
5 y51, y52, y53, y54, . . . , y5m x5
n yn1, yn2, yn3, yn4, . . . , ynm xn
should be based on time. If the output measure is based on observations (like the
waiting time of entities in a queue), then the batch interval is typically based on
the number of observations.
Table 8.5 details how a single, long simulation run (one replication), like
the one shown in Figure 8.5, is partitioned into batch intervals for the purpose of
obtaining (approximately) independent and identically distributed observations
of a simulation model’s output response. Note the similarity between Table 8.3
of section 8.3 and Table 8.5. As in Table 8.3, the observations represent the time
an entity waited in queues during the simulation. The waiting time for an entity is
denoted by yij, where the subscript i denotes the interval of time (batch interval)
from which the observation came and the subscript j denotes the value of the
counter used to count entities as they exit the system during a particular batch
interval. For example, y23 is the waiting time for the third entity that was processed
through the system during the second batch interval of time.
Because these observations are all from a single run (replication), they are
not statistically independent. For example, the waiting time of the mth entity exit-
ing the system during the first batch interval, denoted y1m, is correlated with the
har01307_ch08_211-242.indd 235 20/12/10 1:52 PM

waiting time of the first entity to exit the system during the second batch interval,
denoted y21. This is because if the waiting time observed for one entity is long, it
is likely that the waiting time for the next entity observed is going to be long, and
vice versa. Therefore, adjacent observations in Table 8.5 will usually be autocor-
related. However, the value of y14 can be somewhat uncorrelated with the value
of y24 if they are spaced far enough apart such that the conditions that resulted
in the waiting time value of y14 occurred so long ago that they have little or no
influence on the waiting time value of y24. Therefore, most of the observations
within the interior of one batch interval can become relatively uncorrelated with
most of the observations in the interior of other batch intervals, provided they are
spaced sufficiently far apart. Thus the goal is to extend the batch interval length
until there is very little correlation (you cannot totally eliminate it) between the
observations appearing in different batch intervals. When this occurs, it is reason-
able to assume that observations within one batch interval are independent of the
observations within other batch intervals.
With the independence assumption in hand, the observations within a batch
interval can be treated in the same manner as we treated observations within a
replication in Table 8.3. Therefore, the values in the Average column in Table 8.5,
which represent the average amount of time the entities processed during the ith
batch interval waited in queues, are computed as follows:
mj1 yij
xi _ m
where m is the number of entities processed through the system during the batch
interval of time and yij is the time that the jth entity processed through the system
waited in queues during the ith batch interval. The xi values in the Average column
are approximately independent and identically distributed and can be used to com-
pute a sample mean for estimating the true average time an entity waits in queues:
ni1 xi
x̄ _ n
The sample standard deviation is also computed as before:

___
i1 [xi x̄]

n 2
s __
n1
And, if we assume that the observations (x1 through xn) are normally distributed
(or at least approximately normally distributed), a confidence interval is com-
puted by
(tn1,␣
2 )s
x̄ _ _
n
where tn1,␣
2 is a value from the Student’s t table in Appendix B for an ␣ level of
significance.
ProModel automatically computes point estimates and confidence intervals
using the above method. Figure 8.6 is a ProModel output report from a single,
har01307_ch08_211-242.indd 236 20/12/10 1:52 PM

FIGURE 8.6
Batch means technique used on ATM simulation in Lab Chapter 1. Warm-up period is from 0 to 1,000 hours. Each
batch interval is 500 hours in length.
-----------------------------------------------------------------
General Report
Output from C:\Bowden\3rd Edition\ATM System Ch8.MOD [ATM System]
Date: Nov/26/2010 Time: 04:44:24 PM
-----------------------------------------------------------------
Scenario : Normal Run
Replication : 1 of 1
Period : All
Warmup Time : 1000 hr
Simulation Time : 3500 hr
-----------------------------------------------------------------
LOCATIONS
Average
Location Scheduled Total Minutes Average Maximum Current
Name Hours Capacity Entries Per Entry Contents Contents Contents % Util
--------- --------- -------- ------- --------- -------- -------- -------- ------
ATM Queue 500 999999 9903 9.457 3.122 26 18 0.0 (Batch 1)
ATM Queue 500 999999 10065 9.789 3.284 24 0 0.0 (Batch 2)
ATM Queue 500 999999 9815 8.630 2.823 23 16 0.0 (Batch 3)
ATM Queue 500 999999 9868 8.894 2.925 24 1 0.0 (Batch 4)
ATM Queue 500 999999 10090 12.615 4.242 34 0 0.0 (Batch 5)
ATM Queue 500 999999 9948.2 9.877 3.279 26.2 7 0.0 (Average)
ATM Queue 0 0 122.441 1.596 0.567 4.494 9.165 0.0 (Std. Dev.)
ATM Queue 500 999999 9796.19 7.895 2.575 20.620 -4.378 0.0 (95% C.I. Low)
ATM Queue 500 999999 10100.2 11.860 3.983 31.779 18.378 0.0 (95% C.I. High)
long run of the ATM simulation in Lab Chapter 1 with the output divided into
five batch intervals of 500 hours each after a warm-up time of 1,000 hours. The
column under the Locations section of the output report labeled “Average Minutes
Per Entry” displays the average amount of time that customer entities waited in
the queue during each of the five batch intervals. These values correspond to the
xi values of Table 8.5. Note that the 95 percent confidence interval automatically
computed by ProModel in Figure 8.6 is comparable, though not identical, to the
95 percent confidence interval in Figure 8.2 for the same output statistic.
Establishing the batch interval length such that the observations x1, x2, . . . , xn
of Table 8.5 (assembled after the simulation reaches steady state) are approxi-
mately independent is difficult and time-consuming. There is no foolproof method
for doing this. However, if you can generate a large number of observations, say
n 100, you can gain some insight about the independence of the observations
by estimating the autocorrelation between adjacent observations (lag-1 autocorre-
lation). See Chapter 5 for an introduction to tests for independence and autocor-
relation plots. The observations are treated as being independent if their lag-1
autocorrelation is zero. Unfortunately, current methods available for estimating the
lag-1 autocorrelation are not very accurate. Thus we may be persuaded that our ob-
servations are almost independent if the estimated lag-1 autocorrelation computed
from a large number of observations falls between −0.20 to 0.20. The word
almost is emphasized because there is really no such thing as almost independent
har01307_ch08_211-242.indd 237 20/12/10 1:52 PM

(the observations are independent or they are not). What we are indicating here is
that we are leaning toward calling the observations independent when the estimate
of the lag-1 autocorrelation is between −0.20 and 0.20. Recall that autocorrela-
tion values fall between 1. Before we elaborate on this idea for determining if an
acceptable batch interval length has been used to derive the observations, let’s talk
about positive autocorrelation versus negative autocorrelation.
Positive autocorrelation is a bigger enemy to us than negative autocorrela-
tion because our sample standard deviation s will be biased low if the observa-
tions from which it is computed are positively correlated. This would result in a
falsely narrow confidence interval (smaller half-width), leading us to believe that
we have a better estimate of the mean ␮ than we actually have. Negatively cor-
related observations have the opposite effect, producing a falsely wide confidence
interval (larger half-width), leading us to believe that we have a worse estimate
of the mean ␮ than we actually have. A negative autocorrelation may lead us
to waste time collecting additional observations to get a more precise (narrow)
confidence interval but will not result in a hasty decision. Therefore, an emphasis
is placed on guarding against positive autocorrelation. We present the following
procedure adapted from Banks et al. (2010) for estimating an acceptable batch
interval length.
For the observations x1, x2, . . . , xn derived from n batches of data assembled
after the simulation has researched steady state as outlined in Table 8.5, compute
an estimate of the lag-1 autocorrelation ␳^1 using
i1 (xi − x̄)(xi1 − x̄)
n−1
␳^1 __
s2(n − 1)
where s2 is the sample variance (sample standard deviation squared) and x̄ is the
sample mean of the n observations. Recall the recommendation that n should be
at least 100 observations. If the estimated lag-1 autocorrelation is not between
0.20 and 0.20, then extend the original batch interval length by 50 to 100
percent, rerun the simulation to collect a new set of n observations, and check
the lag-1 autocorrelation between the new x1 to xn observations. This would be
repeated until the estimated lag-1 autocorrelation falls between 0.20 and 0.20
(or you give up trying to get within the range). Note that falling below 0.20 is
less worrisome than exceeding 0.20.
Achieving ⴚ0.20 ␳^1 0.20: Upon achieving an estimated lag-1
autocorrelation within the desired range, “rebatch” the data by combining
adjacent batch intervals of data into a larger batch. This produces a
smaller number of observations that are based on a larger batch interval
length. The lag-1 autocorrelation for this new set of observations will
likely be closer to zero than the lag-1 autocorrelation of the previous set
of observations because the new set of observations is based on a larger
batch interval length. Note that at this point, you are not rerunning the
simulation with the new, longer batch interval but are “rebatching” the
output data from the last run. For example, if a batch interval length of
125 hours produced 100 observations (x1 to x100) with an estimated lag-1
har01307_ch08_211-242.indd 238 20/12/10 1:52 PM

autocorrelation ␳^1 0.15, you might combine four contiguous batches

of data into one batch with a length of 500 hours (125 hours 4 500
hours). “Rebatching” the data with the 500-hour batch interval length
would leave you 25 observations (x1 to x25) from which to construct the
confidence interval. We recommend that you “rebatch” the data into 10 to
30 batch intervals.
Not Achieving ⴚ0.20 ␳^1 0.20: If obtaining an estimated lag-1
autocorrelation within the desired range becomes impossible because you
cannot continue extending the length of the simulation run, then rebatch
the data from your last run into no more than about 10 batch intervals and
construct the confidence interval. Interpret the confidence interval with the
apprehension that the observations may be significantly correlated.
Lab Chapter 8 provides an opportunity for you to gain experience apply-
ing these criteria to a ProModel simulation experiment. However, remember that
there is no universally accepted method for determining an appropriate batch in-
terval length. See Banks et al. (2010) for additional details and variations on this
approach, such as a concluding hypothesis test for independence on the final set
of observations. The danger is in setting the batch interval length too short, not
too long. This is the reason for extending the length of the batch interval in the
last step. A starting point for setting the initial batch interval length from which to
begin the process of evaluating the lag-1 autocorrelation estimates is provided in
section 8.6.3.
In summary, the statistical methods in this chapter are applied to the data
compiled from batch intervals in the same manner they were applied to the data
compiled from replications. And, as before, the accuracy of point and interval
estimates generally improves as the sample size (number of batch intervals) in-
creases. In this case we increase the number of batch intervals by extending the
length of the simulation run. As a general guideline, the simulation should be run
long enough to create around 10 batch intervals and possibly more, depending on
the desired confidence interval half-width. If a trade-off must be made between
the number of batch intervals and the batch interval length, then err on the side of
increasing the batch interval length. It is better to have a few independent observa-
tions than to have several autocorrelated observations when using the statistical
methods presented in this text.
8.6.3 Determining Run Length

Determining run length for terminating simulations is quite simple because a
natural event or time point defines it. Determining the run length for a steady-
state simulation is more difficult because the simulation could be run indefinitely.
Obviously, running extremely long simulations can be very time-consuming, so
the objective is to determine an appropriate run length to ensure a representative
sample of the steady-state response of the system.
The recommended length of the simulation run for a steady-state simulation
depends on the interval between the least frequently occurring event and the type
har01307_ch08_211-242.indd 239 20/12/10 1:52 PM

FIGURE 8.7 Replication 3

Batch intervals
compared with
replications. Replication 2
Replication 1
Batch
Batch Batch Batch

interval 1 interval 2 interval 3
Simulation time
of sampling method (replication or interval batching) used. If running indepen-

dent replications, it is usually a good idea to run the simulation long enough past
the warm-up point to let every type of event (including rare ones) happen many
times and, if practical, several hundred times. Remember, the longer the model
is run, the more confident you can become that the results represent the steady-
state behavior. A guideline for determining the initial run length for the interval
batching method is much the same as for the replication method. Essentially, the
starting point for selecting the length of each batch interval is the length of time
chosen to run the simulation past the warm-up period when using the replication
method. The total run length is the sum of the time for each batch interval plus the
initial warm-up time. This guideline is illustrated in Figure 8.7.
The choice between running independent replications or batch intervals can
be made based on how long it takes the simulation to reach steady state. If the
warm-up period is short (that is, it doesn’t take long to simulate through the warm-
up period), then running replications is preferred over batch intervals because run-
ning replications guarantees that the observations are independent. However, if
the warm-up period is long, then the batch interval method is preferred because
the time required to get through the warm-up period is incurred only once, which
reduces the amount of time to collect the necessary number of observations for a
desired confidence interval. Also, because a single, long run is made, chances are
that a better estimate of the model’s true steady-state behavior is obtained.
8.7 Summary
Simulation experiments can range from a few replications (runs) to a large num-
ber of replications. While “ballpark” decisions require little analysis, more precise
decision making requires more careful analysis and extensive experimentation.
har01307_ch08_211-242.indd 240 20/12/10 1:52 PM

Correctly configuring the model to produce valid output is essential to making

sound decisions.
This chapter has identified the issues involved in producing and analyzing
simulation output. Key issues include determining run length, number of replica-
tions, number of batch intervals, batch interval length, and warm-up times. When
running multiple replications, the modeler needs to understand the concepts of
random number streams and random number seeds. Several guidelines were pre-
sented throughout the chapter to help guide the analyst in interpreting simulation
output data using point and interval estimates. Remember that the guidelines are
not rigid rules that must be followed but rather are common practices that have
worked well in the past.

1. Define population, sample, and sample size.
2. If you run the same simulation model twice, would you expect to see
any difference in the results? Explain.
3. How are multiple simulation replications produced?
4. What do x and s define, and why are they called point estimators?
5. What does a confidence interval express?
6. What does the confidence interval on simulation output reveal about the
validity of the model?
7. In order to apply statistical measures, such as standard deviation
and confidence interval estimation, what assumptions must be made
regarding simulation output?
8. What is the central limit theorem and how is it applied to simulation
output?
9. Give an example of a system in a transient state; in a steady state.
10. How do you determine the length of a terminating simulation run?
11. How many replications should be made for terminating simulations?
12. Verify that the 95 percent confidence interval has been correctly
computed by ProModel for the Average Minutes Per Entry statistic
shown in Figure 8.2. The ProModel half-width of 1.408 minutes
represents an approximate 14.3 percent error in the point estimate
x̄ 9.820 minutes. Determine the approximate number of additional
replications needed to estimate the average time that customer entities
wait in the ATM queue with a percentage error of 5 percent and a
confidence level of 95 percent.
13. Describe two methods for obtaining sample observations for a
nonterminating simulation.
14. When determining the batch interval length, what is the danger in
making the interval too small?
har01307_ch08_211-242.indd 241 20/12/10 1:52 PM

15. Given the five batch means for the Average Minutes Per Entry
of customer entities at the ATM queue location in Figure 8.6,
estimate the lag-1 autocorrelation ␳^1. Note that five observations are
woefully inadequate for obtaining an accurate estimate of the lag-1
autocorrelation. Normally you will want to base the estimate on at
least 100 observations. The question is designed to give you some
experience using the ␳^1 equation so that you will understand what the
Stat::Fit software is doing when you use it in Lab Chapter 8 to crunch
through an example with 100 observations.
16. Construct a Welch moving average with a window of 2 (w 2) using
the data in Table 8.4 and compare it to the Welch moving average with
a window of 6 (w 6) presented in Table 8.4.
References
Bateman, Robert E.; Royce O. Bowden; Thomas J. Gogg; Charles R. Harrell; and Jack
R. A. Mott. System Improvement Using Simulation. Orem, UT: PROMODEL Corp.,
1997.
Hines, William W.; Douglas C. Montgomery; David M. Goldsman; and Connie M. Borror.
Probability and Statistics in Engineering. 4th ed. New York: John Wiley and Sons,
2003.
Law, Averill M. Simulation Modeling and Analysis. New York: McGraw-Hill, 2007.
Montgomery, Douglas C. Design and Analysis of Experiments. New York: John Wiley &
Sons, 1991.
Petersen, Roger G. Design and Analysis of Experiments. New York: Marcel Dekker, 1985.
har01307_ch08_211-242.indd 242 20/12/10 1:52 PM

C H A P T E R
9 COMPARING SYSTEMS
“The method that proceeds without analysis is like the groping of a blind man.”
—Socrates
9.1 Introduction
In many cases, simulations are conducted to compare two or more alternative de-
signs of a system with the goal of identifying the superior system relative to some
performance measure. Comparing alternative system designs requires careful
analysis to ensure that differences being observed are attributable to actual differ-
ences in performance and not to statistical variation. This is where running either
multiple replications or batches is required. Suppose, for example, that method A
for deploying resources yields a throughput of 100 entities for a given time period
while method B results in 110 entities for the same period. Is it valid to conclude
that method B is better than method A, or might additional replications actually
lead to the opposite conclusion?
You can evaluate alternative configurations or operating policies by perform-
ing several replications of each alternative and comparing the average results
from the replications. Statistical methods for making these types of comparisons
are called hypotheses tests. For these tests, a hypothesis is first formulated (for ex-
ample, that methods A and B both result in the same throughput) and then a test is
made to see whether the results of the simulation lead us to reject the hypothesis.
The outcome of the simulation runs may cause us to reject the hypothesis that
methods A and B both result in equal throughput capabilities and conclude that
the throughput does indeed depend on which method is used.
This chapter extends the material presented in Chapter 8 by providing statisti-
cal methods that can be used to compare the output of different simulation models
that represent competing designs of a system. The concepts behind hypothesis
testing are introduced in section 9.2. Section 9.3 addresses the case when two
243
har01307_ch09_243-272.indd 243 12/17/10 12:01 PM

alternative system designs are to be compared, and section 9.4 considers the case
when more than two alternative system designs are to be compared. Additionally,
a technique called common random numbers is described in section 9.5 that can
sometimes improve the accuracy of the comparisons.
9.2 Hypothesis Testing

An inventory allocation example will be used to further explore the use of hypoth-
esis testing for comparing the output of different simulation models. Suppose a
production system consists of four machines and three buffer storage areas. Parts
entering the system require processing by each of the four machines in a serial
fashion (Figure 9.1). A part is always available for processing at the first machine.
After a part is processed, it moves from the machine to the buffer storage area for
the next machine, where it waits to be processed. However, if the buffer is full,
the part cannot move forward and remains on the machine until a space becomes
available in the buffer. Furthermore, the machine is blocked and no other parts
can move to the machine for processing. The part exits the system after being
processed by the fourth machine.
The question for this system is how best to allocate buffer storage between the
machines to maximize the throughput of the system (number of parts completed
per hour). The production control staff has identified two candidate strategies for
allocating buffer capacity (number of parts that can be stored) between machines,
and simulation models have been built to evaluate the proposed strategies.
FIGURE 9.1
Production system with
four workstations and
three buffer storage
areas.
.72 4.56
har01307_ch09_243-272.indd 244 12/17/10 12:01 PM

Chapter 9 Comparing Systems 245
Suppose that Strategy 1 and Strategy 2 are the two buffer allocation strate-
gies proposed by the production control staff. We wish to identify the strategy that
maximizes the throughput of the production system (number of parts completed
per hour). Of course, the possibility exists that there is no significant difference in
the performance of the two candidate strategies. That is to say, the mean through-
put of the two proposed strategies is equal. A starting point for our problem is
to formulate our hypotheses concerning the mean throughput for the production
system under the two buffer allocation strategies. Next we work out the details
of setting up our experiments with the simulation models built to evaluate each
strategy. For example, we may decide to estimate the true mean performance of
each strategy (1 and 2) by simulating each strategy for 16 days (24 hours per
day) past the warm-up period and replicating the simulation 10 times. After we
run experiments, we would use the simulation output to evaluate the hypotheses
concerning the mean throughput for the production system under the two buffer
allocation strategies.
In general, a null hypothesis, denoted H0, is drafted to state that the value of
1 is not significantly different than the value of 2 at the level of significance.
An alternate hypothesis, denoted H1, is drafted to oppose the null hypothesis H0.
For example, H1 could state that 1 and 2 are different at the level of signifi-
cance. Stated more formally:
H0: 1 2 or equivalently H0: 1 2 0
H1: 1 2 or equivalently H1: 1 2 0
In the context of the example problem, the null hypothesis H0 states that the
mean throughputs of the system due to Strategy 1 and Strategy 2 do not differ.
The alternate hypothesis H1 states that the mean throughputs of the system due to
Strategy 1 and Strategy 2 do differ. Hypothesis testing methods are designed such
that the burden of proof is on us to demonstrate that H0 is not true. Therefore, if our
analysis of the data from our experiments leads us to reject H0, we can be confident
that there is a significant difference between the two population means. In our
example problem, the output from the simulation model for Strategy 1 represents
possible throughput observations from one population, and the output from the
simulation model for Strategy 2 represents possible throughput observations from
another population.
The level of significance in these hypotheses refers to the probability of
making a Type I error. A Type I error occurs when we reject H0 in favor of H1
when in fact H0 is true. Typically is set at a value of 0.05 or 0.01. However, the
choice is yours, and it depends on how small you want the probability of making
a Type I error to be. A Type II error occurs when we fail to reject H0 in favor of
H1 when in fact H1 is true. The probability of making a Type II error is denoted as
. Hypothesis testing methods are designed such that the probability of making
a Type II error, , is as small as possible for a given value of . The relationship
between and is that increases as decreases. Therefore, we should be care-
ful not to make too small.
We will test these hypotheses using a confidence interval approach to
determine if we should reject or fail to reject the null hypothesis in favor of the
har01307_ch09_243-272.indd 245 12/17/10 12:01 PM

alternative hypothesis. The reason for using the confidence interval method is
that it is equivalent to conducting a two-tailed test of hypothesis with the added
benefit of indicating the magnitude of the difference between 1 and 2 if they
are in fact significantly different. The first step of this procedure is to construct a
confidence interval to estimate the difference between the two means (1 2).
This can be done in different ways depending on how the simulation experiments
are conducted (we will discuss this later). For now, let’s express the confidence
interval on the difference between the two means as
P[(x̄ 1 x̄ 2) hw 1 2 (x̄ 1 x̄ 2) hw] 1
where hw denotes the half-width of the confidence interval. Notice the similari-
ties between this confidence interval expression and the one given on page 217 in
Chapter 8. Here we have replaced x̄ with x̄ 1 x̄ 2 and with 1 2.
If the two population means are the same, then 1 2 0, which is our
null hypothesis H0. If H0 is true, our confidence interval should include zero with
a probability of 1 . This leads to the following rule for deciding whether to
reject or fail to reject H0. If the confidence interval includes zero, we fail to reject
H0 and conclude that the value of 1 is not significantly different than the value
of 2 at the level of significance (the mean throughput of Strategy 1 is not sig-
nificantly different than the mean throughput of Strategy 2). However, if the con-
fidence interval does not include zero, we reject H0 and conclude that the value
of 1 is significantly different than the value of 2 at the level of significance
(throughput values for Strategy 1 and Strategy 2 are significantly different).
Figure 9.2(a) illustrates the case when the confidence interval contains zero,
leading us to fail to reject the null hypothesis H0 and conclude that there is no
significant difference between 1 and 2. The failure to obtain sufficient evidence
to pick one alternative over another may be due to the fact that there really is no
difference, or it may be a result of the variance in the observed outcomes being too
high to be conclusive. At this point, either additional replications may be run or
one of several variance reduction techniques might be employed (see section 9.5).
Figure 9.2(b) illustrates the case when the confidence interval is completely to the
FIGURE 9.2
Three possible
(a) [------|------] Fail to
reject H0
positions of a
confidence interval
relative to zero.
(b) [------|------] Reject H0
(c) [------|------] Reject H0
␮1 ⫺ ␮2 ⫽ 0
[------|------] denotes confidence interval (x1 ⫺ x 2) ⫾ hw
har01307_ch09_243-272.indd 246 12/17/10 12:01 PM

left of zero, leading us to reject H0. This case suggests that 1 2 0 or, equiva-
lently, 1 2. Figure 9.2(c) illustrates the case when the confidence interval is
completely to the right of zero, leading us to also reject H0. This case suggests that
1 2 0 or, equivalently, 1 2. These rules are commonly used in practice
to make statements about how the population means differ (1 2 or 1 2)
when the confidence interval does not include zero (Banks et al. 2010; Hoover
and Perry 1989).
9.3 Comparing Two Alternative System Designs

In this section, we will present two methods based on the confidence interval ap-
proach that are commonly used to compare two alternative system designs. To fa-
cilitate our understanding of the confidence interval methods, the presentation will
relate to the buffer allocation example in section 9.2, where the production control
staff has identified two candidate strategies believed to maximize the throughput
of the system. We seek to discover if the mean throughputs of the system due to
Strategy 1 and Strategy 2 are significantly different. We begin by estimating the
mean performance of the two proposed buffer allocation strategies (1 and 2) by
simulating each strategy for 16 days (24 hours per day) past the warm-up period.
The simulation experiment was replicated 10 times for each strategy. Therefore,
we obtained a sample size of 10 for each strategy (n1 n2 10). The average
hourly throughput achieved by each strategy is shown in Table 9.1.
As in Chapter 8, the methods for computing confidence intervals in this chap-
ter require that our observations be independent and normally distributed. The
10 observations in column B (Strategy 1 Throughput) of Table 9.1 are independent
TABLE 9.1 Comparison of Two Buffer Allocation Strategies
(A) (B) (C)

Strategy 1 Strategy 2
Replication Throughput x1 Throughput x2
1 54.48 56.01
2 57.36 54.08
3 54.81 52.14
4 56.20 53.49
5 54.83 55.49
6 57.69 55.00
7 58.33 54.88
8 57.19 54.47
9 56.84 54.93
10 55.29 55.84
Sample mean x̄i, for i 1, 2 56.30 54.63

Sample standard deviation si, for i 1, 2 1.37 1.17
Sample variance s2i for i 1, 2 1.89 1.36
har01307_ch09_243-272.indd 247 12/17/10 12:01 PM

because a unique segment (stream) of random numbers from the random number
generator was used for each replication. The same is true for the 10 observations
in column C (Strategy 2 Throughput). The use of random number streams is dis-
cussed in Chapters 3 and 8 and later in this chapter. At this point we are assuming
that the observations are also normally distributed. The reasonableness of assum-
ing that the output produced by our simulation models is normally distributed is
discussed at length in Chapter 8. For this data set, we should also point out that
two different sets of random numbers were used to simulate the 10 replications
of each strategy. Therefore, the observations in column B are independent of the
observations in column C. Stated another way, the two columns of observations
are not correlated. Therefore, the observations are independent within a popula-
tion (strategy) and between populations (strategies). This is an important distinc-
tion and will be employed later to help us choose between different methods for
computing the confidence intervals used to compare the two strategies.
From the observations in Table 9.1 of the throughput produced by each strat-
egy, it is not obvious which strategy yields the higher throughput. Inspection of
the summary statistics indicates that Strategy 1 produced a higher mean through-
put for the system; however, the sample variance for Strategy 1 was higher than
for Strategy 2. Recall that the variance provides a measure of the variability of the
data and is obtained by squaring the standard deviation. Equations for computing
the sample mean x̄, sample variance s2, and sample standard deviation s are given
in Chapter 8. Because of this variation, we should be careful when making con-
clusions about the population of throughput values (1 and 2) by only inspecting
the point estimates (x̄ 1 and x̄ 2). We will avoid the temptation and use the output
from the 10 replications of each simulation model along with a confidence inter-
val to make a more informed decision.
We will use an 0.05 level of significance to compare the two candidate
strategies using the following hypotheses:
H0: 1 2 0
H1: 1 2 0
where the subscripts 1 and 2 denote Strategy 1 and Strategy 2, respectively. As
stated earlier, there are two common methods for constructing a confidence inter-
val for evaluating hypotheses. The first method is referred to as the Welch confi-
dence interval (Law 2007; Miller 1986) and is a modified two-sample-t confidence
interval. The second method is the paired-t confidence interval (Miller et al.
1990). We’ve chosen to present these two methods because their statistical as-
sumptions are more easily satisfied than are the assumptions for other confidence
interval methods.
9.3.1 Welch Confidence Interval

for Comparing Two Systems
The Welch confidence interval method requires that the observations drawn from
each population (simulated system) be normally distributed and independent
har01307_ch09_243-272.indd 248 12/17/10 12:01 PM

within a population and between populations. Recall that the observations in

Table 9.1 are independent and are assumed normal. However, the Welch confi-
dence interval method does not require that the number of samples drawn from
one population (n1) equal the number of samples drawn from the other popula-
tion (n2) as we did in the buffer allocation example. Therefore, if you have more
observations for one candidate system than for the other candidate system, then
by all means use them. Additionally, this approach does not require that the two
populations have equal variances (21 22 2) as do other approaches. This is
useful because we seldom know the true value of the variance of a population.
Thus we are not required to judge the equality of the variances based on the sam-
ple variances we compute for each population (s21 and s22) before using the Welch
confidence interval method.
The Welch confidence interval for an level of significance is
P[(x̄ 1 x̄ 2) hw 1 2 (x̄ 1 x̄ 2) hw] 1
where x̄ 1 and x̄ 2 represent the sample means used to estimate the population
means 1 and 2; hw denotes the half-width of the confidence interval and is
computed by
_______
s21 __
s22
n1 n2
hw tdf,2 __
where df (degrees of freedom) is estimated by

[s21n1 s22n2]2
df ______________________________
[s1n1] (n1 1) [s22n2]2(n2 1)
2 2
and tdf,2 is a factor obtained from the Student’s t table in Appendix B based on
the value of 2 and the estimated degrees of freedom. Note that the degrees of
freedom term in the Student’s t table is an integer value. Given that the estimated
degrees of freedom will seldom be an integer value, you will have to use interpo-
lation to compute the tdf,2 value.
For the example buffer allocation problem with an 0.05 level of signifi-
cance, we use these equations and data from Table 9.1 to compute
[1.8910 1.3610]2
df ____________________________________ 17.5
[1.8910] (10 1) [1.3610]2(10 1)
2
and
___________
_____
____
hw t17.5,0.025 1.89 1.36
____ 2.106 0.325 1.20 parts per hour
10 10
where tdf,2 t17.5,0.025 2.106 is determined from Student’s t table in Appendix B
by interpolation. Now the 95 percent confidence interval is
(x̄ 1 x̄ 2) hw 1 2 (x̄ 1 x̄ 2) hw
(56.30 54.63) 1.20 1 2 (56.30 54.63) 1.20
0.47 1 2 2.87
har01307_ch09_243-272.indd 249 12/17/10 12:01 PM

With approximately 95 percent confidence, we conclude that there is a signifi-

cant difference between the mean throughputs of the two strategies because the
interval excludes zero. The confidence interval further suggests that the mean
throughput 1 of Strategy 1 is higher than the mean throughput 2 of Strategy 2
(an estimated 0.47 to 2.87 parts per hour higher).
9.3.2 Paired- t Confidence Interval

for Comparing Two Systems
Like the Welch confidence interval method, the paired-t confidence interval
method requires that the observations drawn from each population (simulated
system) be normally distributed and independent within a population. How-
ever, the paired-t confidence interval method does not require that the observa-
tions between populations be independent. This allows us to use a technique
called common random numbers to force a positive correlation between the
two populations of observations in order to reduce the half-width of our con-
fidence interval without increasing the number of replications. Recall that the
smaller the half-width, the better our estimate. The common random num-
bers technique is discussed in Section 9.5. Unlike the Welch method, the
paired-t confidence interval method does require that the number of samples
drawn from one population (n1) equal the number of samples drawn from the
other population (n2) as we did in the buffer allocation example. And like the
Welch method, the paired-t confidence interval method does not require that
the populations have equal variances (21 22 2). This is useful because
we seldom know the true value of the variance of a population. Thus we are
not required to judge the equality of the variances based on the sample vari-
ances we compute for each population (s21 and s22) before using the paired-t
confidence interval method.
Given n observations (n1 n2 n), we pair the observations from each
population (x1 j and x2 j) to define a new random variable x(12) j x1 j x2 j, for
j 1, 2, 3, . . . , n. The x1 j denotes the jth observation from the first population
sampled (the output from the jth replication of the simulation model for the first
alternative design), x2 j denotes the jth observation from the second population
sampled (the output from the jth replication of the simulation model for the
second alternative design), and x(12) j denotes the difference between the jth ob-
servations from the two populations. The point estimators for the new random
variable are
j1 x(12) j
n
Sample mean x̄ (12) _________

n
_________________

j1 [x(12) j x̄ (12)]
n 2
Sample standard deviation s(12) _________________
n1
where x̄ (12) estimates (12) and s(12) estimates (12).
har01307_ch09_243-272.indd 250 12/17/10 12:01 PM

The half-width equation for the paired-t confidence interval is

(tn 1,2)s(12)
hw ___________
__
n
where tn1,2 is a factor that can be obtained from the Student’s t table in Appendix B
based on the value of 2 and the degrees of freedom (n 1). Thus the paired-t
confidence interval for an level of significance is
P(x̄ (12) hw (12) x̄ (12) hw) 1
Notice that this is basically the same confidence interval expression presented in
Chapter 8 with x̄ (12) replacing x̄ and (12) replacing .
Let’s create Table 9.2 by restructuring Table 9.1 to conform to our new
“paired” notation and paired-t method before computing the confidence interval
necessary for testing the hypotheses
H0: 1 2 0 or using the new “paired” notation (12) 0
H1: 1 2 0 or using the new “paired” notation (12) 0
where the subscripts 1 and 2 denote Strategy 1 and Strategy 2, respectively.
The observations in Table 9.2 are identical to the observations in Table 9.1.
However, we added a fourth column. The values in the fourth column (col-
umn D) are computed by subtracting column C from column B. This fourth
column represents the 10 independent observations (n 10) of our new random
variable x(12) j.

Based on the Paired Differences
(A) (B) (C) (D)
Throughput
Strategy 1 Strategy 2 Difference (B C)
Replication (j) Throughput x1j Throughput x2j x(12) j x1j x2j
1 54.48 56.01 1.53

2 57.36 54.08 3.28
3 54.81 52.14 2.67
4 56.20 53.49 2.71
5 54.83 55.49 0.66
6 57.69 55.00 2.69
7 58.33 54.88 3.45
8 57.19 54.47 2.72
9 56.84 54.93 1.91
10 55.29 55.84 0.55
Sample mean x̄(12) 1.67

Sample standard deviation s(12) 1.85
2
Sample variance s(12) 3.42
har01307_ch09_243-272.indd 251 12/17/10 12:01 PM

Now, for an 0.05 level of significance, the confidence interval on the

new random variable x(12) j in the Throughput Difference column of Table 9.2 is
computed using our equations as follows:
j1 x(12) j
10
x̄ (12) _________ 1.67 parts per hour

10
_________________

j1 [x(12) j 1.67]
10 2
s(12) _________________ 1.85 parts per hour
10 1
)1.85 __________
(t9,0.025___ (2.262)1.85
hw __________ ___ 1.32 parts per hour
10 10
where tn1,2 t9,0.025 2.262 is determined from the Student’s t table in Appen-
dix B. The 95 percent confidence interval is
x̄ (12) hw (12) x̄ (12) hw
1.67 1.32 (12) 1.67 1.32
0.32 (12) 2.99
cant difference between the mean throughputs of the two strategies given that
the interval excludes zero. The confidence interval further suggests that the
mean throughput 1 of Strategy 1 is higher than the mean throughput 2 of
Strategy 2 (an estimated 0.35 to 2.99 parts per hour higher). This is basically the
same conclusion reached using the Welch confidence interval method presented
in section 9.3.1.
Now let’s walk back through the assumptions made when we used the paired-t
method. The main requirement is that the observations in the Throughput Dif-
ference column be independent and normally distributed. Pairing the throughput
observations for Strategy 1 with the throughput observations for Strategy 2 and
subtracting the two values formed the Throughput Difference column of Table 9.2.
The observations for Strategy 1 are independent because nonoverlapping streams
of random numbers were used to drive each replication. The observations for Strat-
egy 2 are independent for the same reason. The use of random number streams is
discussed in Chapters 3 and 8 and later in this chapter. Therefore, there is no doubt
that these observations meet the independence requirement. It then follows that the
observations under the Throughput Difference column are also statistically inde-
pendent. The assumption that has been made, without really knowing if it’s true,
is that the observations in the Throughput Difference column are normally distrib-
uted. The reasonableness of assuming that the output produced by our simulation
models is normally distributed is discussed at length in Chapter 8.
9.3.3 Welch versus the Paired-t Confidence Interval

It is difficult to say beforehand which method would produce the smaller con-
fidence interval half-width for a given comparison problem. If, however, the
har01307_ch09_243-272.indd 252 12/17/10 12:01 PM

observations between populations (simulated systems) are not independent, then

the Welch method cannot be used to compute the confidence interval. For this
case, use the paired-t method. If the observations between populations are in-
dependent, the Welch method would be used to construct the confidence inter-
val should you have an unequal number of observations from each population
(n1 n2) and you do not wish to discard any of the observations in order to pair
them up as required by the paired-t method.
9.4 Comparing More Than Two Alternative System Designs

Sometimes we use simulation to compare more than two alternative designs of
a system with respect to a given performance measure. And, as in the case of
comparing two alternative designs of a system, several statistical methods can be
used for the comparison. We will present two of the most popular methods used
in simulation. The Bonferroni approach is presented in section 9.4.1 and is use-
ful for comparing three to about five designs. A class of linear statistical models
useful for comparing any number of alternative designs of a system is presented
in section 9.4.2. A brief introduction to factorial design and optimization experi-
ments is presented in section 9.4.3.
9.4.1 The Bonferroni Approach for Comparing More Than

Two Alternative Systems
The Bonferroni approach is useful when there are more than two alternative
system designs to compare with respect to some performance measure. Given
K alternative system designs to compare, the null hypothesis H0 and alternative
hypothesis H1 become
H0: 1 2 3 . . . K for K alternative systems
H1: i i for at least one pair i i
where i and i are between 1 and K and i < i . The null hypothesis H0 states that the
means from the K populations (mean output of the K different simulation models)
are not different, and the alternative hypothesis H1 states that at least one pair of
the means are different.
The Bonferroni approach is very similar to the two confidence interval meth-
ods presented in section 9.3 in that it is based on computing confidence intervals
to determine if the true mean performance of one system (i) is significantly dif-
ferent than the true mean performance of another system (i ). In fact, either the
paired-t confidence interval or the Welch confidence interval can be used with the
Bonferroni approach. However, we will describe it in the context of using paired-t
confidence intervals, noting that the paired-t confidence interval method can be used
when the observations across populations are either independent or correlated.
The Bonferroni method is implemented by constructing a series of con-
fidence intervals to compare all system designs to each other (all pairwise
har01307_ch09_243-272.indd 253 12/17/10 12:01 PM

comparisons). The number of pairwise comparisons for K candidate designs is

computed by K(K 1)2. A paired-t confidence interval is constructed for each
pairwise comparison. For example, four candidate designs, denoted D1, D2, D3,
and D4, require the construction of six [4(4 1)2] confidence intervals to evalu-
ate the differences (D1D2), (D1D3), (D1D4), (D2D3), (D2D4), and (D3D4). The
six paired-t confidence intervals are
P(x̄ (D1D2) hw (D1D2) x̄ (D1D2) hw) 1 1
P(x̄ (D1D3) hw (D1D3) x̄ (D1D3) hw) 1 2
P(x̄ (D1D4) hw (D1D4) x̄ (D1D4) hw) 1 3
P(x̄ (D2D3) hw (D2D3) x̄ (D2D3) hw) 1 4
P(x̄ (D2D4) hw (D2D4) x̄ (D2D4) hw) 1 5
P(x̄ (D3D4) hw (D3D4) x̄ (D3D4) hw) 1 6
The rule for deciding if there is a significant difference between the true mean
performance of two system designs is the same as before. Confidence intervals
that exclude zero indicate a significant difference between the mean performance
of the two systems being compared.
In a moment, we will gain some experience using the Bonferroni ap-
proach on our example problem. However, we should discuss an important
issue about the approach first. Notice that the number of confidence intervals
quickly grows as the number of candidate designs K increases [number of con-
fidence intervals K(K 1)2]. This increases our computational workload,
but, more importantly, it has a rather dramatic effect on the overall confidence
we can place in our conclusions. Specifically, the overall confidence in the
correctness of our conclusions goes down as the number of candidate designs
increases. If we pick any one of our confidence intervals, say the sixth one, and
evaluate it separately from the other five confidence intervals, the probabil-
ity that the sixth confidence interval statement is correct is equal to (1 6).
Stated another way, we are 100(1 6) percent confident that the true but un-
known mean ((D3D4)) lies within the interval (x̄ (D3D4) hw) to (x̄ (D3D4) hw).
Although each confidence interval is computed separately, it is the simultane-
ous interpretation of all the confidence intervals that allows us to compare
the competing designs for a system. The Bonferroni inequality states that the
probability of all six confidence intervals being simultaneously correct is at
least equal to (1 i1 i). Stated more generally,
6
P(all m confidence inteval statements are correct)

(1 ) (1 i1 i)
m
K(K1)
where i1 i and is the overall level of significance and m _____
m
2
and is
the number of confidence interval statements.
If, in this example for comparing four candidate designs, we set 1 2
3 4 5 6 0.05, then the overall probability that all our conclusions
are correct is as low as (1 0.30), or 0.70. Being as low as 70 percent confident
in our conclusions leaves much to be desired. To combat this, we simply lower
the values of the individual significance levels (1 2 3 . . . m) so
har01307_ch09_243-272.indd 254 12/17/10 12:01 PM

their sum is not so large. However, this does not come without a price, as we
shall see later in our example.
One way to assign values to the individual significance levels is to first estab-
lish an overall level of significance and then divide it by the number of pairwise
comparisons. That is,

i ___________ for i 1, 2, 3, . . ., K(K 1)2
K(K 1 )2
Note, however, that it is not required that the individual significance levels be as-
signed the same value. This is useful in cases where the decision maker wants to
place different levels of significance on certain comparisons.
Practically speaking, the Bonferroni inequality limits the number of system
designs that can be reasonably compared to about five designs or less. This is
because controlling the overall significance level for the test requires the assign-
ment of small values to the individual significance levels (1 2 3 . . .
m) if more than five designs are compared. This presents a problem because the
width of a confidence interval quickly increases as the level of significance is
reduced. Recall that the width of a confidence interval provides a measure of the
accuracy of the estimate. Therefore, we pay for gains in the overall confidence of
our test by reducing the accuracy of our individual estimates (wide confidence
intervals). When accurate estimates (tight confidence intervals) are desired,
we recommend not using the Bonferroni approach when comparing more than five
system designs. For comparing more than five system designs, we recommend that
the analysis of variance technique be used in conjunction with perhaps the Fisher’s
least significant difference test. These methods are presented in section 9.4.2.
Let’s return to the buffer allocation example from the previous section and
apply the Bonferroni approach using paired-t confidence intervals. In this case,
the production control staff has devised three buffer allocation strategies to com-
pare. And, as before, we wish to determine if there are significant differences
between the throughput levels (number of parts completed per hour) achieved by
the strategies. Although we will be working with individual confidence intervals,
the hypotheses for the overall level of significance are
H0: 1 2 3
H1: 1 2 or 1 3 or 2 3
where the subscripts 1, 2, and 3 denote Strategy 1, Strategy 2, and Strategy 3,
respectively.
To evaluate these hypotheses, we estimated the performance of the three
strategies by simulating the use of each strategy for 16 days (24 hours per day)
past the warm-up period. And, as before, the simulation was replicated 10 times
for each strategy. The average hourly throughput achieved by each strategy is
shown in Table 9.3.
The evaluation of the three buffer allocation strategies (K 3) requires that
three [3(3 1)2] pairwise comparisons be made. The three pairwise comparisons
har01307_ch09_243-272.indd 255 12/17/10 12:01 PM

TABLE 9.3 Comparison of Three Buffer Allocation Strategies (K ⴝ 3) Based on Paired Differences
(A) (B) (C) (D) (E) (F) (G)

Strategy 1 Strategy 2 Strategy 3 Difference (B C) Difference (B D) Difference (C D)
Rep. Throughput Throughput Throughput Strategy 1 Strategy 2 Strategy 1 Strategy 3 Strategy 2 Strategy 3
(j) x1 j x2 j x3 j x(12) j x(13) j x(23) j
1 54.48 56.01 57.22 1.53 2.74 1.21

2 57.36 54.08 56.95 3.28 0.41 2.87
3 54.81 52.14 58.30 2.67 3.49 6.16
4 56.20 53.49 56.11 2.71 0.09 2.62
5 54.83 55.49 57.00 0.66 2.17 1.51
6 57.69 55.00 57.83 2.69 0.14 2.83
7 58.33 54.88 56.99 3.45 1.34 2.11
8 57.19 54.47 57.64 2.72 0.45 3.17
9 56.84 54.93 58.07 1.91 1.23 3.14
10 55.29 55.84 57.81 0.55 2.52 1.97
x̄(ii ), for all i and i between 1 and 3, with i < i 1.67 1.09 2.76
s(ii ), for all i and i between 1 and 3, with i < i 1.85 1.58 1.37
are shown in columns E, F, and G of Table 9.3. Also shown in Table 9.3 are the
sample means x̄(ii ) and sample standard deviations s(ii ) for each pairwise comparison.
Let’s say that we wish to use an overall significance level of 0.06 to
evaluate our hypotheses. For the individual levels of significance, let’s set 1
2 3 0.02 by using the equation
____
i __ 0.06 0.02 for i 1, 2, 3
3 3
The computation of the three paired-t confidence intervals using the method out-
lined in section 9.3.2 and data from Table 9.3 follows:
Comparing (12): 1 0.02
tr1,12 t9,0.01 2.821 from Appendix B
(t9,0.01__)s(12) __________
(2.821)1.85
hw _________
n
___
10
hw 1.65 parts per hour
The approximate 98 percent confidence interval is
x̄ (12) hw (12) x̄ (12) hw
1.67 1.65 (12) 1.67 1.65
0.02 (12) 3.32
Comparing (13): 2 0.02
tn1,22 t9,0.01 2.821 from Appendix B
(t9,0.01__)s(13) __________
(2.821)1.58
hw _________
n
___
10
hw 1.41 parts per hour
har01307_ch09_243-272.indd 256 12/17/10 12:01 PM

x̄ (13) hw (13) x̄ (13) hw

1.09 1.41 (13) 1.09 1.41
2.50 (13) 0.32
Comparing (23): The approximate 98 percent confidence interval is
3.98 (23) 1.54
Given that the confidence interval about (12) excludes zero, we conclude
that there is a significant difference in the mean throughput produced by Strate-
gies 1 (1) and 2 (2). The confidence interval further suggests that the mean
(an estimated 0.02 parts per hour to 3.32 parts per hour higher). This conclu-
sion should not be surprising because we concluded that Strategy 1 resulted in a
higher throughput than Strategy 2 earlier in sections 9.3.1 and 9.3.2. However,
notice that this confidence interval is wider than the one computed in section
9.3.2 for comparing Strategy 1 to Strategy 2 using the same data. This is because
the earlier confidence interval was based on a significance level of 0.05 and this
one is based on a significance level of 0.02. Notice that we went from using
t9,0.025 2.262 for the paired-t confidence interval in section 9.3.2 to t9,0.01
2.821 for this confidence interval, which increased the width of the interval to
the point that it is very close to including zero. This is the price that you pay,
mentioned earlier, for reducing the level of significance.
Given that the confidence interval about (13) includes zero, we conclude
that there is no significant difference in the mean throughput produced by Strate-
gies 1 (1) and 3 (3). And from the final confidence interval about (23), we
conclude that there is a significant difference in the mean throughput produced by
Strategies 2 (2) and 3 (3). This confidence interval suggests that the throughput
of Strategy 3 is higher than the throughput of Strategy 2 (an estimated 1.54 parts
per hour to 3.98 parts per hour higher).
Recall that our overall confidence for these conclusions is approximately
94 percent. Based on these results, we may be inclined to believe that Strategy 2
is the least favorable with respect to mean throughput while Strategies 1 and 3 are
the most favorable with respect to mean throughput. Additionally, the difference
in the mean throughput of Strategy 1 and Strategy 3 is not significant. Therefore,
we recommend that you implement Strategy 3 in place of your own Strategy 1
because Strategy 3 was the boss’s idea.
In this case, the statistical assumptions for using the Bonferroni approach are
the same as for the paired-t confidence interval. Because we used the Student’s t
distribution to build the confidence intervals, the observations in the Throughput
Difference columns of Table 9.3 must be independent and normally distributed.
It is reasonable to assume that these two assumptions are satisfied here for the
Bonferroni test using the same logic presented at the conclusion of section 9.3.2
for the paired-t confidence interval.
har01307_ch09_243-272.indd 257 12/17/10 12:01 PM

9.4.2 Advanced Statistical Models for Comparing

More Than Two Alternative Systems
Analysis of variance (ANOVA) in conjunction with a multiple comparison test
provides a means for comparing a much larger number of alternative system
designs than does the Welch confidence interval, paired-t confidence inter-
val, or Bonferroni approach. The major benefit that the ANOVA procedure
has over the Bonferroni approach is that the overall confidence level of the
test of hypothesis does not decrease as the number of candidate system designs
increases. There are entire textbooks devoted to ANOVA and multiple com-
parison tests used for a wide range of experimental designs. However, we will
limit our focus to using these techniques for comparing the performance of
multiple system designs. As with the Bonferroni approach, we are interested in
evaluating the hypotheses
H0: 1 2 3 . . . K for K alternative systems
H1: i i for at least one pair i i
where i and i are between 1 and K and i < i . After defining some new terminol-
ogy, we will formulate the hypotheses differently to conform to the statistical
model used in this section.
An experimental unit is the system to which treatments are applied. The
simulation model of the production system is the experimental unit for the
buffer allocation example. A treatment is a generic term for a variable of in-
terest and a factor is a category of the treatment. We will consider only the
single-factor case with K levels. Each factor level corresponds to a differ-
ent system design. For the buffer allocation example, there are three factor
levels—Strategy 1, Strategy 2, and Strategy 3. Treatments are applied to the
experimental unit by running the simulation model with a specified factor
level (strategy).
An experimental design is a plan that causes a systematic and efficient ap-
plication of treatments to an experimental unit. We will consider the completely
randomized (CR) design—the simplest experimental design. The primary as-
sumption required for the CR design is that experimental units (simulation mod-
els) are homogeneous with respect to the response (model’s output) before the
treatment is applied. For simulation experiments, this is usually the case because
a model’s logic should remain constant except to change the level of the factor
under investigation. We first specify a test of hypothesis and significance level,
say an value of 0.05, before running experiments. The null hypothesis for the
buffer allocation problem would be that the mean throughputs due to the applica-
tion of treatments (Strategies 1, 2, and 3) do not differ. The alternate hypothesis
states that the mean throughputs due to the application of treatments (Strategies 1,
2, and 3) differ among at least one pair of strategies.
We will use a balanced CR design to help us conduct this test of hypothesis.
In a balanced design, the same number of observations are collected for each fac-
tor level. Therefore, we executed 10 simulation runs to produce 10 observations
har01307_ch09_243-272.indd 258 12/17/10 12:01 PM

TABLE 9.4 Experimental Results and Summary Statistics for a Balanced

Experimental Design
Strategy 1 Strategy 2 Strategy 3

Throughput Throughput Throughput
Replication (j) (x1j) (x2j) (x3j)
1 54.48 56.01 57.22

2 57.36 54.08 56.95
3 54.81 52.14 58.30
4 56.20 53.49 56.11
5 54.83 55.49 57.00
6 57.69 55.00 57.83
7 58.33 54.88 56.99
8 57.19 54.47 57.64
9 56.84 54.93 58.07
10 55.29 55.84 57.81
Sum xi j1 xi j j1 xi j, for i 1, 2, 3

n 10
563.02 546.33 573.92

n 10
j1 i j x j1 i jx
Sample mean x̄i _______
n ______ , for i 1, 2, 3 56.30 54.63 57.39
10
of throughput for each strategy. Table 9.4 presents the experimental results and
summary statistics for this problem. The response variable x̄i j is the observed
throughput for the treatment (strategy). The subscript i refers to the factor level
(Strategy 1, 2, or 3) and j refers to an observation (output from replication j) for
that factor level. For example, the mean throughput response of the simulation
model for the seventh replication of Strategy 2 is 54.88 in Table 9.4. Parameters
for this balanced CR design are
Number of factor levels number of alternative system designs K 3
Number of observations for each factor level n 10
Total number of observations N nK (10)3 30
Inspection of the summary statistics presented in Table 9.4 indicates that
Strategy 3 produced the highest mean throughput and Strategy 2 the lowest.
Again, we should not jump to conclusions without a careful analysis of the experi-
mental data. Therefore, we will use analysis of variance (ANOVA) in conjunction
with a multiple comparison test to guide our decision.
Analysis of Variance
Analysis of variance (ANOVA) allows us to partition the total variation in the out-
put response from the simulated system into two components—variation due to the
effect of the treatments and variation due to experimental error (the inherent vari-
ability in the simulated system). For this problem case, we are interested in knowing
if the variation due to the treatment is sufficient to conclude that the performance of
one strategy is significantly different than the other with respect to mean throughput
of the system. We assume that the observations are drawn from normally distributed
har01307_ch09_243-272.indd 259 12/17/10 12:01 PM

populations and that they are independent within a strategy and between strategies.
Therefore, the variance reduction technique based on common random numbers
(CRN) presented in section 9.5 cannot be used with this method.
The fixed-effects model is the underlying linear statistical model used for the
analysis because the levels of the factor are fixed and we will consider each pos-
sible factor level. The fixed-effects model is written as

for i 1, 2, 3, . . . , K
xi j i ij
for j 1, 2, 3, . . . , n
where i is the effect of the ith treatment (ith strategy in our example) as a devia-
tion from the overall (common to all treatments) population mean and i j is the
error associated with this observation. In the context of simulation, the i j term
represents the random variation of the response xi j that occurred during the jth rep-
lication of the ith treatment. Assumptions for the fixed-effects model are that the
sum of all i equals zero and that the error terms i j are independent and normally
distributed with a mean of zero and common variance. There are methods for test-
ing the reasonableness of the normality and common variance assumptions. How-
ever, the procedure presented in this section is reported to be somewhat insensitive
to small violations of these assumptions (Miller et al. 1990). Specifically, for the
buffer allocation example, we are testing the equality of three treatment effects
(Strategies 1, 2, and 3) to determine if there are statistically significant differences
among them. Therefore, our hypotheses are written as
H0: 1 2 3 0
H1: i 0 for at least one i, for i 1, 2, 3
Basically, the previous null hypothesis that the K population means are all
equal (1 2 3 . . . K ) is replaced by the null hypothesis 1
2 3 . . . K 0 for the fixed-effects model. Likewise, the alternative
hypothesis that at least two of the population means are unequal is replaced by
i 0 for at least one i. Because only one factor is considered in this problem, a
simple one-way analysis of variance is used to determine FCALC, the test statistic
that will be used for the hypothesis test. If the computed FCALC value exceeds a
threshold value called the critical value, denoted FCRITICAL, we shall reject the null
hypothesis that states that the treatment effects do not differ and conclude that
there are statistically significant differences among the treatments (strategies in
our example problem).
To help us with the hypothesis test, let’s summarize the experimental results
shown in Table 9.4 for the example problem. The first summary statistic that we
will compute is called the sum of squares (SSi) and is calculated for the ANOVA
for each factor level (Strategies 1, 2, and 3 in this case). In a balanced design
where the number of observations n for each factor level is a constant, the sum of
squares is calculated using the formula
( j 1 xi j )2
n
SSi ( j 1 x2i j ) ________ for i 1, 2, 3, . . . , K

n
n
har01307_ch09_243-272.indd 260 12/17/10 12:01 PM

For this example, the sums of squares are
( j 1 x1 j )
10
SS1 ( j 1 x21 j ) _________

10
2
10
(563.02)2
SS1 [(54.48)2 (57.36)2 . . . (55.29)2] ________ 16.98
10
SS2 12.23
SS3 3.90
The grand total of the N observations (N nK) collected from the output response
of the simulated system is computed by
Grand total x . . i1 j1 xij i1 xi
K n K
The overall mean of the N observations collected from the output response of the
simulated system is computed by
i1 j1 xi j x . .
K n
Overall mean x̄ . . __________ ___

N N
Using the data in Table 9.4 for the buffer allocation example, these statistics are
Grand total x . . i1 xi 563.02 546.33 573.92 1,683.27
3
1,683.27
x . . ________
Overall mean x̄ . . ___ 56.11
N 30
Our analysis is simplified because a balanced design was used (equal observa-
tions for each factor level). We are now ready to define the computational formulas
for the ANOVA table elements (for a balanced design) needed to conduct the hy-
pothesis test. As we do, we will construct the ANOVA table for the buffer allocation
example. The computational formulas for the ANOVA table elements are
Degrees of freedom total (corrected) df(total corrected) N 1
30 1 29
Degrees of freedom treatment df(treatment) K 1 3 1 2
Degrees of freedom error df(error) N K 30 3 27
and
Sum of squares error SSE i1 SSi 16.98 12.23 3.90 33.11
K
x2
Sum of squares treatment SST _1n ( i1 x2i ) __. .
K
K
1 ((563.02)2 (546.33)2 (573.92)2) __________
SST ___
10
(1,683.27)2
3
38.62
Sum of squares total (corrrected) SSTC SST SSE
38.62 33.11 71.73
and
SST
Mean square treatment MST ___________ 38.62 19.31
_____
df(treatment) 2
har01307_ch09_243-272.indd 261 12/17/10 12:01 PM

TABLE 9.5 Analysis of Variance Table
Source of Degrees of Sum of Mean

Variation Freedom Squares Square FCALC
Total (corrected) N 1 29 SSTC 71.73

Treatment (strategies) K12 SST 38.62 MST 19.31 15.70
Error N K 27 SSE 33.11 MSE 1.23
SSF _____
Mean square error MSE ________ 33.11 1.23
df(error) 27
and finally
Calculated F statistic FCALC _____ 19.31 15.70
MST _____
MSE 1.23
Table 9.5 presents the ANOVA table for this problem. We will compare the value
of FCALC with a value from the F table in Appendix C to determine whether to reject
or fail to reject the null hypothesis H0: 1 2 3 0. The values obtained from
the F table in Appendix C are referred to as critical values and are determined by
F(df(treatment), df(error); ). For this problem, F(2,27; 0.05) 3.35 FCRITICAL, using a signifi-
cance level () of 0.05. Therefore, we will reject H0 since FCALC FCRITICAL at the
0.05 level of significance. If we believe the data in Table 9.4 satisfy the assump-
tions of the fixed-effects model, then we would conclude that the buffer allocation
strategy (treatment) significantly affects the mean throughput of the system. We now
have evidence that at least one strategy produces better results than the other strate-
gies. Next, a multiple comparison test will be conducted to determine which strategy
(or strategies) causes the significance.
Multiple Comparison Test

Our final task is to conduct a multiple comparison test. The hypothesis test sug-
gested that not all strategies are the same with respect to throughput, but it did
not identify which strategies performed differently. We will use Fisher’s least
significant difference (LSD) test to identify which strategies performed differ-
ently. It is generally recommended to conduct a hypothesis test prior to the LSD
test to determine if one or more pairs of treatments are significantly different.
If the hypothesis test failed to reject the null hypothesis, suggesting that all i
were the same, then the LSD test would not be performed. Likewise, if we re-
ject the null hypothesis, we should then perform the LSD test. Because we first
performed a hypothesis test, the subsequent LSD test is often called a protected
LSD test.
The LSD test requires the calculation of a test statistic used to evalu-
ate all pairwise comparisons of the sample mean from each population (x̄ 1, x̄ 2,
x̄ 3, . . . , x̄ K). In our example buffer allocation problem, we are dealing with the
sample mean throughput computed from the output of our simulation models for
the three strategies (x̄ 1, x̄ 2, x̄ 3). Therefore, we will make three pairwise compari-
sons of the sample means for our example, recalling that the number of pairwise
har01307_ch09_243-272.indd 262 12/17/10 12:01 PM

comparisons for K candidate designs is computed by K(K 1)2. The LSD test
statistic is calculated as _______

2(MSE)
LSD () t(df (error),2) _______
n
The decision rule states that if the difference in the sample mean response values
exceeds the LSD test statistic, then the population mean response values are sig-
nificantly different at a given level of significance. Mathematically, the decision
rule is written as
If xi xi LSD (), then i and i are significantly different at the level of
significance.
For this problem, the LSD test statistic is determined at the 0.05 level of
significance:
_______ _______

2(MSE) 2(1.23)
LSD(0.05) t27,0.025 _______
n 2.052 _______ 1.02
10
Table 9.6 presents the results of the three pairwise comparisons for the LSD analy-
sis. With 95 percent confidence, we conclude that each pair of means is different
(1 2, 1 3, and 2 3). We may be inclined to believe that the best strategy
is Strategy 3, the second best strategy is Strategy 1, and the worst strategy is Strategy 2.
Recall that the Bonferroni approach in section 9.4.1 did not detect a signifi-
cant difference between Strategy 1 (1) and Strategy 3 (3). One possible expla-
nation is that the LSD test is considered to be more liberal in that it will indicate
a difference before the more conservative Bonferroni approach. Perhaps if the
paired-t confidence intervals had been used in conjunction with common random
numbers (which is perfectly acceptable because the paired-t method does not re-
quire that observations be independent between populations), then the Bonfer-
roni approach would have also indicated a difference. We are not suggesting here
that the Bonferroni approach is in error (or that the LSD test is in error). It could
be that there really is no difference between the performances of Strategy 1 and
Strategy 3 or that we have not collected enough observations to be conclusive.
There are several multiple comparison tests from which to choose. Other
tests include Tukey’s honestly significant difference (HSD) test, Bayes LSD
(BLSD) test, and a test by Scheffe. The LSD and BLSD tests are considered to be
TABLE 9.6 LSD Analysis
Strategy 2 Strategy 1
x̄2 54.63 x̄1 56.30
Strategy 3 |x̄2 x̄3| 2.76 |x̄1 x̄3| 1.09

x̄3 57.39 Significant Significant
(2.76 > 1.02) (1.09 1.02)
|x̄1 x̄2| 1.67
Strategy 1
Significant
x̄1 56.30
(1.67 1.02)
har01307_ch09_243-272.indd 263 12/17/10 12:01 PM

liberal in that they will indicate a difference between i and i before the more
conservative Scheffe test. A book by Petersen (1985) provides more information
on multiple comparison tests.
9.4.3 Factorial Design and Optimization

In simulation experiments, we are sometimes interested in finding out how differ-
ent decision variable settings impact the response of the system rather than simply
comparing one candidate system to another. For example, we may want to measure
how the mean time a customer waits in a bank changes as the number of tellers (the
decision variable) is increased from 1 through 10. There are often many decision
variables of interest for complex systems. And rather than run hundreds of experi-
ments for every possible variable setting, experimental design techniques can be
used as a shortcut for finding those decision variables of greatest significance (the
variables that significantly influence the output of the simulation model). Using
experimental design terminology, decision variables are referred to as factors and
the output measures are referred to as responses (Figure 9.3). Once the response of
interest has been identified and the factors that are suspected of having an influence
on this response defined, we can use a factorial design method that prescribes how
many runs to make and what level or value to use for each factor.
The natural inclination when experimenting with multiple factors is to test
the impact that each individual factor has on system response. This is a simple and
straightforward approach, but it gives the experimenter no knowledge of how fac-
tors interact with each other. It should be obvious that experimenting with two or
more factors together can affect system response differently than experimenting
with only one factor at a time and keeping all other factors the same.
One type of experiment that looks at the combined effect of multiple factors
on system response is referred to as a two-level, full-factorial design. In this type
of experiment, we simply define a high and a low setting for each factor and, since
it is a full-factorial experiment, we try every combination of factor settings. This
means that if there are five factors and we are testing two different levels for each
factor, we would test each of the 25 32 possible combinations of high and low
factor levels. For factors that have no range of values from which a high and a low
can be chosen, the high and low levels are arbitrarily selected. For example, if one
of the factors being investigated is an operating policy (like first come, first served
or last come, last served), we arbitrarily select one of the alternative policies as the
high-level setting and a different one as the low-level setting.
FIGURE 9.3
Relationship between Factors (X1, X2, . . . , Xn) Simulation Output responses
factors (decision
variables) and output model
responses.
har01307_ch09_243-272.indd 264 12/17/10 12:01 PM

For experiments in which a large number of factors are considered, a two-

level, full-factorial design would result in an extremely large number of combi-
nations to test. In this type of situation, a fractional-factorial design is used to
strategically select a subset of combinations to test in order to “screen out” factors
with little or no impact on system performance. With the remaining reduced num-
ber of factors, more detailed experimentation such as a full-factorial experiment
can be conducted in a more manageable fashion.
After fractional-factorial experiments and even two-level, full-factorial ex-
periments have been performed to identify the most significant factor level com-
binations, it is often desirable to conduct more detailed experiments, perhaps over
the entire range of values, for those factors that have been identified as being the
most significant. This provides more precise information for making decisions
regarding the best, or optimal, factor values or variable settings for the system.
For a more detailed treatment of factorial design in simulation experimentation,
see Law (2007).
There are excellent software tools available for assisting in statistical analysis. To
facilitate the use of these tools, several discrete-event simulation software packages
export output reports in formats readable by statistical software. For example, Pro-
Model exports files readable by Excel and Minitab. SAS and MegaStat are examples
of other software tools useful for the analysis and interpretation of simulation output.
At this writing, Wikipedia lists over 50 proprietary statistical software packages.
In many cases, the number of factors of interest prohibits the use of even
fractional-factorial designs because of the many combinations to test. If this is
the case and you are seeking the best, or optimal, factor values for a system, an
alternative is to employ an optimization technique to search for the best combina-
tion of values. Several optimization techniques are useful for searching for the
combination that produces the most desirable response from the simulation model
without evaluating all possible combinations. This is the subject of simulation
optimization and is discussed in Chapter 10.
9.5 Variance Reduction Techniques

One luxury afforded to model builders is that the variance of a performance mea-
sure computed from the output of simulations can be reduced. This is a luxury
because reducing the variance allows us to estimate the mean value of a random
variable within a desired level of precision and confidence with fewer replications
(independent observations). The reduction in the required number of replications
is achieved by controlling how random numbers are used to “drive” the events in
the simulation model. These time-saving techniques are called variance reduction
techniques. The use of common random numbers (CRN) is perhaps one of the
most popular variance reduction techniques. This section provides an introduction
to the CRN technique, presents an example application of CRN, and discusses
how CRN works. For additional details about CRN and a review of other variance
reduction techniques, see Law (2007).
har01307_ch09_243-272.indd 265 12/17/10 12:01 PM

9.5.1 Common Random Numbers

The common random numbers (CRN) technique was invented for comparing alter-
native system designs. Recall the proposed buffer allocation strategies for the pro-
duction system presented in section 9.2. The objective was to decide which strategy
yielded the highest throughput for the production system. The CRN technique was
not used to compare the performance of the strategies using the paired-t confidence
interval method in section 9.3.2. However, it would have been a good idea because
the CRN technique provides a means for comparing alternative system designs under
more equal experimental conditions. This is helpful in ensuring that the observed dif-
ferences in the performance of two system designs are due to the differences in the
designs and not to differences in experimental conditions. The goal is to evaluate
each system under the exact same circumstances to ensure a fair comparison.
Suppose a system is simulated to measure the mean time that entities wait in a
queue for service under different service policies. The mean time between arrivals
of entities to the system is exponentially distributed with a mean of 5.5 minutes.
The exponentially distributed variates are generated using a stream of numbers
that are uniformly distributed between zero and 1, having been produced by the
random number generator. (See Chapter 3 for a discussion on generating random
numbers and random variates.) If a particular segment of the random number
stream resulted in several small time values being drawn from the exponential
distribution, then entities would arrive to the system faster. This would place a
heavier workload on the workstation servicing the entities, which would tend
to increase how long entities wait for service. Therefore, the simulation of each
policy should be driven by the same stream of random numbers to ensure that the
differences in the mean waiting times are due only to differences in the policies
and not because some policies were simulated with a stream of random numbers
that produced more extreme conditions.
The goal is to use the exact random number from the stream for the exact
purpose in each simulated system. To help achieve this goal, the random number
stream can be seeded at the beginning of each independent replication to keep it
synchronized across simulations of each system. For example, in Figure 9.4, the
first replication starts with a seed value of 9, the second replication starts with a
seed value of 5, the third with 3, and so on. If the same seed values for each replica-
tion are used to simulate each alternative system, then the same stream of random
numbers will drive each of the systems. This seems simple enough. However, care
has to be taken not to pick a seed value that places us in a location on the stream
that has already been used to drive the simulation in a previous replication. If this
were to happen, the results from replicating the simulation of a system would not
be independent because segments of the random number stream would have been
shared between replications, and this cannot be tolerated. Therefore, some simula-
tion software provides a CRN option that, when selected, automatically assigns
seed values to each replication to minimize the likelihood of this happening.
A common practice that helps to keep random numbers synchronized across
systems is to assign a different random number stream to each stochastic element
har01307_ch09_243-272.indd 266 12/17/10 12:01 PM

FIGURE 9.4 Stream

Unique seed value Rep.1
assigned for each Seed 9
replication. .83
.
.
.
.12
Rep. 2
Seed 5
.93
.
.
.
.79
Rep. 3
Seed 3
.28
.
.
.
Distributions for arrival times, service times, and so on
Simulation models for each alternative
in the model. For this reason, most simulation software provides several unique
streams of uniformly distributed random numbers to drive the simulation. This
concept is illustrated in Figure 9.5 in that separate random number streams are
used to generate service times at each of the four machines. If an alternative de-
sign for the system was the addition of a fifth machine, the effects of adding the
fifth machine could be measured while holding the behavior of the original four
machines constant.
In ProModel, up to 100 streams (1 through 100) are available to assign to any
random variable specified in the model. Each stream can have one of 100 differ-
ent initial seed values assigned to it. Each seed number starts generating random
numbers at an offset of 100,000 from the previous seed number (seed 1 gener-
ates 100,000 random numbers before it catches up to the starting point in the
cycle of seed 2). For most simulations, this ensures that streams do not overlap.
If you do not specify an initial seed value for a stream that is used, ProModel will
use the same seed number as the stream number (stream 3 uses the third seed).
har01307_ch09_243-272.indd 267 12/17/10 12:01 PM

FIGURE 9.5 Stream 1 Stream 2 Stream 5 Stream 7

Unique random number
stream assigned to each Rep. 1 Rep. 1 Rep. 1 Rep. 1
stochastic element in Seed 99 Seed 51 Seed 89 Seed 67
system. .83 .27 .56 .25
. . . .
. . . .
. . . .
.12 .19 .71 .99
Rep. 2 Rep. 2 Rep. 2 Rep. 2

Seed 75 Seed 33 Seed 7 Seed 23
.93 .87 .45 .69
. . . .
. . . .
. . . .
.79 .35 .74 .42
Rep. 3 Rep. 3 Rep. 3 Rep. 3

Seed 3 Seed 79 Seed 49 Seed 13
.28 .21 .53 .82
. . . .
. . . .
. . . .
Service time Service time Service time

distribution distribution distribution
Machine Machine Machine Machine

1 2 3 4
A detailed explanation of how random number generators work and how they
produce unique streams of random numbers is provided in Chapter 3.
Complete synchronization of the random numbers across different models is
sometimes difficult to achieve. Therefore, we often settle for partial synchroni-
zation. At the very least, it is a good idea to set up two streams with one stream
of random numbers used to generate an entity’s arrival pattern and the other
stream of random numbers used to generate all other activities in the model.
That way, activities added to the model will not inadvertently alter the arrival
pattern because they do not affect the sample values generated from the arrival
distribution.
9.5.2 Example Use of Common Random Numbers

In this section the buffer allocation problem of section 9.2 will be simulated using
CRN. Recall that the amount of time to process an entity at each machine is a
random variable. Therefore, a unique random number stream will be used to draw
samples from the processing time distributions for each machine. There is no
need to generate an arrival pattern for parts (entities) because it is assumed that a
part (entity) is always available for processing at the first machine. Therefore, the
har01307_ch09_243-272.indd 268 12/17/10 12:01 PM

assignment of random number streams is similar to that depicted in Figure 9.5.

Next, we will use the paired-t confidence interval to compare the two competing
buffer allocation strategies identified by the production control staff. Also, note
that the common random numbers technique can be used with the Bonferroni
approach when paired-t confidence intervals are used.
There are two buffer allocation strategies, and the objective is to determine
if one strategy results in a significantly different average throughput (number of
parts completed per hour). Specifically, we will test at an 0.05 level of signifi-
cance the following hypotheses:
H0: 1 2 0 or using the paired notation (12) 0
H1: 1 2 0 or using the paired notation (12) 0
where the subscripts 1 and 2 denote Strategy 1 and Strategy 2, respectively. As
before, each strategy is evaluated by simulating the system for 16 days (24 hours
per day) past the warm-up period. For each strategy, the experiment is replicated
10 times. The only difference is that we have assigned individual random number
streams to each process and have selected seed values for each replication. This
way, both buffer allocation strategies will be simulated using identical streams
of random numbers. The average hourly throughput achieved by each strategy is
paired by replication as shown in Table 9.7.
Using the equations of section 9.3.2 and an 0.05 level of significance, a
paired-t confidence interval using the data from Table 9.7 can be constructed:
x̄ (12) 2.67 parts per hour
s(12) 1.16 parts per hour

Using Common Random Numbers
(A) (B) (C) (D)

Throughput
Strategy 1 Strategy 2 Difference (B C)
Replication (j) Throughput x1 j Throughput x2 j x(12) j x1 j x2 j
1 79.05 75.09 3.96

2 54.96 51.09 3.87
3 51.23 49.09 2.14
4 88.74 88.01 0.73
5 56.43 53.34 3.09
6 70.42 67.54 2.88
7 35.71 34.87 0.84
8 58.12 54.24 3.88
9 57.77 55.03 2.74
10 45.08 42.55 2.53
Sample mean x̄(12) 2.67

Sample standard deviation s(12) 1.16
2
Sample variance s(12) 1.35
har01307_ch09_243-272.indd 269 12/17/10 12:01 PM

and
)s(12) ___________
(t9,0.025__ (2.262) 1.16
hw __________ ___ 0.83 parts per hour
n 10
where tn1,2 t9,0.0025 2.262 is determined from the Student’s t table in

Appendix B. The 95 percent confidence interval is
x̄ (12) hw (12) x̄ (12) hw
2.67 0.83 (12) 2.67 0.83
1.84 (12) 3.50
cant difference between the mean throughputs of the two strategies because the
interval excludes zero. The confidence interval further suggests that the mean
(an estimated 1.84 to 3.50 parts per hour higher).
The interesting point here is that the half-width of the confidence interval
computed from the CRN observations is considerably shorter than the half-width
computed in section 9.3.2 without using CRN. In fact, the half-width using CRN
is approximately 37 percent shorter. This is due to the reduction in the sample
standard deviation s(12). Thus we have a more precise estimate of the true mean
difference without making additional replications. This is the benefit of using
variance reduction techniques.
9.5.3 Why Common Random Numbers Work

Using CRN does not actually reduce the variance of the output from the simula-
tion model. It is the variance of the observations in the Throughput Difference
column (column D) in Table 9.7 that is reduced. This happened because the ob-
servations in the Strategy 1 column are positively correlated with the observations
in the Strategy 2 column. This resulted from “driving” the two simulated systems
with exactly the same (or as close as possible) stream of random numbers. The
effect of the positive correlation is that the variance of the observations in the
Throughput Difference column will be reduced.
Although the observations between the strategy columns are correlated, the
observations down a particular strategy’s column are independent. Therefore, the
observations in the Throughput Difference column are also independent. This is
because each replication is based on a different segment of the random number
stream. Thus the independence assumption for the paired-t confidence interval still
holds. Note, however, that you cannot use data produced by using CRN to calculate
the Welch confidence interval or to conduct an analysis of variance. These proce-
dures require that the observations between populations (the strategy columns in
this case) be independent, which they are not when the CRN technique is used.
As the old saying goes, “there is no such thing as a free lunch,” and there is a hitch
with using CRN. Sometimes the use of CRN can produce the opposite effect and in-
crease the sample standard deviation of the observations in the Throughput Difference
har01307_ch09_243-272.indd 270 12/17/10 12:01 PM

column. Without working through the mathematics, know that this occurs when a
negative correlation is created between the observations from each system instead
of a positive correlation. Unfortunately, there is really no way of knowing before-
hand if this will happen. However, the likelihood of realizing a negative correlation in
practice is low, so the ticket to success lies in your ability to synchronize the random
numbers. If good synchronization is achieved, then the desired result of reducing the
standard deviation of the observations in the Difference column will likely be realized.
9.6 Summary
An important point to make here is that simulation, by itself, does not solve a
problem. Simulation merely provides a means to evaluate proposed solutions by
estimating how they behave. The user of the simulation model has the responsibil-
ity to generate candidate solutions either manually or by use of automatic opti-
mization techniques and to correctly measure the utility of the solutions based on
the output from the simulation. This chapter presented several statistical methods
for comparing the output produced by simulation models representing candidate
solutions or designs.
When comparing two candidate system designs, we recommend using either
the Welch confidence interval method or the paired-t confidence interval. Also,
a variance reduction technique based on common random numbers can be used
in conjunction with the paired-t confidence interval to improve the precision of
the confidence interval. When comparing between three and five candidate sys-
tem designs, the Bonferroni approach is useful. For more than five designs, the
ANOVA procedure in conjunction with Fisher’s least significance difference test
is a good choice, assuming that the population variances are approximately equal.
Additional methods useful for comparing the output produced by different simu-
lation models can be found in Goldsman and Nelson (1998).

1. The following simulation output was generated to compare four
candidate designs of a system.
a. Use the paired-t confidence interval method to compare Design 1 with
Design 3 using a 0.05 level of significance. What is your conclusion?
What statistical assumptions did you make to use the paired-t
confidence interval method?
b. Use the Bonferroni approach with Welch confidence intervals to
compare all four designs using a 0.06 overall level of significance.
What are your conclusions? What statistical assumptions did you
make to use the Bonferroni approach?
c. Use a one-way analysis of variance (ANOVA) with 0.05 to
determine if there is a significant difference between the designs.
What is your conclusion? What statistical assumptions did you make
to use the one-way ANOVA?
har01307_ch09_243-272.indd 271 12/17/10 12:01 PM

Waiting Time in System
Replication Design 1 Design 2 Design 3 Design 4
1 53.9872 58.1365 58.5438 60.1208

2 58.4636 57.6060 57.3973 59.6515
3 55.5300 58.5968 57.1040 60.5279
4 56.3602 55.9631 58.7105 58.1981
5 53.8864 58.3555 58.0406 60.3144
6 57.2620 57.0748 56.9654 59.1815
7 56.9196 56.0899 57.2882 58.3103
8 55.7004 59.8942 57.3548 61.6756
9 55.3685 57.5491 58.2188 59.6011
10 56.9589 58.0945 59.5975 60.0836
11 55.0892 59.2632 60.5354 61.1175
12 55.4580 57.4509 57.9982 59.5142
2. Why is the Bonferroni approach to be avoided when comparing more

than about five alternative designs of a system?
3. What is a Type I error in hypothesis testing?
4. What is the relationship between a Type I and a Type II error?
5. How do common random numbers reduce variation when comparing
two models?
6. Analysis of variance (ANOVA) allows us to partition the total variation in
the output from a simulation model into two components. What are they?
7. Why can common random numbers not be used if the Welch confidence
interval method is used?
References
Bateman, Robert E.; Royce O. Bowden; Thomas J. Gogg; Charles R. Harrell; and Jack R. A.
Mott. System Improvement Using Simulation. Orem, UT: PROMODEL Corp., 1997.
Goldsman, David, and Berry L. Nelson. “Comparing Systems via Simulation.” Chapter 8
in Handbook of Simulation. New York: John Wiley and Sons, 1998.
Hines, William W., and Douglas C. Montgomery. Probability and Statistics in Engineering
and Management Science. New York: John Wiley & Sons, 1990.
Hoover, Stewart V., and Ronald F. Perry. Simulation: A Problem-Solving Approach. Read-
ing, MA: Addison-Wesley, 1989.
Law, Averill M. Simulation Modeling and Analysis. New York: McGraw-Hill, 2007.
Miller, Irwin R.; John E. Freund; and Richard Johnson. Probability and Statistics for Engi-
neers. Englewood Cliffs, NJ: Prentice Hall, 1990.
Miller, Rupert G. Beyond ANOVA, Basics of Applied Statistics, New York: Wiley, 1986.
Montgomery, Douglas C. Design and Analysis of Experiments. 7th ed. New York: John
Wiley & Sons, 2008.
Petersen, Roger G. Design and Analysis of Experiments. New York: Marcel Dekker, 1985.
har01307_ch09_243-272.indd 272 12/17/10 12:01 PM

C H A P T E R
10 SIMULATION
OPTIMIZATION
“Man is a goal seeking animal. His life only has meaning if he is reaching out
and striving for his goals.”
—Aristotle
10.1 Introduction
Simulation models of systems are built for many reasons. Some models are built
to gain a better understanding of a system, to forecast the output of a system, or to
compare one system to another. If the reason for building simulation models is to
find answers to questions like “What are the optimal settings for to mini-
mize (or maximize) ?” then optimization is the appropriate technology to
combine with simulation. Optimization is the process of trying different combina-
tions of values for the variables that can be controlled to seek the combination of
values that provides the most desirable output from the simulation model.
For convenience, let us think of the simulation model as a black box that
imitates the actual system. When inputs are presented to the black box, it produces
output that estimates how the actual system responds. In our question, the first
blank represents the inputs to the simulation model that are controllable by the
decision maker. These inputs are often called decision variables or factors. The
second blank represents the performance measures of interest that are computed
from the stochastic output of the simulation model when the decision variables are
set to specific values (Figure 10.1). In the question, “What is the optimal number
of material handling devices needed to minimize the time that workstations are
starved for material?” the decision variable is the number of material handling
devices and the performance measure computed from the output of the simulation
model is the amount of time that workstations are starved. The objective, then, is
to seek the optimal value for each decision variable that minimizes, or maximizes,
the expected value of the performance measure(s) of interest. The performance
273
har01307_ch10_273-298.indd 273 12/17/10 12:02 PM

FIGURE 10.1
Optimization Decision variables (X1, X2, . . . , Xn)
Relationship between Simulation
optimization algorithm algorithm model
and simulation model.
Output responses
measure is traditionally called the objective function. Note that the expected value
of the objective function is estimated by averaging the model’s output over mul-
tiple replications or batch intervals. The simulation optimization problem is more
formally stated as
Min or Max E [ f (X1, X2, . . . , Xn)]
Subject to
Lower Boundi ⱕ Xi ⱕ Upper Boundi for i ⫽ 1, 2, . . . , n
where E [ f (X1, X2, . . . , Xn)] denotes the expected value of the objective function,
which is estimated.
The search for the optimal solution can be conducted manually or automated
with algorithms specifically designed to seek the optimal solution without evalu-
ating all possible solutions. Interfacing optimization algorithms that can automati-
cally generate solutions and evaluate them in simulation models is a worthwhile
endeavor because
• It automates part of the analysis process, saving the analyst time.
• A logical method is used to efficiently explore the realm of possible
solutions, seeking the best.
• The method often finds several exemplary solutions for the analyst to
consider.
The latter is particularly important because, within the list of optimized solutions,
there may be solutions that the decision maker may have otherwise overlooked.
In 1995, PROMODEL Corporation and Decision Science, Incorporated, de-
veloped SimRunner based on the research of Bowden (1992) on the use of modern
optimization algorithms for simulation-based optimization and machine learning.
SimRunner helps those wishing to use advanced optimization concepts to seek
better solutions from their simulation models. SimRunner uses an optimization
method based on evolutionary algorithms. It is the first widely used, commer-
cially available simulation optimization package designed for major simulation
packages (ProModel, MedModel, ServiceModel, and ProcessModel). Although
SimRunner is relatively easy to use, it can be more effectively used with a basic
understanding of how it seeks optimal solutions to a problem. Therefore, the pur-
pose of this chapter is fourfold:
• To provide an introduction to simulation optimization, focusing on
the latest developments in integrating simulation and a class of direct
optimization techniques called evolutionary algorithms.
har01307_ch10_273-298.indd 274 12/17/10 12:02 PM

Chapter 10 Simulation Optimization 275
• To give the reader an appreciation of the advantages and disadvantages of

using evolutionary algorithms for optimizing simulated systems.
• To discuss tactical issues involved in the use of evolutionary algorithms
for simulation optimization.
• To present examples highlighting how simulation optimization can help
analysts in their quest to identify optimal solutions.
There are many approaches to optimizing simulated systems other than those
based on evolutionary algorithms. For a more complete review of other ap-
proaches to simulation optimization, see Azadivar (1992), Jacobson and Schruben
(1989), and Safizadeh (1990).
10.2 In Search of the Optimum

Finding the so-called optimal solution is not an easy task. In fact, it is somewhat
like finding a needle in a haystack. The surest way to find the optimal solution is
to follow these steps (Akbay 1996):
Step 1. Identify all possible decision variables that affect the output of the
system.
Step 2. Based on the possible values of each decision variable, identify all
possible solutions.
Step 3. Evaluate each of these solutions accurately.
Step 4. Compare each solution fairly.
Step 5. Record the best answer.
If the output from the simulation model for all possible values of the decision
variables is recorded and plotted, the resulting plot would be called the response
surface. For problems involving only one or two decision variables, the optimal
solution can be identified quickly on the response surface. For example, the op-
timal solution would be located at the highest peak on the response surface for a
maximization problem (Figure 10.2). However, it becomes difficult to view the
response surface when there are more than two decision variables. Additionally,
in many cases, there are simply too many solutions to evaluate (search through) in
a reasonable amount of time. Therefore, a balance must be struck between finding
the optimal solution and the amount of time to allocate to the search.
Fortunately, researchers have developed several optimization methods that
can find the optimal solution for well-posed problems quickly without complete
enumeration of all possible alternatives. Example techniques include the Newton-
Raphson method and linear programming. Unfortunately, most realistic problems
are not well posed and do not lend themselves to being solved by these optimiza-
tion methods. For example, the response surfaces produced by stochastic simu-
lation models can be highly nonlinear, multimodal, and noisy; can contain both
discrete and continuous decision variables; and may not provide independent and
identically distributed observations. However, many good heuristic optimization
techniques relax the requirement that problems be well posed.
har01307_ch10_273-298.indd 275 12/17/10 12:02 PM

FIGURE 10.2
SimRunner plots the
output responses
generated by a
ProModel simulation
model as it seeks the
optimal solution,
which occurred at the
highest peak.
Heuristic techniques consistently provide good, or near optimal, solutions

within a reasonable amount of search time. In fact, the use of a heuristic tech-
nique does not preclude the chance of finding the optimal solution. It may find the
optimal solution, but there is no guarantee that it always will do so. For example,
the Nelder Mead Simplex Search technique (Nelder and Mead 1965) has been
extensively, and successfully, used for solving simulation optimization problems
(Barton and Ivey 1996). This technique often finds the optimal solution even
though there is no mathematical proof that it will converge to the optimal solution
for these types of problems.
10.3 Combining Direct Search Techniques with Simulation

Direct search techniques are a class of search techniques designed to seek optimal
values for the decision variables for a system so as to minimize, or maximize, the
output measure of interest from the system. These techniques work directly with
the output generated from the system of interest and require no additional infor-
mation about the function that generates the output. They are ideally suited for
optimization tasks when mathematical models do not exist from which gradient
information can be computed to guide the search for the optimal solution or if the
cost of estimating the gradient is computationally prohibitive. Such is often the
case when dealing with stochastic simulation models.
Early on, researchers realized the benefits of combining simulation and direct
search techniques. For example, Pegden and Gately (1977) developed an optimi-
zation module for the GASP IV simulation software package. Pegden and Gately
(1980) later developed another optimization module for use with the SLAM simu-
lation software package. Their optimization packages were based on a variant
of a direct search method developed by Hooke and Jeeves (1961). After solving
har01307_ch10_273-298.indd 276 12/17/10 12:02 PM

several problems, Pegden and Gately concluded that their packages extended the
capabilities of the simulation language by “providing for automatic optimization
of decision.”
The direct search algorithms available today for simulation optimization are
much better than those available in the late 1970s. Using these newer algorithms,
the SimRunner simulation optimization tool was developed in 1995. Following
SimRunner, two other simulation software vendors soon added an optimization
feature to their products. These products are OptQuest, which was introduced in
1996 to be used with simulation models built with Micro Saint software, and WIT-
NESS Optimizer, which was introduced in 1997 to be used with simulation mod-
els built with Witness software. The optimization module in OptQuest is based
on scatter search, which has links to Tabu Search and the popular evolutionary
algorithm called the Genetic Algorithm (Glover 1994; Glover et al. 1996). In ad-
dition to SimRunner, ProModel also works with OptQuest. WITNESS Optimizer
is based on a search algorithm called Simulated Annealing (Markt and Mayer
1997). Today most major simulation software packages include an optimization
feature.
SimRunner has an optimization module and a module for determining the
required sample size (replications) and a model’s warm-up period (in the case
of a steady-state analysis). The optimization module can optimize integer and
real decision variables. The design of the optimization module in SimRunner was
influenced by optima-seeking techniques such as Tabu Search (Glover 1990) and
evolutionary algorithms (Fogel 1992; Goldberg 1989; Schwefel 1981), though it
most closely resembles an evolutionary algorithm (SimRunner 1996b).
10.4 Evolutionary Algorithms

Evolutionary algorithms (EAs) are a class of direct search techniques that are
based on concepts from the theory of evolution. They mimic the underlying evo-
lutionary process in that entities adapt to their environment in order to survive.
EAs manipulate a population of solutions to a problem in such a way that poor
solutions fade away and good solutions continually evolve in their search for the
optimum (Bowden and Bullington 1996). Search techniques based on this con-
cept have proved robust in that they have been successfully used to solve a wide
variety of difficult problems (Biethahn and Nissen 1994; Bowden and Bullington
1995; Usher and Bowden 1996).
Evolutionary algorithms differ from traditional nonlinear optimization tech-
niques in many ways. The most significant difference is that they search the re-
sponse surface using a population of solutions as opposed to a single solution.
This allows an EA to collect information about the response surface from many
different points simultaneously. The EA uses the information reported back from
multiple locations on the response surface, as opposed to a single point, to guide
the search for the optimal solution. This population-based approach increases
the chance of finding the global optimal solution (Biethahn and Nissen 1994).
har01307_ch10_273-298.indd 277 12/17/10 12:02 PM

Research has been conducted to demonstrate the convergence characteristics of

EAs and to develop mathematical proofs for global convergence (for example, see
Bäck and Schwefel 1993).
The most popular EAs are genetic algorithms (Goldberg 1989), evolutionary
programming (Fogel 1992), and evolution strategies (Schwefel 1981). Although
these three EAs have some significant differences in the way they implement
ideas from evolution, they share the same basic methodology. Four major algo-
rithmic steps are needed to apply an EA:
Step 1. Generate an initial population of solutions to the problem by
distributing them throughout the solution space.
Step 2. Accurately compute the fitness (response) of each solution.
Step 3. Based on the fitness of the solutions, select the best solutions and
apply idealized-type genetic operators to produce a new generation of
offspring solutions.
Step 4. Return to Step 2 as long as the algorithm is locating improved
solutions.
10.4.1 Combining Evolutionary Algorithms with Simulation

The deviations mentioned previously from traditional nonlinear optimization
techniques give EAs several advantages for solving simulation optimization prob-
lems. For example, Biethahn and Nissen (1994), Bäck et al. (1995), and Bäck and
Schwefel (1993) report that EAs are very suitable for simulation optimization
because
• They require no restrictive assumptions or prior knowledge about the
topology of the response surface being searched, thereby making them
broadly applicable techniques.
• They are well suited for problems with response surfaces that are
highly dimensional, multimodal, discontinuous, nondifferentiable, and
stochastic, and even for problems with moving response surfaces.
• They are very reliable search techniques and are relatively easy to use.
A potential disadvantage of using EAs for simulation optimization is that
they sometimes require the evaluation of many solutions (calls to the simulation
model). This could be prohibitive if computational time is limited. However, there
are procedures that can be used to reduce the number of times the simulation
model is called. For example, Hall and Bowden (1996) developed a method that
combines an EA with a more locally oriented direct search technique that results
in substantially fewer calls to the simulation model.
Several examples of using EAs for simulation optimization have appeared in
the research literature. Lacksonen and Anussornnitisarn (1995) compared a genetic
algorithm (GA) with three other direct search techniques (Hooke Jeeves Pattern
Search, Nelder Mead Simplex Search, and Simulated Annealing) on 20 simula-
tion optimization test problems and ranked GA as best. Stuckman, Evans, and
har01307_ch10_273-298.indd 278 12/17/10 12:02 PM

Mollaghasemi (1991) suggest that GAs and Simulated Annealing are the algorithms
of choice when dealing with a large number of decision variables. Tompkins and
Azadivar (1995) recommend using GAs when the optimization problem involves
qualitative (logical) decision variables. The authors have extensively researched
the use of genetic algorithms, evolutionary programming, and evolution strategies
for solving manufacturing simulation optimization problems and simulation-based
machine learning problems (Bowden 1995; Bowden, Neppalli, and Calvert 1995;
Bowden, Hall, and Usher 1996; Hall, Bowden, and Usher 1996).
Reports have also appeared in the trade journal literature on how the EA-
based optimizer in SimRunner helped to solve real-world problems. For example,
IBM; Sverdrup Facilities, Inc.; and Baystate Health Systems report benefits from
using SimRunner as a decision support tool (Akbay 1996). The simulation group
at Lockheed Martin used SimRunner to help determine the most efficient lot sizes
for parts and when the parts should be released to the system to meet schedules
(Anonymous 1996).
10.4.2 Illustration of an Evolutionary Algorithm’s

Search of a Response Surface
In this section, we demonstrate how EAs perform when applied to problems that
are characterized by multimodal and stochastic response surfaces that can be pro-
duced by a simulation model’s output. For convenience of presentation, a one-
decision-variable version of a function developed by Ackley (1987) is used to gain
insight on how EAs seek the optimal solution. First Ackley’s function is trans-
formed from a deterministic function to a stochastic function. This is done in the
usual way by adding stochastic perturbations (noise) to the response generated by
Ackley’s function. Figure 10.3 illustrates Ackley’s function when noise is added
FIGURE 10.3
Ackley’s function with noise and the ES’s progress over eight generations.
20 Ackley's Function
18
Generation 1
16
Measured response
Generation 2
14
12 Generation 3
10 Generation 4
8 Generation 5
6
Generation 6
4
2 Generation 7
0 Generation 8
⫺10 ⫺5 0 5 10
Decision variable X
har01307_ch10_273-298.indd 279 12/17/10 12:02 PM

to it by sampling from a uniform (⫺1, 1) distribution. This simulates the variation

that can occur in the output from a stochastic simulation model. The function is
shown with a single decision variable X that takes on real values between ⫺10 and
10. The response surface has a minimum expected value of zero, occurring when X
is equal to zero, and a maximum expected value of 19.37 when X is equal to ⫺9.54
or ⫹9.54. Ackley’s function is multimodal and thus has several locally optimal
solutions (optima) that occur at each of the low points (assuming a minimization
problem) on the response surface. However, the local optimum that occurs when X
is equal to zero is the global optimum (the lowest possible response). This is a use-
ful test function because search techniques can prematurely converge and end their
search at one of the many optima before finding the global optimum.
According to Step 1 of the four-step process outlined in section 10.4, an initial
population of solutions to the problem is generated by distributing them through-
out the solution space. Using a variant of Schwefel’s (1981) Evolution Strategy
(ES) with two parent solutions and 10 offspring solutions, 10 different values for
the decision variable between −10 and 10 are randomly picked to represent an
initial offspring population of 10 solutions. However, to make the search for the
optimal solution more challenging, the 10 solutions in the initial offspring popula-
tion are placed far from the global optimum to see if the algorithm can avoid being
trapped by one of the many local optima. Therefore, the 10 solutions for the first
generation were randomly picked between −10 and −8. So the test is to see if the
population of 10 solutions can evolve from one generation to the next to find the
global optimal solution without being trapped by one of the many local optima.
Figure 10.3 illustrates the progress that the ES made by following the four-
step process from section 10.4. To avoid complicating the graph, only the re-
sponses for the two best solutions (parents) in each generation are plotted on the
response surface. Clearly, the process of selecting the best solutions and applying
idealized genetic operators allows the algorithm to focus its search toward the
optimal solution from one generation to the next. Although the ES samples many
of the local optima, it quickly identifies the region of the global optimum and is
beginning to hone in on the optimal solution by the eighth generation. Notice that
in the sixth generation, the ES has placed solutions to the right side of the search
space (X ⬎ 0) even though it was forced to start its search at the far left side of
the solution space. This ability of an evolutionary algorithm allows it to conduct
a more globally oriented search.
When a search for the optimal solution is conducted in a noisy simulation en-
vironment, care should be taken in measuring the response generated for a given
input (solution) to the model. This means that to get an estimate of the expected
value of the response, multiple observations (replications) of a solution’s per-
formance should be averaged. But to test the idea that EAs can deal with noisy
response surfaces, the previous search was conducted using only one observation
of a solution’s performance. Therefore, the potential existed for the algorithm to
become confused and prematurely converge to one of the many local optima be-
cause of the noisy response surface. Obviously, this was not the case with the EA
in this example.
har01307_ch10_273-298.indd 280 12/17/10 12:02 PM

The authors are not advocating that analysts can forget about determining
the number of replications needed to satisfactorily estimate the expected value of
the response. However, to effectively conduct a search for the optimal solution,
an algorithm must be able to deal with noisy response surfaces and the resulting
uncertainties that exist even when several observations (replications) are used to
estimate a solution’s true performance.
10.5 Strategic and Tactical Issues of Simulation Optimization

When optimizing a simulated system for the purpose of applying the solution to
an actual system, the analyst must first ensure that the simulation model accu-
rately mimics the real system. This subject is discussed in Chapter 7. Additionally,
the analyst must determine the appropriate amount of time to simulate the system,
the warm-up period for the model (if applicable), and the required number of in-
dependent observations (model replications). A detailed treatment of these topics
can be found in Chapter 8.
Another issue very relevant to simulation optimization is the efficiency of
the simulation study. Here efficiency is measured in terms of the amount of time
necessary to conduct experiments with the simulation model. In the context of
an optimization project, an experiment is the evaluation of a solution to estimate
its utility. Optimization projects often require the evaluation of many solutions,
and a few seconds saved per experiment can translate into a significant reduction
in time during an optimization project. Therefore, the amount of time it takes
to run the model for one replication (operational efficiency) and the number of
replications required to satisfactorily estimate the expected value of the model’s
response for a given input solution to the model (statistical efficiency) are impor-
tant considerations.
10.5.1 Operational Efficiency

When a general-purpose language like C is used to build a simulation model, the
programmer has a great deal of control over how efficiently, or fast, the program
operates. For example, minimizing the number of lines of code in a program is
one way to reduce the time required to execute the program. When using a simu-
lation language or a simulator, the analyst has less control over the program’s
efficiency. However, the analyst still has control over other factors that affect the
time required to run a model such as the level of detail included in the model, the
number of different statistics collected during model execution, and the use of
animation.
A good rule is not to take the approach of including every possible element of
a system in a model, as a way of making sure something important is not left out
of the model. Each element added to the model usually increases the time required
to run the simulation. Therefore, include only those elements of the system that
have a direct bearing on the problem being studied. For these essential elements,
har01307_ch10_273-298.indd 281 12/17/10 12:02 PM

select the minimum number of statistics necessary to estimate the utility of a solu-
tion. ProModel products allow users to turn off statistical tracking for locations,
entities, resources, and variables that are not needed to estimate the performance
of a solution. If, for example, a model is to estimate an entity’s time in the system,
then only entity-based statistics need be recorded. Also, turning the animation
feature off generally reduces the time to run the model.
10.5.2 Statistical Efficiency

One luxury afforded to model builders is that the variance of a performance mea-
sure computed from the output of simulations can be reduced. This is a luxury
because reducing the variance allows us to estimate the mean value of a random
variable within a desired level of precision and confidence with fewer replications
(independent observations). The reduction in the required number of replications
is achieved by controlling how random variates are used to “drive” the events in
the simulation model. These time-saving techniques are called variance reduction
techniques.
One popular variance reduction technique called common random numbers
(CRN) was covered in Chapter 9. The CRN technique was invented to compare
different system configurations, or solutions, in the context of simulation optimiza-
tion. The CRN technique is intuitively appealing because it provides a means for
comparing solutions under more equal experimental conditions than if CRN were
not used. This is helpful in ensuring that the differences in the performance of two
solutions are due to the differences in the solutions themselves and not to differences
in experimental conditions. Therefore, the use of CRN is generally recommended.
10.5.3 General Optimization Procedure

After a credible and efficient (both operationally and statistically) simulation
model has been built, it is time to begin the optimization project. The general
procedure and rules of thumb for carrying out an optimization project with the
EA-based optimization module in SimRunner follow.
Step 1. The decision variables believed to affect the output of the simulation
model are first programmed into the model as variables whose values can be
quickly changed by the EA. Decision variables are typically the parameters whose
values can be adjusted by management, such as the number of nurses assigned to
a shift or the number of machines to be placed in a work cell.
Step 2. For each decision variable, define its numeric data type (integer or real)
and its lower bound (lowest possible value) and upper bound (highest possible
value). During the search, the EA will generate solutions by varying the values of
decision variables according to their data types, lower bounds, and upper bounds.
The number of decision variables and the range of possible values affect the size
of the search space (number of possible solutions to the problem). Increasing the
har01307_ch10_273-298.indd 282 12/17/10 12:02 PM

number of decision variables or their range of values increases the size of the
search space, which can make it more difficult and time-consuming to identify
the optimal solution. As a rule, include only those decision variables known to
significantly affect the output of the simulation model and judiciously define the
range of possible values for each decision variable. Also, care should be taken
when defining the lower and upper bounds of the decision variables to ensure that
a combination of values will not be created that lead to a solution not envisioned
when the model was built.
Step 3. After selecting the decision variables, construct the objective function
to measure the utility of the solutions tested by the EA. Actually, the foundation
for the objective function would have already been established when the goals for
the simulation project were set. For example, if the goal of the modeling project is
to find ways to minimize a customer’s waiting time in a bank, then the objective
function should measure an entity’s (customer’s) waiting time in the bank. The
objective function is built using terms taken from the output report generated at
the end of the simulation run. Objective function terms can be based on entity sta-
tistics, location statistics, resource statistics, variable statistics, and so on. The user
specifies whether a term is to be minimized or maximized as well as the overall
weighting of that term in the objective function. Some terms may be more or less
important to the user than other terms. Remember that as terms are added to the
objective function, the complexity of the search space may increase, which makes
a more difficult optimization problem. From a statistical point of view, single-term
objective functions are also preferable to multiterm objective functions. Therefore,
strive to keep the objective function as specific as possible.
The objective function is a random variable, and a set of initial experiments
should be conducted to estimate its variability (standard deviation). Note that there
is a possibility that the objective function’s standard deviation differs from one
solution to the next. Therefore, the required number of replications necessary to es-
timate the expected value of the objective function may change from one solution
to the next. Thus the objective function’s standard deviation should be measured
for several different solutions and the highest standard deviation recorded used to
compute the number of replications necessary to estimate the expected value of
the objective function. When selecting the set of test solutions, choose solutions
that are very different from one another. For example, form solutions by setting the
decision variables to their lower bounds, middle values, or upper bounds.
A better approach for controlling the number of replications used to estimate
the expected value of the objective function for a given solution would be to in-
corporate a rule into the model that schedules additional replications until the es-
timate reaches a desired level of precision (confidence interval half-width). Using
this technique can help to avoid running too many replications for some solutions
and too few replications for others.
Step 4. Select the size of the EA’s population (number of solutions) and begin
the search. The size of the population of solutions used to conduct the search
har01307_ch10_273-298.indd 283 12/17/10 12:02 PM

affects both the likelihood that the algorithm will locate the optimal solution and
the time required to conduct the search. In general, as the population size is in-
creased, the algorithm finds better solutions. However, increasing the population
size generally increases the time required to conduct the search. Therefore, a bal-
ance must be struck between finding the optimum and the amount of available
time to conduct the search.
Step 5. After the EA’s search has concluded (or halted due to time constraints),
the analyst should study the solutions found by the algorithm. In addition to the
best solution discovered, the algorithm usually finds many other competitive
solutions. A good practice is to rank each solution evaluated based on its utility
as measured by the objective function. Next, select the most highly competitive
solutions and, if necessary, make additional model replications of those solutions
to get better estimates of their true utility. And, if necessary, refer to Chapter 9
for background on statistical techniques that can help you make a final decision
between competing solutions. Also, keep in mind that the database of solutions
evaluated by the EA represents a rich source of information about the behavior,
or response surface, of the simulation model. Sorting and graphing the solutions
can help you interpret the “meaning” of the data and gain a better understanding
of how the system behaves.
If the general procedure presented is followed, chances are that a good course
of action will be identified. This general procedure is easily carried out using Pro-
Model simulation products. Analysts can use SimRunner to help
• Determine the length of time and warm-up period (if applicable) for
running a model.
• Determine the required number of replications for obtaining estimates
with a specified level of precision and confidence.
• Search for the optimal values for the important decision variables.
Even though it is easy to use SimRunner and other modern optimizers, do
not fall into the trap of letting them become the decision maker. Study the top
solutions found by the optimizers as you might study the performance records of
different cars for a possible purchase. Kick their tires, look under their hoods, and
drive them around the block before buying. Always remember that the optimizer
is not the decision maker. It only suggest a possible course of action. It is the
user’s responsibility to make the final decision.
10.6 Formulating an Example Optimization Problem

The optimal allocation of buffer storage between workstations is a common indus-
trial problem and has been extensively studied by academic researchers. If a pro-
duction line consists of a network of tightly linked workstations, an inventory of
parts (work in progress) between workstations can be used to minimize the chance
har01307_ch10_273-298.indd 284 12/17/10 12:02 PM

that disruptions (machine failures, line imbalances, quality problems, or the like)
will shut down the production line. Several strategies have been developed for
determining the amount of buffer storage needed between workstations. However,
these strategies are often developed based on simplifying assumptions, made for
mathematical convenience, that rarely hold true for real production systems.
One way to avoid oversimplifying a problem for the sake of mathematical
convenience is to build a simulation model of the production system and use it
to help identify the amount of buffer storage space needed between workstations.
However, the number of possible solutions to the buffer allocation problem grows
rapidly as the size of the production system (number of possible buffer storage
areas and their possible sizes) increases, making it impractical to evaluate all so-
lutions. In such cases, it is helpful to use simulation optimization software like
SimRunner to identify a set of candidate solutions.
This example is loosely based on the example production system presented
in Chapter 9. It gives readers insight into how to formulate simulation optimiza-
tion problems when using SimRunner. The example is not fully solved. Its com-
pletion is left as an exercise in Lab Chapter 10.
10.6.1 Problem Description

The production system consists of four machines and three buffers. Parts enter-
ing the system require processing by each of the four machines in a serial fashion
(Figure 10.4). A part is always available for processing at the first machine. After a
part is processed, it moves from the machine to the buffer storage area for the next
FIGURE 10.4
Production system
with four workstations
and three buffer
storage areas.
.72 4.56
har01307_ch10_273-298.indd 285 12/17/10 12:02 PM

machine, where it waits to be processed. However, if the buffer is full, the part
cannot move forward and remains on the machine until a space becomes available
in the buffer. Furthermore, the machine is blocked and no other parts can move
to the machine for processing. The part exits the system after being processed by
the fourth machine. Note that parts are selected from the buffers to be processed
by a machine in a first-in, first-out order. The processing time at each machine is
exponentially distributed with a mean of 1.0 minute, 1.3 minutes, 0.7 minute, and
1.0 minute for machines one, two, three, and four, respectively. The time to move
parts from one location to the next is negligible.
For this problem, three decision variables describe how buffer space is al-
located (one decision variable for each buffer to signify the number of parts that
can be stored in the buffer). The Goal is to find the optimal value for each decision
variable to maximize the profit made from the sale of the parts. The manufacturer
collects $10 per part produced. The limitation is that each unit of space provided
for a part in a buffer costs $1,000. So the buffer storage has to be strategically al-
located to maximize the throughput of the system. Throughput will be measured
as the number of parts completed during a 30-day period.
10.6.2 Demonstration of the General Optimization Procedure

The general optimization procedure outlined in section 10.5.3 will be used to
guide the formulation of the problem as a SimRunner simulation optimization
project. The assumption is that a credible and efficient simulation model has been
built of the production system.
Step 1. In this step, the decision variables for the problem are identified and
defined. Three decision variables are needed to represent the number of parts that
can be stored in each buffer (buffer capacity). Let Q1, Q2, and Q3 represent the
number of parts that can be stored in buffers 1, 2, and 3, respectively.
Step 2. The numeric data type for each Qi is integer. If it is assumed that each
buffer will hold a minimum of one part, then the lower bound for each Qi is 1. The
upper bound for each decision variable could arbitrarily be set to, say, 20. How-
ever, keep in mind that as the range of values for a decision variable is increased,
the size of the search space also increases and will likely increase the time to con-
duct the search. Therefore, this is a good place to apply existing knowledge about
the performance and design of the system.
For example, physical constraints may limit the maximum capacity of each
buffer to no greater than 15 parts. If so, there is no need to conduct a search with
a range of values that produce infeasible solutions (buffer capacities greater than
15 parts). Considering that the fourth machine’s processing time is larger than the
third machine’s processing time, parts will tend to queue up at the third buffer.
Therefore, it might be a good idea to set the upper bound for Q3 to a larger value
than the other two decision variables. However, be careful not to assume too much
har01307_ch10_273-298.indd 286 12/17/10 12:02 PM

because it could prevent the optimization algorithm from exploring regions in the
search space that may contain good solutions.
With this information, it is decided to set the upper bound for the capacity of
each buffer to 15 parts. Therefore, the bounds for each decision variable are
1 ⱕ Q1 ⱕ 15 1 ⱕ Q2 ⱕ 15 1 ⱕ Q3 ⱕ 15
Given that each of the three decision variable has 15 different values, there are
153, or 3,375, unique solutions to the problem.
Step 3. Here the objective function is formulized. The model was built to in-
vestigate buffer allocation strategies to maximize the throughput of the system.
Given that the manufacturer collects $10 per part produced and that each unit of
space provided for a part in a buffer costs $1,000, the objective function for the
optimization could be stated as
Maximize[$10(Throughput) ⫺ $1,000(Q1 ⫹ Q2 ⫹ Q3)]
where Throughput is the total number of parts produced during a 30-day
period.
Next, initial experiments are conducted to estimate the variability of the ob-
jective function in order to determine the number of replications the EA-based
optimization algorithm will use to estimate the expected value of the objective
function for each solution it evaluates. While doing this, it was also noticed that
the throughput level increased very little for buffer capacities beyond a value
of nine. Therefore, it was decided to change the upper bound for each decision
variable to nine. This resulted in a search space of 93, or 729, unique solutions, a
reduction of 2,646 solutions from the original formulation. This will likely reduce
the search time.
Step 4. Select the size of the population that the EA-based optimization al-
gorithm will use to conduct its search. SimRunner allows the user to select an
optimization profile that influences the degree of thoroughness used to search
for the optimal solution. The three optimization profiles are aggressive, moder-
ate, and cautious, which correspond to EA population sizes of small, medium,
and large. The aggressive profile generally results in a quick search for locally
optimal solutions and is used when computer time is limited. The cautious pro-
file specifies that a more thorough search for the global optimum be conducted
and is used when computer time is plentiful. At this point, the analyst knows the
amount of time required to evaluate a solution. Only one more piece of informa-
tion is needed to determine how long it will take the algorithm to conduct its
search. That is the fraction of the 729 solutions the algorithm will evaluate before
converging to a final solution. Unfortunately, there is no way of knowing this in
advance. With time running out before a recommendation for the system must be
given to management, the analyst elects to use a small population size by select-
ing SimRunner’s aggressive optimization profile.
har01307_ch10_273-298.indd 287 12/17/10 12:02 PM

Step 5. After the search concludes, the analyst selects for further evaluation
some of the top solutions found by the optimization algorithm. Note that this does
not necessarily mean that only those solutions with the best objective function
values are chosen, because the analyst should conduct both a quantitative and a
qualitative analysis. On the quantitative side, statistical procedures presented in
Chapter 9 are used to gain a better understanding of the relative differences in per-
formance between the candidate solutions. On the qualitative side, one solution
may be preferred over another based on factors such as ease of implementation.
Figure 10.5 illustrates the results from a SimRunner optimization of the buffer
allocation problem using an aggressive optimization profile. The warm-up time for
the simulation was set to 10 days, with each day representing a 24-hour production
period. After the warm-up time of 10 days (240 hours), the system is simulated for
an additional 30 days (720 hours) to determine throughput. The estimate for the
expected value of the objective function was based on five replications of the simu-
lation. The smoother line that is always at the top of the Performance Measures Plot
in Figure 10.5 represents the value of the objective function for the best solution
identified by SimRunner during the optimization. Notice the rapid improvement in
the value of the objective function during the early part of the search as SimRunner
identifies better buffer capacities. The other, more irregular line represents the value
of the objective function for all the solutions that SimRunner tried.
SimRunner’s best solution to the problem specifies a Buffer 1 capacity of
nine, a Buffer 2 capacity of seven, and a Buffer 3 capacity of three and was the
33rd solution (experiment) evaluated by SimRunner. The best solution is located
at the top of the table in Figure 10.5. SimRunner sorts the solutions in the table
FIGURE 10.5
SimRunner results for
the buffer allocation
problem.
har01307_ch10_273-298.indd 288 12/17/10 12:02 PM

from best to worst. SimRunner evaluated 82 out of the possible 729 solutions.
Note that there is no guarantee that SimRunner’s best solution is in fact the opti-
mum solution to the problem. However, it is likely to be one of the better solutions
to the problem and could be the optimum one.
The last two columns in the SimRunner table shown in Figure 10.5 display
the lower and upper bounds of a 95 percent confidence interval for each solution
evaluated. Notice that there is significant overlap between the confidence inter-
vals. Although this is not a formal hypothesis-testing procedure, the overlapping
confidence intervals suggest the possibility that there is not a significant differ-
ence in the performance of the top solutions displayed in the table. Therefore, it
would be wise to run additional replications of the favorite solutions from the list
and/or use one of the hypothesis-testing procedures in Chapter 9 before selecting
a particular solution as the best. The real value here is that SimRunner automati-
cally conducted the search, without the analyst having to hover over it, and re-
ported back several good solutions to the problem for the analyst to consider.
10.7 Real-World Simulation Optimization Project

In this section, we provide an account of how simulation combined with optimization
was used to solve a kanban sizing problem. The objective for this problem is to find
the minimum amount of inventory required to support production in a just-in-time
(JIT) environment where parts are pulled through the system using kanban cards. This
presentation is based on a problem faced by a major appliance factory in the United
States. Two different techniques are used to solve the problem. One approach uses
a technique developed by Toyota. The other approach uses an EA-based simulation
optimization algorithm similar to the one used in SimRunner. The presentation con-
cludes with a comparative analysis of the two solutions generated by these techniques.
10.7.1 Problem Description

A kanban is a card that authorizes a workstation to begin producing a part. Kan-
ban cards are passed from downstream workstations to upstream workstations,
signaling the upstream workstation to begin producing a part. This results in parts
being pulled through the system as downstream workstations issue kanban cards
to upstream workstations.
In a pull production system, the number of kanban cards directly influences
the number of material-handling containers required for transporting (pulling) parts
through the system and thus the maximum amount of work-in-progress (WIP) in-
ventory. Trigger values represent the number of kanban cards that must accumulate
before a workstation is authorized to begin production. Therefore, trigger values
control the frequency of machine setups. While the ideal minimum trigger value
equals 1, many processes do not allow for this due to the time required for machine
setup or maintenance. Solving for the optimal number of kanban cards and cor-
responding trigger values presents a difficult task for production planners.
har01307_ch10_273-298.indd 289 12/17/10 12:02 PM

FIGURE 10.6
The two-stage pull Trigger ⫽ 1 Trigger ⫽ 2
production system. Customer

Raw
demand Assembly line
Stage Stage materials
One Stage One Two Stage Two
WIP process WIP line
Assembly line
Kanban posts
Final product Component
Subassembly Kanban card
Figure 10.6 illustrates the relationship of the processes in the two-stage pull
production system of interest that produces several different types of parts. Cus-
tomer demand for the final product causes containers of subassemblies to be pulled
from the Stage One WIP location to the assembly lines. As each container is with-
drawn from the Stage One WIP location, a production-ordering kanban card rep-
resenting the number of subassemblies in a container is sent to the kanban post for
the Stage One processing system. When the number of kanban cards for a given
subassembly meets its trigger value, the necessary component parts are pulled
from the Stage Two WIP to create the subassemblies. Upon completing the Stage
One process, subassemblies are loaded into containers, the corresponding kanban
card is attached to the container, and both are sent to Stage One WIP. The container
and card remain in the Stage One WIP location until pulled to an assembly line.
In Stage Two, workers process raw materials to fill the Stage Two WIP loca-
tion as component parts are pulled from it by Stage One. As component parts are
withdrawn from the Stage Two WIP location and placed into the Stage One process,
a production-ordering kanban card representing the quantity of component parts in
a container is sent to the kanban post for the Stage Two line. When the number of
kanban cards for a given component part meets a trigger value, production orders
equal to the trigger value are issued to the Stage Two line. As workers move com-
pleted orders of component parts from the Stage Two line to WIP, the corresponding
kanban cards follow the component parts to the Stage Two WIP location.
While an overall reduction in WIP is sought, production planners desire a
solution (kanban cards and trigger values for each stage) that gives preference to
minimizing the containers in the Stage One WIP location. This requirement is due
to space limitations.
10.7.2 Simulation Model and Performance Measure

Without sacrificing the presentation of the simulation optimization subject, only
a sketch of the model is given. The pull production system was represented in a
har01307_ch10_273-298.indd 290 12/17/10 12:02 PM

discrete-event simulation model. The model assumes constant interarrival times for
customer orders to the assembly lines. However, workers at the assembly lines may
find defects in the subassemblies, which increases subassembly demand above that
needed for scheduled customer demand. The number of defective subassemblies
found on the assembly lines follows a stochastic function. Production times for the
Stage One process and Stage Two line are constant because they are automated pro-
cesses. Setup time, time between failures, and time to repair for the Stage Two line
are triangularly distributed. Each simulation run consists of a 10-day (two-week)
warm-up period followed by a 20-day (one-month) steady-state period.
The performance measure to be maximized is
f (a) ⫽ W1 (AvgPct ⫹ MinPct) ⫺ W2 (TK1) ⫺ W3 (TK2)
where AvgPct is the average throughput percentage (percentage of demand satis-
fied) for all product types coming off the assembly lines; MinPct is the minimum
throughput percentage for all product types coming off the assembly lines; TK1
is the total number of kanban cards in Stage One; TK2 is the total number of
kanban cards in Stage Two; and W1, W2, and W3 are coefficients that reflect the
level of importance (weighting) assigned to each term by production planners.
The weighting levels assigned to W1, W2, and W3 are 20.0, 0.025, and 0.015, re-
spectively. Using this performance function, the objective is to find a solution that
will maximize throughput while minimizing the total number of kanban cards in
the system.
The production planners decided to fix the trigger values in the Stage One
process to 1; therefore, the optimal number of kanban cards in the Stage One
process and the optimal number of kanban cards and corresponding trigger values
in the Stage Two line for each of the 11 different part types being produced need
to be determined. Thus the problem contains 33 integer decision variables. Each
solution is a vector a ⫽ (K11, K12, K13, . . . , K111; K21, K22, K23, . . . , K211; T 21,
T 22, T 23, . . . , T 211) where K1i represents the number of kanban cards for part type
i circulating in the Stage One process, K2i represents the number of kanban cards
for part type i circulating in the Stage Two line, and T2i represents the correspond-
ing trigger quantities for part type i used in the Stage Two line. Considering the
range of possible values for each of the decision variables (omitted for brevity),
there are 4.81 ⫻ 1041 unique solutions to this problem.
10.7.3 Toyota Solution Technique

Because of their ease of use, the kanban sizing equations developed by Toyota are
often used by manufacturers to estimate the number of kanban cards needed. For this
problem, the constant-quantity, nonconstant-cycle withdrawal system as described
by Monden (1993) was used. The number of kanbans required is determined by
Kanbans ⫽
Monthly setups demand coefficient
Monthly demand
______________ ⫹
Daily
⫻
Safety
______________________________________
Container capacity
har01307_ch10_273-298.indd 291 12/17/10 12:02 PM

where the safety coefficient represents the amount of WIP needed in the system.
Production planners assumed a safety coefficient that resulted in one day of WIP
for each part type. Additionally, they decided to use one setup per day for each part
type. Although this equation provides an estimate of the minimum number of kanban
cards, it does not address trigger values. Therefore, trigger values for the Stage Two
line were set at the expected number of containers consumed for each part type in one
day. The Toyota equation recommended using a total of 243 kanban cards. The details
of the calculation are omitted for brevity. When evaluated in the simulation model,
this solution yielded a performance score of 35.140 (using the performance function
defined in section 10.7.2) based on four independent simulation runs (replications).
10.7.4 Simulation Optimization Technique

Here an EA-based optimization module like the one in SimRunner is used to
search for an optimal solution to the problem. Given that it is not practical to as-
sume that there is an unlimited amount of time for conducting a search, the soft-
ware allows the user to select an optimization profile that influences the degree
of thoroughness used to search for the optimal solution. The three optimization
profiles are aggressive, moderate, and cautious, which correspond to EA popula-
tion sizes of small, medium, and large. The aggressive profile generally results in
a quick search and is used when computer time is limited. The cautious profile
specifies that a more thorough search be conducted. The moderate profile was
used to search for an optimal number of kanban cards and trigger values.
Each time the simulation model is called to evaluate a solution, the model runs
four replications to get four independent observations of a solution’s performance.
The performance of a solution is computed at the end of each simulation run using
the performance function defined in section 10.7.2. The final performance score
assigned to each solution is the average of the four independent observations.
Figure 10.7 illustrates the progress of the optimization module’s search. Be-
cause it is difficult to visualize a response surface with 33 dimension variables,
only the performance score of the best solution found per generation is plotted.
The points are connected with a line to highlight the rate at which the algorithm
finds solutions with improved performance scores. Note that the optimization
module quickly finds solutions that are better (based on the four independent ob-
servations) than the solution found using the Toyota technique, and the optimiza-
tion module steadily finds better solutions after that point. The search was halted
at 150 generations because the solution was not improving significantly. The solu-
tion recommended using a total of 110 kanban cards and yielded a performance
score of 37.945 based on four replications.
10.7.5 Comparison of Results

Final solutions generated by the optimization module and Toyota equation were
run in the simulation model for 20 independent replications to get a better estimate
of the true mean performance of the solutions. Table 10.1 presents a comparison
har01307_ch10_273-298.indd 292 12/17/10 12:02 PM

FIGURE 10.7 39 Optimization module

After the second Toyota
generation, the 38
Measured response
solutions found by the
optimization module 37
are better than the
solutions generated 36
using the Toyota
method.
35
34
33
0 50 100 150
Generation
TABLE 10.1 The Optimization Module’s Solution Meets Production

Throughput Goals with 54 Percent Fewer Kanban Cards
Than Toyota
Technique Kanban Cards Performance Scores
Optimization module 110 37.922

Toyota solution 243 35.115
of the solutions. Both solutions achieved the required throughput using a substan-
tially different number of kanban cards.
An application of Fisher’s LSD statistical multiple comparison test presented
in Chapter 9 suggests that the performance scores for the methods in Table 10.1
are different at a 0.05 level of significance. Results indicate that the optimization
module produced a significantly better solution to the kanban sizing problem than
did the production planners using the Toyota technique. The solution that the op-
timization module found requires approximately 54 percent fewer kanban cards
(54 percent less WIP) than the solution generated using the Toyota technique—a
considerable reduction in inventory carrying costs.
Considering the time required to run the simulation model on a personal com-
puter and the fact that there are 4.81 ⫻ 1041 unique solutions to the problem, it
would take years to evaluate all solutions to the problem. Obviously, this is not a
practical alternative. The optimization module required only a few hours to com-
plete its search. It evaluated only a small fraction of the total number of solutions
and provided a much better starting point than the Toyota solution for making the
final decision regarding the actual pull production system, a noteworthy contribu-
tion to the decision-making process whether or not the solution found is the global
optimum.
har01307_ch10_273-298.indd 293 12/17/10 12:02 PM

10.8 Summary
In recent years, major advances have been made in the development of user-
friendly simulation software packages. However, progress in developing simu-
lation output analysis tools has been especially slow in the area of simulation
optimization because conducting simulation optimization with traditional tech-
niques has been as much an art as a science (Greenwood, Rees, and Crouch 1993).
There are such an overwhelming number of traditional techniques that only in-
dividuals with extensive backgrounds in statistics and optimization theory have
realized the benefits of integrating simulation and optimization concepts. Using
newer optimization techniques, it is now possible to narrow the gap with user-
friendly, yet powerful, tools that allow analysts to combine simulation and opti-
mization for improved decision support. SimRunner is one such tool.
Our purpose is not to argue that evolutionary algorithms are the panacea for solv-
ing simulation optimization problems. Rather, our purpose is to introduce the reader
to evolutionary algorithms by illustrating how they work and how to use them for
simulation optimization and to expose the reader to the wealth of literature that dem-
onstrates that evolutionary algorithms are a viable choice for reliable optimization.
There will always be debate on what are the best techniques to use for simula-
tion optimization. Debate is welcome as it results in better understanding of the
real issues and leads to the development of better solution procedures. It must be
remembered, however, that the practical issue is not that the optimization tech-
nique guarantees that it locates the optimal solution in the shortest amount of time
for all possible problems that it may encounter; rather, it is that the optimization
technique consistently finds good solutions to problems that are better than the
solutions analysts are finding on their own. Newer techniques such as evolution-
ary algorithms and scatter search meet this requirement because they have proved
robust in their ability to solve a wide variety of problems, and their ease of use
makes them a practical choice for simulation optimization today (Boesel et al.
2001; Brady and Bowden 2001).

1. What is a response surface?
2. What does it mean if an optimization algorithm is classified as a heuristic
technique?
3. What is the most significant difference between the way evolutionary
algorithms (EAs) search a response surface and the way many traditional
nonlinear optimization algorithms conduct a search? What advantage
does this difference give to an EA over many traditional nonlinear
optimization algorithms?
4. Assume you are working on a project with five decision variables (A, B,
C, D, E). Each decision variable’s numeric data type is integer with
the following values: 1, 2, 3, 4, 5, 6. How many different solutions can
har01307_ch10_273-298.indd 294 12/17/10 12:02 PM

be generated from these five decision variables? If it takes one minute

of computer time to evaluate a solution (run the simulation model for
the required number of replications), how many hours would it take to
evaluate all solutions?
5. Define operational and statistical efficiency in the context of simulation
optimization.
References
Ackley, D. A Connectionist Machine for Genetic Hill Climbing. Boston, MA: Kluwer, 1987.
Akbay, K. “Using Simulation Optimization to Find the Best Solution.” IIE Solutions, May
1996, pp. 24–29.
Anonymous, “Lockheed Martin.” IIE Solutions, December, 1996, pp. SS48–SS49.
Azadivar, F. “A Tutorial on Simulation Optimization.” 1992 Winter Simulation Conference.
Arlington Virginia, ed. Swain, J., D. Goldsman, R. Crain and J. Wilson. Institute of
Electrical and Electronics Engineers, Piscataway, NJ: 1992, pp. 198–204.
Bäck, T.; T. Beielstein; B. Naujoks; and J. Heistermann. “Evolutionary Algorithms for the
Optimization of Simulation Models Using PVM.” Euro PVM 1995—Second European
PVM 1995, User’s Group Meeting. Hermes, Paris, ed. Dongarra, J., M. Gengler,
B. Tourancheau and X. Vigouroux, 1995, pp. 277–282.
Bäck, T., and H.-P. Schwefel. “An Overview of Evolutionary Algorithms for Parameter
Optimization.” Evolutionary Computation 1, no. 1 (1993), pp. 1–23.
Barton, R., and J. Ivey. “Nelder-Mead Simplex Modifications for Simulation Optimization.”
Management Science 42, no. 7 (1996), pp. 954–73.
Biethahn, J., and V. Nissen. “Combinations of Simulation and Evolutionary Algorithms
in Management Science and Economics.” Annals of Operations Research 52 (1994),
pp. 183–208.
Boesel, J.; Bowden, R. O.; Glover, F.; and J. P. Kelly. “Future of Simulation Optimization.”
Proceedings of the 2001 Winter Simulation Conference, 2001, pp. 1466–69.
Bowden, R. O. “Genetic Algorithm Based Machine Learning Applied to the Dynamic
Routing of Discrete Parts.” Ph.D. Dissertation, Department of Industrial Engineering,
Mississippi State University, 1992.
Bowden, R. “The Evolution of Manufacturing Control Knowledge Using Reinforcement
Learning.” 1995 Annual International Conference on Industry, Engineering, and
Management Systems, Cocoa Beach, FL, ed. G. Lee. 1995, pp. 410–15.
Bowden, R., and S. Bullington. “An Evolutionary Algorithm for Discovering Manufactur-
ing Control Strategies.” In Evolutionary Algorithms in Management Applications, ed.
Biethahn, J. and V. Nissen. Berlin: Springer, 1995, pp. 124–38.
Bowden, R., and S. Bullington. “Development of Manufacturing Control Strategies Using
Unsupervised Learning.” IIE Transactions 28 (1996), pp. 319–31.
Bowden, R.; J. Hall; and J. Usher. “Integration of Evolutionary Programming and Simula-
tion to Optimize a Pull Production System.” Computers and Industrial Engineering 31,
no. 1/2 (1996), pp. 217–20.
Bowden, R.; R. Neppalli; and A. Calvert. “A Robust Method for Determining Good
Combinations of Queue Priority Rules.” Fourth International Industrial Engineering
Research Conference, Nashville, TN, ed. Schmeiser, B. and R. Uzsoy. Norcross, GA:
IIE, 1995, pp. 874–80.
har01307_ch10_273-298.indd 295 12/17/10 12:02 PM

Brady, T. F., and R. O. Bowden. “The Effectiveness of Generic Optimization Modules

in Computer Simulation Languages.” Proceedings of the Tenth Annual Industrial
Engineering Research Conference, CD-ROM, 2001.
Fogel, D. “Evolving Artificial Intelligence.” Ph.D. Thesis, University of California, 1992.
Glover, F. “Tabu Search: A Tutorial.” Interfaces, 20, no. 4 (1990), pp. 41–52.
Glover, F. “Genetic Algorithms and Scatter Search: Unsuspected Potential,” Statistics and
Computing 4 (1994), pp. 131–40.
Glover, F.; P. Kelly, and M. Laguna. “New Advances and Applications of Combining
Simulation and Optimization.” 1996 Winter Simulation Conference, Coronado, CA:
ed. Charnes, J., Morrice, D., Brunner, D., and J. Swain. Piscataway, NJ: Institute of
Electrical and Electronics Engineers, 1996, pp. 144–52.
Goldberg, D. Genetic Algorithms in Search, Optimization, and Machine Learning. Read-
ing, MA: Addison Wesley, 1989.
Greenwood, A.; L. Rees; and I. Crouch. “Separating the Art and Science of Simulation
Optimization: A Knowledge-Based Architecture Providing for Machine Learning.”
IIE Transactions 25, no. 6 (1993), pp. 70–83.
Hall, J., and R. Bowden. “Simulation Optimization for a Manufacturing Problem.” Pro-
ceedings of the Southeastern Simulation Conference, Huntsville, AL, ed. J. Gauthier.
San Diego, CA: The Society for Computer Simulation, 1996, pp. 135–40.
Hall, J.; R. Bowden; and J. Usher. “Using Evolution Strategies and Simulation to Optimize
a Pull Production System.” Journal of Materials Processing Technology 61 (1996),
pp. 47–52.
Hooke, R., and T. Jeeves. “Direct Search Solution of Numerical Statistical Problems.”
Journal of the Association of Computer Machines 8 (1961), pp. 212–29.
Jacobson, S., and L. Schruben. “Techniques for Simulation Response Optimization.” Op-
erations Research Letters 8 (1989), pp. 1–9.
Lacksonen, T., and P. Anussornnitisarn. “Empirical Comparison of Discrete Event Simu-
lation Optimization Techniques.” Proceedings of the Summer Computer Simulation
Conference, Ottawa, Canada. San Diego, CA: The Society for Computer Simulation,
1995, pp. 69–101.
Law, A. Simulation Modeling and Analysis. New York: McGraw-Hill, 4th ed., 2007.
Markt, P., and M. Mayer. “Witness Simulation Software: A Flexible Suite of Simula-
tion Tools.” 1997 Winter Simulation Conference, Atlanta, GA, ed. Andradóttir, S.,
K. Healy, D. Withers and B. Nelson. Piscataway, NJ: Institute of Electrical and Elec-
tronics Engineers, 1997, pp. 127–33.
Monden, Y. Toyota Production System. Norcross, GA: IIE, 1993.
Nelder, J., and R. Mead. “A Simplex Method for Function Minimization.” Computer Jour-
nal 7 (1965), pp. 308–13.
OptQuest 96 Information Sheet. Boulder, CO: Micro Analysis and Design, Inc., 1996.
Pegden, D., and M. Gately. “Decision Optimization for GASP IV Simulation Models.”
1977 Winter Simulation Conference, Gaitherburg, MD, ed. Highland, H. R. Sargent
and J. Schmidt. Piscataway, NJ: Institute of Electrical and Electronics Engineers,
1977, pp. 127–33.
Pegden, D., and M. Gately. “A Decision-Optimization Module for SLAM.” Simulation 34
(1980), pp. 18–25.
Safizadeh, M. “Optimization in Simulation: Current Issues and the Future Outlook.” Naval
Research Logistics 37 (1990), pp. 807–25.
Schwefel, H.-P. Numerical Optimization of Computer Models. Chichester: John Wiley &
Sons, 1981.
har01307_ch10_273-298.indd 296 12/17/10 12:02 PM

SimRunner Online Software Help. Ypsilanti, MI: Decision Science, Inc., 1996b.
SimRunner User’s Guide ProModel Edition. Ypsilanti, MI: Decision Science, Inc., 1996a.
Stuckman, B.; G. Evans; and M. Mollaghasemi. “Comparison of Global Search Methods
for Design Optimization Using Simulation.” 1991 Winter Simulation Conference,
Phoenix, AZ, ed. Nelson, B., W. Kelton and G. Clark. Piscataway, NJ: Institute of
Electrical and Electronics Engineers, 1991, pp. 937–44.
Tompkins, G., and F. Azadivar. “Genetic Algorithms in Optimizing Simulated Systems.”
1991 Winter Simulation Conference, Arlington, VA, ed. Alexopoulos, C., K. Kang,
W. Lilegdon, and D. Goldsman. Piscataway, NJ: Institute of Electrical and Electronics
Engineers, 1995, pp. 757–62.
Usher, J., and R. Bowden. “The Application of Genetic Algorithms to Operation Sequenc-
ing for Use in Computer-Aided Process Planning.” Computers and Industrial Engi-
neering Journal 30, no. 4 (1996), pp. 999–1013.
har01307_ch10_273-298.indd 297 12/17/10 12:02 PM

har01307_ch10_273-298.indd 298 12/17/10 12:02 PM
C H A P T E R
11 MODELING
MANUFACTURING
SYSTEMS
“We no longer have the luxury of time to tune and debug new manufacturing
systems on the floor, since the expected economic life of a new system, before
major revision will be required, has become frighteningly short.”
—Conway and Maxwell
11.1 Introduction
In Chapter 6 we discussed general procedures for modeling the basic operation of
manufacturing and service systems. In this chapter we discuss design and operat-
ing issues that are more specific to manufacturing systems. Different applications
of simulation in manufacturing are presented together with how specific manufac-
turing issues are addressed in a simulation model. Most manufacturing systems
have material handling systems that, in some instances, have a major impact on
overall system performance. We touch on a few general issues related to material
handling systems in this chapter; however, a more thorough treatment of material
handling systems is given in Chapter 12.
Manufacturing systems are processing systems in which raw materials are
transformed into finished products through a series of operations performed at
workstations. In the rush to get new manufacturing systems on line, engineers
and planners often become overly preoccupied with the processes and methods
without fully considering the overall coordination of system components.
Many layout and improvement decisions in manufacturing are left to chance
or are driven by the latest management fad with little knowledge of how much
improvement will result or whether a decision will result in any improvement
at all. For example, work-in-process (WIP) reduction that is espoused by lean
just-in-time (JIT) often disrupts operations because it merely uncovers the rocks
(variability, long setups, or the like) hidden beneath the inventory water level that
necessitated the WIP in the first place. To accurately predict the effect of lowering
299
har01307_ch11_299-322.indd 299 20/12/10 1:51 PM

WIP levels requires sonar capability. Ideally you would like to identify and re-
move production rocks before arbitrarily lowering inventory levels and exposing
production to these hidden problems. “Unfortunately,” note Hopp and Spearman
(2001), “JIT, as described in the American literature, offers neither sonar (models
that predict the effects of system changes) nor a sense of the relative economics of
level reduction versus rock removal.”
Another popular management technique is the theory of constraints. In this
approach, a constraint or bottleneck is identified and a best-guess solution is imple-
mented, aimed at either eliminating that particular constraint or at least mitigating
the effects of the constraint. The implemented solution is then evaluated and, if
the impact was underestimated, another solution is attempted. As one manufactur-
ing manager expressed, “Contraint-based management can’t quantify investment
justification or develop a remedial action plan” (Berdine 1993). It is merely a
trial-and-error technique in which a best-guess solution is implemented with the
hope that enough improvement is realized to justify the cost of the solution. Even
Deming’s plan–do–check–act (PDCA) cycle of process improvement implicitly
prescribes checking performance after implementation. What the PDCA cycle
lacks is an evaluation step to test or simulate the plan before it is implemented.
While eventually leading to a better solution, this trial-and-error approach ends up
being costly, time-consuming, and disruptive to implement.
In this chapter we address the following topics:
• What are the special characteristics of manufacturing systems?
• What terminology is used to describe manufacturing operations?
• How is simulation applied to manufacturing systems?
• What techniques are used to model manufacturing systems?
11.2 Characteristics of Manufacturing Systems

Manufacturing systems represent a class of processing systems in which entities
(raw materials, components, subassemblies, pallet or container loads, and so on)
are routed through a series of workstations, queues, and storage areas. As opposed
to service systems where the entities are usually human, entities in manufactur-
ing systems are inanimate objects that have a more controlled entry and routing
sequence. Entity production is frequently performed according to schedules to
fill predefined quotas. In other instances, production is triggered by low finished
goods inventories (make-to-stock production) or by customer orders (make/
assemble-to-order production). Manufacturing systems utilize varying degrees of
mechanization and automation for entity processing and material movement. In the
most highly automated facilities, there may be very little, if any, human involve-
ment to be considered in the simulation study. These characteristics have the follow-
ing implications for modeling manufacturing systems:
• Operation times often have little, if any, variability.
• Entity arrivals frequently occur at set times or conditions.
har01307_ch11_299-322.indd 300 20/12/10 1:51 PM

Chapter 11 Modeling Manufacturing Systems 301
• Routing sequences are usually fixed from the start.

• Entities are often processed in batches.
• Equipment reliability is frequently a key factor.
• Material handling can significantly impact entity flow.
• We are usually interested in the steady-state behavior of the system.
Of course, actual modeling considerations will vary depending on the type of
system being modeled and the purpose of the simulation.
11.3 Manufacturing Terminology

Manufacturing systems share many common elements and issues. They also share
a vernacular that can seem foreign to those unfamiliar with manufacturing activi-
ties. Understanding how to model manufacturing systems requires a knowledge
of the terminology that is used in manufacturing industries. Here are some of the
terms that are useful to know when simulating manufacturing systems:
Operation—An operation is the activity performed on a product at a
workstation. Operations are usually performed to produce some physical
change to a part such as machining or molding. Assembly operations
combine component parts. Nontransformational operations such as
inspection or testing are also frequently performed.
Workstation—A workstation or work center is the place or location where
an operation is performed. A workstation consists of one or more machines
and/or personnel.
NC machine—An NC (numerically controlled) machine is a machine
tool whose spindle and table action are automatically controlled by a
computer.
Machining center—A machining center is essentially an NC machine
with a tool exchanger that provides greater capability than a stand-alone
NC machine. Machining centers have been defined as “multifunctional,
numerically controlled machine tools with automatic tool changing
capabilities and rotating cutting tools that permit the programmed,
automated production of a wide variety of parts” (Wick 1987).
Master production schedule—The master production schedule defines what
end products are to be produced for a given period, such as the number of
office tables to produce for a week.
Production plan—The production plan is a detailed schedule of production
for each individual component comprising each end product.
Bottleneck—The bottleneck is traditionally thought of as the workstation
that has the highest utilization or largest ratio of required processing time
to time available. In a broader sense, a bottleneck may be any constraint
or limiting factor (an operator, a conveyor system) in the production of
goods.
har01307_ch11_299-322.indd 301 20/12/10 1:51 PM

Setup—Setup is the activity required to prepare a workstation to produce a

different part type.
Job—Depending on the industry, a job may be any of the following:
• In a general sense, a job refers to the activity or task being
performed.
• In mass production, a job refers to each individual part.
• In a job shop, a job refers to a customer order to be produced.
Machine cycle time—Machine cycle time is the time required to perform

a single operation. If load and unload times are included, it is sometimes
referred to as floor-to-floor time.
Capacity—Capacity refers to either holding capacity, as in the case of a
tank or storage bin, or production capacity, as in the case of a machine.
Production capacity is usually expressed in one of the following four
ways:
• Theoretical capacity is the rate at which a machine is capable of
producing if it were in constant production, assuming 100 percent
efficiency.
• Effective capacity is the actual capacity factored by the reliability of the
machine with allowances for load and unload, chip removal, and so forth.
• Expected capacity is the effective capacity of the machine factored
by anticipated allowances for system influences such as shortages and
blockages (this is what simulation can tell you).
• Actual capacity is what the machine actually produces.
Scrap rate—Scrap rate is the percentage of defective parts that are removed
from the system following an operation. The opposite of scrap rate is yield.
Reliability—Reliability is usually measured in terms of mean time
between failures (MTBF), which is the average time a machine or piece of
equipment is in operation before it fails.
Maintainability—Maintainability is usually measured in terms of mean time
to repair (MTTR) and is the average time required to repair a machine or
piece of equipment whenever it fails.
Availability—Availability is the percentage of total scheduled time that a
resource is actually available for production. Availability is uptime, which is
a function of reliability and maintainability.
Preventive or scheduled maintenance—Preventive or scheduled
maintenance is periodic maintenance (lubrication, cleaning, or the like)
performed on equipment to keep it in running condition.
Unit load—A unit load is a consolidated group of parts that is containerized
or palletized for movement through the system. The idea of a unit load is to
minimize handling through consolidation and provide a standardized pallet
or container as the movement item.
har01307_ch11_299-322.indd 302 20/12/10 1:51 PM

11.4 Use of Simulation in Manufacturing

Simulation has proved effective in helping to sort through the complex issues sur-
rounding manufacturing decisions. Kochan (1986) notes that, in manufacturing
systems,
The possible permutations and combinations of workpieces, tools, pallets, transport
vehicles, transport routes, operations, etc., and their resulting performance, are almost
endless. Computer simulation has become an absolute necessity in the design of prac-
tical systems, and the trend toward broadening its capabilities is continuing as systems
move to encompass more and more of the factory.
As companies move toward greater vertical and horizontal integration and look
for ways to improve the entire value chain, simulation will continue to be an
essential tool for effectively planning the production and delivery processes.
Simulation in manufacturing covers the range from real-time cell control to
long-range technology assessment, where it is used to assess the feasibility of
new technologies prior to committing capital funds and corporate resources. Fig-
ure 11.1 illustrates this broad range of planning horizons.
Simulations used to make short-term decisions usually require more detailed
models with a closer resemblance to current operations than what would be found in
long-term decision-making models. Sometimes the model is an exact replica of the
current system and even captures the current state of the system. This is true in real-
time control and detailed scheduling applications. As simulation is used for more
long-term decisions, the models may have little or no resemblance to current opera-
tions. The model resolution becomes coarser, usually because higher-level decisions
are being made and data are too fuzzy and unreliable that far out into the future.
Simulation helps evaluate the performance of alternative designs and the ef-
fectiveness of alternative operating policies. A list of typical design and opera-
tional decisions for which simulation might be used in manufacturing includes the
following:
Design Decisions
1. What type and quantity of machines, equipment, and tooling should
be used?
2. How many operating personnel are needed?
3. What is the production capability (throughput rate) for a given
configuration?
FIGURE 11.1
Decision range
in manufacturing
simulation.
Cell Job Resource Production Process System Technology
manage- sequen- loading scheduling change configu- assess-
ment cing studies ration ment
Seconds Hours Days Weeks Months 1–2 years 2–5 years
har01307_ch11_299-322.indd 303 20/12/10 1:51 PM

4. What type and size of material handling system should be used?

5. How large should buffer and storage areas be?
6. What is the best layout of the factory?
7. What automation controls work the best?
8. What is the optimum unit load size?
9. What methods and levels of automation work best?
10. How balanced is a production line?
11. Where are the bottlenecks (bottleneck analysis)?
12. What is the impact of machine downtime on production (reliability
analysis)?
13. What is the effect of setup time on production?
14. Should storage be centralized or localized?
15. What is the effect of vehicle or conveyor speed on part flow?
Operational Decisions
1. What is the best way to schedule preventive maintenance?
2. How many shifts are needed to meet production requirements?
3. What is the optimum production batch size?
4. What is the optimum sequence for producing a set of jobs?
5. What is the best way to allocate resources for a particular set of tasks?
6. What is the effect of a preventive maintenance policy as opposed to a
corrective maintenance policy?
7. How much time does a particular job spend in a system (makespan or
throughput time analysis)?
8. What is the best production control method (kanban, MRP, or other)
to use?
In addition to being used for specific design or operational decisions, simula-

tion has been used to derive general design rules so that the most cost-effective
decisions can be made for a range of operating conditions. One study, for example,
used simulation to estimate how the output capacity of a manufacturing system
changes as the various components of the system are increased or improved for
a known cost. These data were then used to develop cost versus output curves
for a given part or product mix. By comparing these curves for different types of
production systems, it was possible to identify break-even points where a particu-
lar type of system was economically preferred over other systems for particular
products or product mixes (Renbold and Dillmann 1986).
11.5 Applications of Simulation in Manufacturing

There are numerous applications for simulation in manufacturing. Many of them are
the same as those found in any type of system, such as capacity analysis and cycle
time reduction. Applications that tend to be more specific to manufacturing include
har01307_ch11_299-322.indd 304 20/12/10 1:51 PM

• Methods analysis.
• Plant layout.
• Batch sizing.
• Production control.
• Inventory control.
• Supply chain planning.
• Production scheduling.
• Real-time control.
• Emulation.
Each of these application areas is discussed here.
11.5.1 Methods Analysis

Methods analysis looks at alternative ways of processing and handling material.
An assembly process, for example, can be performed all at a single operation
with several parallel stations or be broken down into multiple stages performed
on a single assembly line. Production and assembly lines may be either paced, in
which product movement occurs at a fixed rate with the operator keeping pace,
or unpaced, in which the rate of flow is determined by the speed of the worker.
The relative performance of these alternative methods of production can be easily
demonstrated using simulation.
One of the decisions that addresses manufacturing methods pertains to the
type and level of automation to use in the factory. Simulation has been especially
helpful in designing automated systems to safeguard against suboptimal perfor-
mance. As noted by Glenney and Mackulak (1985), “Computer simulation will
permit evaluation of new automation alternatives and eliminate the problem of
building a six-lane highway that terminates onto a single-lane dirt road.” Often an
unbalanced design occurs when the overall performance is overlooked because of
preoccupation with a single element. “Islands of automation” that are not properly
integrated and balanced lead to suboptimal performance. For example, an auto-
mated machine capable of producing 200 parts per hour fed by a manual handler
at the rate of 50 parts per hour will produce only 50 parts per hour. The expression
that “a chain is no stronger than its weakest link” has particular application to
automated manufacturing systems.
There are three types of automation to consider in manufacturing: operation
automation, material handling automation, and information automation. The de-
gree of automation ranges from manual to fully automatic. Figure 11.2 shows the
types and degrees of automation.
Although information flow is frequently overlooked in manufacturing simu-
lations, it may be a contributing factor to the overall system performance and
therefore may merit some analysis. As noted in a Manufacturing Engineering
article (1993), “It’s easier to make the product than to adjust the paper and data
flow required to track the product.”
har01307_ch11_299-322.indd 305 20/12/10 1:51 PM

FIGURE 11.2 Automated

Axes of automation.
Source: Adapted from
G. Boehlert and W. J.
Trybula, “Successful
Factory Automation,” IEEE
Transactions, September
1984, p. 218.
Automated
Operations
Semiautomated
n
io
at
rm Semiautomated
fo
In
Manual Manual Material handling

Manual Semiautomated Automated
Automation itself does not solve performance problems. If automation is un-

suited for the application, poorly designed, or improperly implemented and inte-
grated, then the system is not likely to succeed. The best approach to automation
is to first simplify, then systematize, and finally automate. Some companies find
that after simplifying and systematizing their processes, the perceived need to
automate disappears.
11.5.2 Plant Layout

With the capital-intensive nature of manufacturing systems, plant layout is im-
portant to the smooth flow of product and movement of resources through the
facility. A good layout results in a streamlined flow with minimal movement and
thus minimizes material handling and storage costs. Even existing systems can
benefit from a layout analysis. Layouts tend to evolve haphazardly over time
with little thought given to work flow. This results in nonuniform, multidirec-
tional flow patterns that cause management and material handling problems.
Simulation helps to identify inefficient flow patterns and to create better layouts
to economize the flow of material.
In planning the layout of a manufacturing system, it is important to have a
total system perspective of material flow. Flow within the manufacturing opera-
tion is often hierarchical, beginning with the overall flow from receiving to manu-
facturing to shipping. At the lowest level, parts might flow from an input buffer to
a machine (see Figure 11.3). Any or all of the levels shown in Figure 11.3 may be
present within a plant. Interaction points occur between levels as well as within
each level. When modeling material flow in a factory it is important to consider
the level of flow and how flows interact.
har01307_ch11_299-322.indd 306 20/12/10 1:51 PM

FIGURE 11.3
Material flow system. Receiving Manufacturing Shipping
Department A Department B Department C
Cell A Cell B Cell C
Station A Station B Station C
Input buffer Machine Output buffer
In laying out equipment in a facility, the designer must choose between a

process layout, a product layout, and a part family layout of machines. In a pro-
cess layout, machines are arranged together based on the process they perform.
Machines performing like processes are grouped together. In a product layout,
machines are arranged according to the sequence of operations that are performed
on a product. This results in a serial flow of the product from machine to machine.
In a part family layout, machines are arranged to process parts of the same family.
Part family layouts are also called group technology cells because group technol-
ogy is used to group parts into families. Figure 11.4 illustrates these three different
types of arrangements and their implications on material flow.
Deciding which layout to use will depend largely on the variety and lot size of
parts that are produced. If a wide variety of parts are produced in small lot sizes,
a process layout may be the most appropriate. If the variety is similar enough in
processing requirements to allow grouping into part families, a cell layout is best.
If the variety is small enough and the volume is sufficiently large, a product layout
har01307_ch11_299-322.indd 307 20/12/10 1:51 PM

FIGURE 11.4
Comparison between Part
(a) process layout, A
(b) product layout, and
(c) cell layout. (a) Process
layout
Part
B
Single (b) Product

product layout
(c) Cell
layout
Family 1 Family 2
is best. Often a combination of topologies is used within the same facility to ac-
commodate mixed requirements. For example, a facility might be predominately
a job shop with a few manufacturing cells. In general, the more the topology
resembles a product layout where the flow can be streamlined, the greater the ef-
ficiencies that can be achieved.
Perhaps the greatest impact simulation can have on plant layout comes from
designing the process itself, before the layout is even planned. Because the pro-
cess plan and method selection provide the basis for designing the layout, hav-
ing a well-designed process means that the layout will likely not require major
changes once the best layout is found for the optimized process.
11.5.3 Batch Sizing

Batching decisions play a major role in meeting the flow and efficiency goals of
a production facility. A batch or lot of parts refers to a quantity of parts grouped
together for some particular purpose. Typically three different batch types are
spoken of in manufacturing: production batch, move batch, and process batch.
The production batch, also referred to as the production lot, consists of all of
the parts of one type that begin production before a new part type is introduced to
the system. The move or transfer batch is the collection of parts that are grouped
har01307_ch11_299-322.indd 308 20/12/10 1:51 PM

FIGURE 11.5
To
Illustration of production
(a) production batch,
(b) move batch, and
(c) process batch.
(a) Production batch
(c) Process batch
Station Station Station

1 2 3
(b) Move batch
and moved together from one workstation to another. A production batch is often
broken down into smaller move batches. This practice is called lot splitting. The
move batch need not be constant from location to location. In some batch manu-
facturing systems, for example, a technique called overlapped production is used
to minimize machine idle time and reduce work in process. In overlapped produc-
tion, a move batch arrives at a workstation where parts are individually processed.
Then instead of accumulating the entire batch before moving on, parts are sent on
individually or in smaller quantities to prevent the next workstation from being idle
while waiting for the entire batch. The process batch is the quantity of parts that are
processed simultaneously at a particular operation and usually consists of a single
part. The relationship between these batch types is illustrated in Figure 11.5.
Deciding which size to use for each particular batch type is usually based on
economic trade-offs between in-process inventory costs and economies of scale as-
sociated with larger batch sizes. Larger batch sizes usually result in lower setup costs,
handling costs, and processing costs. Several commands are provided in ProModel
for modeling batching operations such as GROUP, COMBINE, JOIN and LOAD.
11.5.4 Production Control

Production control governs the flow of material between individual workstations.
Simulation helps plan the most effective and efficient method for controlling ma-
terial flow. Several control methods exist, any of which may work well depending
on the situation. Common methods include
• Push control.
• Pull control.
• Drum–buffer–rope (DBR) control.
har01307_ch11_299-322.indd 309 20/12/10 1:51 PM

Push Control
A push system is one in which production is driven by workstation capacity and
material availability. Each workstation seeks to produce as much as it can, push-
ing finished work forward to the next workstation. In cases where the system is
unconstrained by demand (the demand exceeds system capacity), material can be
pushed without restraint. Usually, however, there is some synchronization of push
with demand. In make-to-stock production, the triggering mechanism is a drop
in finished goods inventory. In make-to-order or assemble-to-order production, a
master production schedule drives production by scheduling through a material
requirements planning (MRP) or other backward or forward scheduling system.
MRP systems determine how much each station should produce for a given
period. Unfortunately, once planned, MRP is not designed to respond to disrup-
tions and breakdowns that occur for that period. Consequently, stations continue
to produce inventory as planned, regardless of whether downstream stations can
absorb the inventory. Push systems such as those resulting from MRP tend to
build up work in process (WIP), creating high inventory-carrying costs and long
flow times.
Pull Control
At the other extreme of a push system is a pull system, in which downstream de-
mand triggers each preceding station to produce a part with no more than one or
two parts at a station at any given time (see Figure 11.6). Pull systems are often
associated with the just-in-time (JIT) or lean manufacturing philosophy, which
advocates the reduction of inventories to a minimum. The basic principle of JIT
is to continuously reduce scrap, lot sizes, inventory, and throughput time as well
as eliminate the waste associated with non–value added activities such as material
handling, storage, machine setup, and rework.
FIGURE 11.6
Push versus pull
system.
1 2 3
(a) Push system

1 2 3
(b) Pull system
har01307_ch11_299-322.indd 310 20/12/10 1:51 PM

A pull system is generally implemented using a method the Japanese call

kanban (literally meaning a visual sign or record) often implemented using cards
or tickets. When an operation is completed at the final location and exits the loca-
tion, a new entity or part is pulled from the preceding WIP buffer. A kanban card
accompanies each part bin in the buffer and is returned to the preceding station
when the bin moves on. The returned card signals the preceding operation to pull
a unit from its own WIP and replenish the kanban. A kanban system is imple-
mented using either a withdrawal kanban, authorizing the retrieval of work items,
or a production kanban, authorizing the production of more parts.
When modeling a kanban control system, it usually isn’t necessary to explic-
itly model the cards or actual triggering mechanism used (some systems use golf
balls, bins, or even a painted area on the floor). For most JIT systems where flow
is serial, a kanban system is easily modeled by simply limiting the capacity of
the workstation input and output buffers. A two-bin kanban system, for example,
can be modeled by limiting the input buffer capacity of the workstation to two.
Ultimately, a kanban system is just a way to limit queue capacities.
Kanban is a queue limiter; that is, it tightly links a provider’s service with a user’s
need. Thus, it limits the queue of work in front of the provider (the quantity/amount of
items waiting and the time each item waits). Queue limitation replaces lax flow con-
trol with discipline, which results in customer satisfaction and competitive advantage
(Schonberger and Knod 1994).
Because ProModel uses capacity limitations to control the flow of entities,
entities will not advance to the next station until there is available capacity at that
location. This effectively provides a built-in pull system in which the emptying of
a buffer or workstation signals the next waiting entity to advance (push systems
actually work the same way, only buffers effectively have infinite capacity). Sev-
eral papers have been published (Mitra and Mitrani 1990; Wang and Wang 1990)
showing the use of limited buffers to model JIT systems.
For certain pull systems, triggers signal the release of WIP further upstream than
just one station. An example of this is another form of pull control called CONWIP
(Hopp and Spearman 2001). A CONWIP (CONstant Work-In-Process) system is a
form of pull system that pulls from the final operation. The completion of a part trig-
gers the release of a new part to the beginning of the line. A CONWIP system limits
work in process just like a kanban system. A new part is introduced to the system
only when a part leaves the system, thus maintaining a constant inventory level. To
simulate this process in ProModel, an ORDER statement should be used whenever
an entity exits the system. This signals the creation of a new entity at the beginning
of the line so that each part that exits the system releases a new part into the system.
CONWIP has the advantage over kanban of being easier to administer while still
achieving nearly the same, and sometimes even superior, results.
Drum–Buffer–Rope (DBR) Control

One approach to production control utilizes a drum–buffer–rope (DBR) meta-
phor. This technique is based on the theory of constraints and seeks to improve
productivity by managing the bottleneck operation. The DBR approach basically
lets the bottleneck operation pace the production rate by acting as a drum. The
har01307_ch11_299-322.indd 311 20/12/10 1:51 PM

buffer is the inventory needed in front of the bottleneck operation to ensure that
it is never starved for parts. The rope represents the tying or synchronizing of the
production release rate to the rate of the bottleneck operation.
As noted earlier in this chapter, a weakness in constraint-based management
is the inability to predict the level of improvement that can be achieved by elimi-
nating a bottleneck. Simulation can provide this evaluative capability missing in
constraint-based management. It can also help detect floating bottlenecks for situa-
tions where the bottleneck changes depending on product mix and production rates.
Simulating DBR control can help determine where bottlenecks are and how much
buffer is needed prior to bottleneck operations to achieve optimum results.
11.5.5 Inventory Control

Inventory control involves the planning, scheduling, and dispatching of inventory
to support the production activity. Because of the high dollar impact of inventory
decisions, developing and maintaining a good inventory control system are critical.
Some common goals associated with inventory control at all stages in the supply
chain include
• Responding promptly to customer demands.
• Minimizing order and setup costs associated with supplying inventory.
• Minimizing the amount of inventory on hand as well as all associated
inventory carrying costs.
Decisions affecting inventory levels and costs include
• How much inventory should be ordered when inventory is replenished
(the economic order quantity)?
• When should more inventory be ordered (the reorder point)?
• What are the optimum inventory levels to maintain?
There are two common approaches to deciding when to order more inventory and
how much: the reorder point system and the MRP system. A reorder point system is
an inventory control system where inventory for a particular item is replenished to a
certain level whenever it falls below a defined level. MRP, on the other hand, is a bit
more complicated in that it is time phased. This means that inventory is ordered to
match the timing of the demand for products over the next planning horizon.
Inventory management must deal with random depletion or usage times and
random lead times, both of which are easily handled in simulation. To model an
inventory control system, one need only define the usage pattern, the demand pat-
tern, and the usage policies (priorities of certain demands, batch or incremental
commitment of inventories, and so on). The demand rate can be modeled using
distributions of interarrival times. Replenishment of raw materials should be
made using an ORDER command that is able to generate additional inventory when
it drops below a particular level.
Running the simulation can reveal the detailed rises and drops in inventory
levels as well as inventory summary statistics (average levels, total stockout time,
har01307_ch11_299-322.indd 312 20/12/10 1:51 PM

and so forth). The modeler can experiment with different replenishment strategies
and usage policies until the best plan is found that meets the established criteria.
Using simulation modeling over traditional, analytic modeling for inventory
planning provides several benefits (Browne 1994):
• Greater accuracy because actual, observed demand patterns and irregular
inventory replenishments can be modeled.
• Greater flexibility because the model can be tailored to fit the situation
rather than forcing the situation to fit the limitations of an existing model.
• Easier to model because complex formulas that attempt to capture the
entire problem are replaced with simple arithmetic expressions describing
basic cause-and-effect relationships.
• Easier to understand because demand patterns and usage conditions are
more descriptive of how the inventory control system actually works.
Results reflect information similar to what would be observed from
operating the actual inventory control system.
• More informative output that shows the dynamics of inventory conditions
over time and provides a summary of supply, demand, and shortages.
• More suitable for management because it provides “what if ” capability so
alternative scenarios can be evaluated and compared. Charts and graphs
are provided that management can readily understand.
Centralized versus Decentralized Storage. Another inventory-related issue has

to do with the positioning of inventory. Traditionally, inventory was placed in
centralized locations for better control. Unfortunately, centralized inventory cre-
ates excessive handling of material and increased response times. More recent
trends are toward decentralized, point-of-use storage, with parts kept where they
are needed. This eliminates needless handling and dramatically reduces response
time. Simulation can effectively assess which method works best and where to
strategically place storage areas to support production requirements.
11.5.6 Supply Chain Management

Supply chain management is a key area for improving efficiencies in a corpo-
ration. A supply chain is made up of suppliers, manufacturers, distributors, and
customers. There are two main goals with respect to supply chain management:
to reduce costs and to improve customer service, which is achieved by reducing
delivery time. Supply chain network analysis tools exist that can look at the prob-
lem in the aggregate and help identify optimum site locations and inventory levels
given a general demand. Simulation, on the other hand, can look at the system on
a more detailed level, analyzing the impact of interdependencies and variation
over time. Simulation “accurately traces each transaction through the network at
a detailed level, incorporating cost and service impacts of changes in inventory
policies, shipment policies, uncertainty in demand, and other time-dependent
factors” (Napolitano 1997).
har01307_ch11_299-322.indd 313 20/12/10 1:51 PM

Simulation of flow between facilities is similar to the simulation of flow

within a facility. One of the most difficult challenges in modeling a supply chain is
making the flow of entities in the model demand-driven. This is done by using the
ORDER statement, with entities representing the orders waiting to be filled. Often
order or customer consolidation must also be taken into account, which requires
some data management using variables and attributes.
11.5.7 Production Scheduling

Production scheduling determines the start and finish times for jobs to be produced.
The use of simulation for production scheduling is called simulation-based sched-
uling. Some scheduling is static and determined ahead of time. Other scheduling
is performed dynamically, such as sequencing a set of jobs currently waiting to be
processed at a particular machine. Simulation can help make both types of sched-
uling decisions and can assess any scheduling rules defined by the modeler. Once
the simulation is run, a report can be generated showing the start and finish times
of each job on each machine. Simulation does not set the schedule itself, but rather
reports the schedule resulting from a given set of operating rules and conditions.
Job sequencing is a frequent challenge for job shops, which tend to be very
due-date-driven in their scheduling and are often faced with the question “In what
sequence should jobs be processed so as to minimize lateness or tardiness?” Sup-
pose, for example, that three jobs are waiting to be processed, each requiring the
same resources but in a different sequence. If the goal is to process the jobs in an
order that will minimize total production time yet still permit on-time delivery,
several simulation runs can be made just before production to find the optimal
sequence in which the jobs should be processed.
In scheduling, we usually try to achieve four principal goals or objectives
(Kiran and Smith 1984):
1. Minimize job lateness or tardiness.
2. Minimize the flow time or time jobs spend in production.
3. Maximize resource utilization.
4. Minimize production costs.
The decision of which job to process next is usually based one of the following
rules:
Rule Definition
Shortest processing time Select the job having the least processing time.
Earliest due date Select the job that is due the soonest.
First-come, first-served Select the job that has been waiting the longest for this
workstation.
First-in-system, first-served Select the job that has been in the shop the longest.
Slack per remaining operations Select the job with the smallest ratio of slack to operations
remaining to be performed.
har01307_ch11_299-322.indd 314 20/12/10 1:51 PM

Traditional scheduling methods fail to account for fluctuations in quantities

and resource capacity. This variation can be built into simulation models. How-
ever, in using simulation for dynamic scheduling, we are dealing with short time
periods and are therefore less concerned about long-term statistical fluctuations in
the system such as machine downtimes. In such instances, typically a determinis-
tic simulation is used that is based on expected operation times. Models built for
simulation-based scheduling have the following characteristics:
• The model captures an initial state.
• Operation times are usually based on expected times.
• Anomaly conditions (machine failures and the like) are ignored.
• The simulation is run only until the required production has been met.
Simulation can work with material requirements planning (MRP), advanced
planning and scheduling (APS), and manufacturing execution systems (MES) to
simulate the plans generated by these systems in order to test their viability. It can
also produce a detailed schedule showing which resources were used and at what
times. This information can then be used for resource scheduling.
It usually isn’t practical to simulate an entire MRP system because the num-
ber of different entity types becomes too unmanageable. You should consider
building a model based only on critical product components. You can forget the
small inexpensive items like nuts and bolts, which are usually managed using a
reorder point system, and focus on the few components that constitute the major
scheduling challenge. This could provide a reasonably simple model that could be
used regularly to test MRP-generated requirements. Such a model could then be used
to achieve true, closed-loop MRP, in which production schedules could be itera-
tively generated and evaluated until the best schedule was found for the available
capacity of the system.
11.5.8 Real-Time Control

During actual production, simulation has been integrated with manufacturing
cells to perform real-time analysis such as next task selection and dynamic rout-
ing decisions. Initially, a model is built to operate with a simulated version of
the cell, and then the simulated cell components are replaced with the actual cell
components. Three benefits of using the simulation for cell management are
• Logic used in the simulator does not need to be recoded for the controller.
• Animation capabilities allow processes to be monitored on the layout.
• Statistical-gathering capabilities of simulation automatically provide
statistics on selected measures of performance.
11.5.9 Emulation
A special use of simulation in manufacturing, particularly in automated systems,
has been in the area of hardware emulation. As an emulator, simulation takes
har01307_ch11_299-322.indd 315 20/12/10 1:51 PM

inputs from the actual control system (such as programmable controllers or mi-
crocomputers), mimics the behavior that would take place in the actual system,
and then provides feedback signals to the control system. The feedback signals
are synthetically created rather than coming from the actual hardware devices.
In using simulation for hardware emulation, the control system is essentially
plugged into the model instead of the actual system. The hardware devices are
then emulated in the simulation model. In this way simulation is used to test,
debug, and even refine the actual control system before any or all of the hardware
has been installed. Emulation can significantly reduce the time to start up new
systems and implement changes to automation.
Emulation has the same real-time requirements as when simulation is used
for real-time control. The simulation clock must be synchronized with the real
clock to mimic actual machining and move times.
11.6 Manufacturing Modeling Techniques

In Chapter 6 we presented modeling techniques that apply to a broad range of
systems, including manufacturing systems. Here we discuss a few modeling tech-
niques that are more specific to manufacturing systems.
11.6.1 Modeling Machine Setup

Manufacturing systems often consist of processing equipment that has particular
settings when working on a particular job. A machine setup or changeover occurs
when tooling or other machine adjustments must be made to accommodate the next
item to be processed. Machine setups usually occur only when there is a change of
part type. For example, after running a lot of 200 parts of type A, the machine is
retooled and adjusted for producing a lot of part type B. In some situations, adjust-
ments must be made between processing parts of the same type. When setup is
performed between different part types, the time and resources required to perform
the setup may depend on both the part type of the new lot and the part type of the
previous lot. For example, the setup time when changing over from part A to part B
may be different than the setup time when changing over from part B to part C.
Setups are easily modeled in ProModel by selecting the Setup option under
the Downtime menu for a location record. Setups can be defined to occur after
every part or only after a change of part type.
11.6.2 Modeling Machine Load and Unload Time

Sometimes a key issue in modeling manufacturing systems is machine load and
unload time. There are several approaches to dealing with load/unload activities:
• Ignore them. If the times are extremely short relative to the operation
time, it may be safe to ignore load and unload times.
• Combine them with the operation time. If loading and unloading the
machine are simply an extension of the operation itself, the load and
har01307_ch11_299-322.indd 316 20/12/10 1:51 PM

unload times can be tacked onto the operation time. A precaution here
is in semiautomatic machines where an operator is required to load and/
or unload the machine but is not required to operate the machine. If the
operator is nearly always available, or if the load and unload activities are
automated, this may not be a problem.
• Model them as a movement or handling activity. If, as just described, an
operator is required to load and unload the machine but the operator is
not always available, the load and unload activities should be modeled
as a separate move activity to and from the location. To be accurate, the
part should not enter or leave the machine until the operator is available.
In ProModel, this would be defined as a routing from the input buffer
to the machine and then from the machine to the output buffer using the
operator as the movement resource.
11.6.3 Modeling Rework and Scrap

It is not uncommon in manufacturing systems to have defective items that require
special treatment. A defective item is usually detected either at an inspection sta-
tion or at the station that created the defect. Sometimes part defects are discovered
at a later downstream operation where, for example, the part can’t be assembled.
In this case, the downstream workstation doubles as a sort of inspection station.
When a defective item is encountered, a decision is usually made to
• Reject the part (in which case it exits the system). This is done by simply
defining a probabilistic routing to EXIT.
• Correct the defect at the current station. To model this approach, simply
add an additional WAIT statement to the operation logic that is executed the
percentage of the time defects are expected to occur.
• Route the part to an offline rework station. This is modeled using a
probabilistic routing to the rework station.
• Recycle the part through the station that caused the defect. This is also
modeled using a probabilistic routing back to the station that caused
the defect. Sometimes an entity attribute needs to be used to designate
whether this is a first pass part or a rework part. This is needed when
reworked parts require a different operation time or if they have a
different routing.
The first consideration when modeling rework operations is to determine
whether they should be included in the model at all. For example, if a part rejec-
tion rate is less than 1 percent, it may not be worth including in the model.
11.6.4 Modeling Transfer Machines

Transfer machines are machines that have multiple serial stations that all transfer
their work forward in a synchronized fashion. Transfer machines may be either
rotary or in-line. On a rotary transfer machine, also called a dial-index machine,
the load and unload stations are located next to each other and parts are usually
har01307_ch11_299-322.indd 317 20/12/10 1:51 PM

FIGURE 11.7
Transfer line system.
Machine Buffer Station
Line
placed directly on the station fixture. In-line transfer machines have a load station
at one end and an unload station at the other end. Some in-line transfer machines
are coupled together to form a transfer line (see Figure 11.7) in which parts are
placed on system pallets. This provides buffering (nonsynchronization) and even
recirculation of pallets if the line is a closed loop.
Another issue is finding the optimum number of pallets in a closed, non-
synchronous pallet system. Such a system is characterized by a fixed number
of pallets that continually recirculate through the system. Obviously, the system
should have enough pallets to at least fill every workstation in the system but not
so many that they fill every position in the line (this would result in a gridlock).
Generally, productivity increases as the number of pallets increases up to a certain
point, beyond which productivity levels off and then actually begins to decrease.
Studies have shown that the optimal point tends to be close to the sum of all of the
workstation positions plus one-half of the buffer pallet positions.
A typical analysis might be to find the necessary buffer sizes to ensure that
the system is unaffected by individual failures at least 95 percent of the time. A
similar study might find the necessary buffer sizes to protect the operation against
the longest tool change time of a downstream operation.
Stations may be modeled individually or collectively, depending on the level
of detail required in the model. Often a series of stations can be modeled as a
single location. Operations in a transfer machine can be modeled as a simple
operation time if an entire machine or block of synchronous stations is modeled
as a single location. Otherwise, the operation time specification is a bit tricky
because it depends on all stations finishing their operation at the same time. One
might initially be inclined simply to assign the time of the slowest operation to
every station. Unfortunately, this does not account for the synchronization of
operations. Usually, synchronization requires a timer to be set up for each station
that represents the operation for all stations. In ProModel this is done by defining
an activated subroutine that increments a global variable representing the cycle
completion after waiting for the cycle time. Entities at each station wait for the
variable to be incremented using a WAIT UNTIL statement.
har01307_ch11_299-322.indd 318 20/12/10 1:51 PM

11.6.5 Continuous Process Systems

A continuous process system manufactures product, but in a different way than
discrete part manufacturing systems. Continuous processing involves the produc-
tion of bulk substances or materials such as chemicals, liquids, plastics, metals,
textiles, and paper (see Figure 11.8).
Pritsker (1986) presents a simulation model in which the level of oil in a stor-
age tank varies depending on whether oil is being pumped into or out of the tank.
If the rate of flow into or out of the tank is constant, then the change in level is the
rate of flow multiplied by the length of time. If, however, the rate of flow varies
continuously, it is necessary to integrate the function defining the rate of flow over
the time interval concerned (see Figure 11.9).
ProModel provides a tank construct and a set of routines for modeling sys-
tems in process industries. Both constant and variable flow rates can be defined.
The routines provided are ideally suited for modeling combined discrete and con-
tinuous activities.
FIGURE 11.8
Continuous flow
system.
FIGURE 11.9 Rate of flow Rate of flow

Quantity of flow when f f
(a) rate is constant and
(b) rate is changing.
F F
t1 t2 t1 t2
Time Time
t1
Q F * (t2 t1) Q =
∫t2
f dt
(a) (b)
har01307_ch11_299-322.indd 319 20/12/10 1:51 PM

11.7 Summary
In this chapter we focused on the issues and techniques for modeling manufactur-
ing systems. Terminology common to manufacturing systems was presented and
design issues were discussed. Different applications of simulation in manufactur-
ing were described with examples of each. Issues related to modeling each type of
system were explained and suggestions offered on how system elements for each
system type might be represented in a model.

1. List four characteristics of manufacturing systems that make them
different from most service systems in the way they are simulated.
2. What is the range of decision making for which simulation is used in
manufacturing systems?
3. How can simulation help in planning the layout of a manufacturing
facility?
4. In modeling automation alternatives, how can simulation reduce the risk of
“building a six-lane highway that terminates into a single-lane dirt road”?
5. Conceptually, how would you model a pull system where the next
downstream operation is pulling production? Where the completion of
a final product triggers the start of a new product?
6. What are the characteristics of a simulation model that is used for
production scheduling?
7. How can emulation help test an actual control system?
8. Why would it be invalid to model a synchronous transfer machine by
simply giving all of the stations the same operation time?
9. In ProModel, how would you model an operation with an 85 percent
yield where defective items are routed to a rework station?
10. In ProModel, how would you model a system where items are moved
in batches of 50, but processed individually at each station?
References
Askin, Ronald G., and Charles R. Standridge. Modeling and Analysis of Manufacturing
Systems. New York: John Wiley & Sons, 1993.
Berdine, Robert A. “FMS: Fumbled Manufacturing Startups?” Manufacturing Engineer-
ing, July 1993, p. 104.
Bevans, J. P. “First, Choose an FMS Simulator.” American Machinist, May 1982,
pp. 144–45.
Blackburn, J., and R. Millen. “Perspectives on Flexibility in Manufacturing: Hardware
versus Software.” In Modelling and Design of Flexible Manufacturing Systems, ed.
Andrew Kusiak, pp. 99–109. Amsterdam: Elsevier, 1986.
har01307_ch11_299-322.indd 320 20/12/10 1:51 PM

Boehlert, G., and W. J. Trybula. “Successful Factory Automation.” IEEE Transactions,

September 1984, p. 218.
Bozer, Y. A., and J. A. White. “Travel-Time Models for Automated Storage/Retrieval
Systems.” IIE Transactions, December 1984, pp. 329–38.
Browne, Jim. “Analyzing the Dynamics of Supply and Demand for Goods and Services.”
Industrial Engineering, June 1994, pp. 18–19.
Cutkosky, M.; P. Fussell; and R. Milligan Jr. “The Design of a Flexible Machining Cell for
Small Batch Production.” Journal of Manufacturing Systems 3, no. 1 (1984), p. 39.
Fitzgerald, K. R. “How to Estimate the Number of AGVs You Need.” Modern Materials
Handling, October 1985, p. 79.
Gayman, David J. “Carving Out a Legend.” Manufacturing Engineering, March 1988,
pp. 75–79.
Glenney, N. E., and G. T. Mackulak. “Modeling and Simulation Provide Key to CIM
Implementation Philosophy.” Industrial Engineering, May 1985, p. 84.
Greene, James H. Production and Inventory Control Handbook. New York: McGraw-Hill,
1987.
Groover, Mikell P. Automation, Production Systems, and Computer-Integrated Manufac-
turing. 3rd ed. Englewood Cliffs, NJ: Prentice Hall, 2007.
Hollier, R. H. “The Grouping Concept in Manufacture.” International Journal of Opera-
tions and Production Management, 1980, pp. 1–71.
Hopp, Wallace J., and M. Spearman, Factory Physics. New York: Irwin/McGraw-Hill,
2001, p. 180.
Kiran, A. S., and M. L. Smith. “Simulation Studies in Job Shop Scheduling: A Survey.”
Computers & Industrial Engineering, vol 8(2). 1984, pp. 87–93.
Kochan, D. CAM Developments in Computer Integrated Manufacturing. Berlin: Springer-
Verlag, 1986.
Langua, Leslie. “Simulation and Emulation.” Material Handling Engineering, May 1997,
pp. 41–43.
Law, A. M. “Introduction to Simulation: A Powerful Tool for Analyzing Complex Manu-
facturing Systems.” Industrial Engineering, May 1986, pp. 57–58.
Material Handling Institute, The. Consideration for Planning and Installing an Automated
Storage/Retrieval System. Pittsburgh, PA: Material Handling Institute, 1977.
Miller, Richard Kendall. Manufacturing Simulation. Lilburn, GA: The Fairmont Press,
1990.
Miner, R. J. “Simulation as a CIM Planning Tool.” In Simulation, ed. R. D. Hurrion, p. 73.
Kempston, Bedford, UK: IFS Publications, 1987.
Mitra, Debasis, and Isi Mitrani. “Analysis of a Kanban Discipline for Cell Coordination in
Production Lines.” Management Science 36, no. 12 (1990), pp. 1548–66.
Napolitano, Maida. “Distribution Network Modeling” IIE Solutions, June 1997, pp. 20–25.
Noaker, Paula M. “Strategic Cell Design.” Manufacturing Engineering, March 1993,
pp. 81–84.
Pritsker, A. A. B. Introduction to Simulation and Slam II. West Lafayette, IN: Systems
Publishing Corporation, 1986, p. 600.
Renbold, E. V., and R. Dillmann. Computer-Aided Design and Manufacturing Methods
and Tools. Berlin: Springer-Verlag, 1986.
Schonberger, Richard J., and Edward M. Knod Jr. Operations Management:Continuous
Improvement. 5th ed. Burr Ridge, IL: Richard D. Irwin, 1994.
Suri, R., and C. K. Whitney. “Decision Support Requirements in Flexible Manufacturing.”
Journal of Manufacturing Systems 3, no. 1 (1984).
har01307_ch11_299-322.indd 321 20/12/10 1:51 PM

Suzaki, Kiyoshi. The New Manufacturing Challenge: Techniques for Continuous Improve-
ment. New York: Free Press, 1987.
Wang, Hunglin, and Hsu Pin Wang. “Determining the Number of Kanbans: A Step Toward
Non-Stock Production.” International Journal of Production Research 28, no. 11 (1990),
pp. 2101–15.
Wick, C. “Advances in Machining Centers.” Manufacturing Engineering,October 1987,
p. 24.
Zisk, B. “Flexibility Is Key to Automated Material Transport System for Manufacturing
Cells.” Industrial Engineering, November 1983, p. 60.
har01307_ch11_299-322.indd 322 20/12/10 1:51 PM

C H A P T E R
12 MODELING MATERIAL
HANDLING SYSTEMS
“Small changes can produce big results, but the areas of highest leverage are
often the least obvious.”
—Peter Senge
12.1 Introduction
Material handling systems utilize resources to move entities from one location to
another. While material handling systems are not uncommon in service systems,
they are found mainly in manufacturing systems. Apple (1977) notes that mate-
rial handling can account for up to 80 percent of production activity. On average,
50 percent of a company’s operation costs are material handling costs (Meyers
1993). Given the impact of material handling on productivity and operation costs,
the importance of making the right material handling decisions should be clear.
This chapter examines simulation techniques for modeling material handling
systems. Material handling systems represent one of the most complicated fac-
tors, yet, in many instances, the most important element, in a simulation model.
Conveyor systems and automatic guided vehicle systems often constitute the
backbone of the material flow process. Because a basic knowledge of material
handling technologies and decision variables is essential to modeling material
handling systems, we will briefly describe the operating characteristics of each
type of material handling system.
12.2 Material Handling Principles

It is often felt that material handling is a necessary evil and that the ideal system
is one in which no handling is needed. In reality, material handling can be a tre-
mendous support to the production process if designed properly. Following is a
323
har01307_ch12_323-344.indd 323 12/17/10 9:48 AM

list of 10 principles published by the Material Handling Institute as a guide to

designing or modifying material handling systems:
1. Planning principle: The plan is the prescribed course of action and how
to get there. At a minimum it should define what, where, and when so
that the how and who can be determined.
2. Standardization principle: Material handling methods, equipment,
controls, and software should be standardized to minimize variety and
customization.
3. Work principle: Material handling work (volume weight or count
per unit of time distance) should be minimized. Shorten distances
and use gravity where possible.
4. Ergonomic principle: Human factors (physical and mental) and safety
must be considered in the design of material handling tasks and
equipment.
5. Unit load principle: Unit loads should be appropriately sized to achieve
the material flow and inventory objectives.
6. Space utilization principle: Effective and efficient use should be made
of all available (cubic) space. Look at overhead handling systems.
7. System principle: The movement and storage system should be
fully integrated to form a coordinated, operational system that spans
receiving, inspection, storage, production, assembly, packaging,
unitizing, order selection, shipping, transportation, and the handling of
returns.
8. Automation principle: Material handling operations should be
mechanized and/or automated where feasible to improve operational
efficiency, increase responsiveness, improve consistency and
predictability, decrease operating costs, and eliminate repetitive or
potentially unsafe manual labor.
9. Environmental principle: Environmental impact and energy
consumption should be considered as criteria when designing or
selecting alternative equipment and material handling systems.
10. Life cycle cost principle: A thorough economic analysis should account
for the entire life cycle of all material handling equipment and resulting
systems.
12.3 Material Handling Classification

With over 500 different types of material handling equipment to choose from, the
selection process can be quite staggering. We shall not attempt here to describe or
even list all of the possible types of equipment. Our primary interest is in model-
ing material handling systems, so we are interested only in the general catego-
ries of equipment having common operating characteristics. Material handling
har01307_ch12_323-344.indd 324 12/17/10 9:48 AM

Chapter 12 Modeling Material Handling Systems 325
systems are traditionally classified into one of the following categories (Tompkins
et al. 1996):
1. Conveyors.
2. Industrial vehicles.
3. Automated storage/retrieval systems.
4. Carousels.
5. Automatic guided vehicle systems.
6. Cranes and hoists.
7. Robots.
Missing from this list is “hand carrying,” which still is practiced widely if for no
other purpose than to load and unload machines.
12.4 Conveyors
A conveyor is a track, rail, chain, belt, or some other device that provides continu-
ous movement of loads over a fixed path. Conveyors are generally used for high-
volume movement over short to medium distances. Some overhead or towline
systems move material over longer distances. Overhead systems most often move
parts individually on carriers, especially if the parts are large. Floor-mounted con-
veyors usually move multiple items at a time in a box or container. Conveyor speeds
generally range from 20 to 80 feet per minute, with high-speed sortation conveyors
reaching speeds of up to 500 fpm in general merchandising operations.
Conveyors may be either gravity or powered conveyors. Gravity conveyors
are easily modeled as queues because that is their principal function. The more
challenging conveyors to model are powered conveyors, which come in a variety
of types.
12.4.1 Conveyor Types

Conveyors are sometimes classified according to function or configuration:
• Accumulation conveyor—a conveyor that permits loads to queue up
behind each other.
• Transport conveyor—a conveyor that transports loads from one point to
another; usually nonaccumulating.
• Sortation conveyor—a conveyor that provides diverting capability to
divert loads off the conveyor.
• Recirculating conveyor—a conveyor loop that causes loads to recirculate
until they are ready for diverting.
• Branching conveyor—a conveyor that is characterized by a main line with
either merging or diverging branches or spurs. Branches may, in turn,
have connected spurs or branches.
har01307_ch12_323-344.indd 325 12/17/10 9:48 AM

There are many different types of conveyors based on physical characteristics.

Among the most common types are the belt, roller, chain, trolley, power-and-free,
tow, and monorail conveyors. A brief description will be given of each of these
types.
Belt Conveyors. Belt conveyors have a circulating belt on which loads are car-
ried. Because loads move on a common span of fabric, if the conveyor stops, all
loads must stop. Belt conveyors are often used for simple load transport.
Roller Conveyors. Some roller conveyors utilize rollers that may be “dead”
if the bed is placed on a grade to enable the effect of gravity on the load to actuate
the rollers. However, many roller conveyors utilize “live” or powered rollers to
convey loads. Roller conveyors are often used for simple load accumulation.
Chain Conveyors. Chain conveyors are much like a belt conveyor in that one
or more circulating chains convey loads. Chain conveyors are sometimes used
in combination with roller conveyors to provide right-angle transfers for heavy
loads. A pop-up section allows the chains running in one direction to “comb”
through the rollers running in a perpendicular direction.
Trolley Conveyors. A trolley conveyor consists of a series of trolleys (small

wheels) that ride along a track and are connected by a chain, cable, or rod. Load
carriers are typically suspended from the trolleys. The trolleys, and hence the load
carriers, continually move so that the entire system must be stopped to bring a
single load to a stop. No accumulation can take place on the conveyor itself. Parts
are frequently loaded and unloaded onto trolley conveyors while the carriers are
in motion (usually not faster than 30 fpm). However, because a trolley conveyor
can be designed with a vertical dip, it can descend to automatically load a part.
Automatic unloading is simpler. A carrier can be tipped to discharge a load or a
cylinder can be actuated to push a load off a carrier onto an accumulation device.
Discharge points can be programmed.
Power-and-Free Conveyors. Power-and-free conveyors consist of two parallel

trolley systems, one of which is powered while trolleys in the other system are
nonpowered and free-moving. Loads are attached to the free trolleys and a “dog”
or engaging pin on the powered trolleys engages and disengages the passive free
trolleys to cause movement and stoppage. In some power-and-free systems, en-
gaged load trolleys will automatically disengage themselves from the power line
when they contact carriers ahead of them. This allows loads to be accumulated.
In the free state, loads can be moved by hand or gravity; they can be switched,
banked, or merely delayed for processing.
Tow Conveyors. A tow conveyor or towline system is an endless chain running

overhead in a track or on or beneath the ground in a track. The chain has a dog
or other device for pulling (towing) wheeled carts on the floor. Towlines are best
har01307_ch12_323-344.indd 326 12/17/10 9:48 AM

suited for applications where precise handling is unimportant. Tow carts are rela-
tively inexpensive compared to powered vehicles; consequently, many of them
can be added to a system to increase throughput and be used for accumulation.
An underfloor towline uses a tow chain in a trough under the floor. The
chain moves continuously, and cart movement is controlled by extending a drive
pin from the cart down into the chain. At specific points along the guideway,
computer-operated stopping mechanisms raise the drive pins to halt cart move-
ment. One advantage of this system is that it can provide some automatic buffering
with stationary carts along the track.
Towline systems operate much like power-and-free systems, and, in fact,
some towline systems are simply power-and-free systems that have been inverted.
By using the floor to support the weight, heavier loads can be transported.
Monorail Conveyors. Automatic monorail systems have self-powered carriers

that can move at speeds of up to 300 fpm. In this respect, a monorail system is
more of a discrete unit movement system than a conveyor system. Travel can be
disengaged automatically in accumulation as the leading “beaver tail” contacts
the limit switch on the carrier in front of it.
12.4.2 Operational Characteristics

The operational characteristics of conveyor systems are distinctly different from
those of other material handling systems. Some of the main distinguishing fea-
tures pertain to the nature of load transport, how capacity is determined, and how
entity pickup and delivery occur.
Load Transport. A conveyor spans the distance of travel along which loads
move continuously, while other systems consist of mobile devices that move one
or more loads as a batch move.
Capacity. The capacity of a conveyor is determined by the speed and load spac-
ing rather than by having a stated capacity. Specifically, it is a function of the
minimum allowable interload spacing (which is the length of a queue position in
the case of accumulation conveyors) and the length of the conveyor. In practice,
however, this may never be reached because the conveyor speed may be so fast
that loads are unable to transfer onto the conveyor at the minimum allowable
spacing. Furthermore, intentional control may be imposed on the number of loads
that are permitted on a particular conveyor section.
Entity Pickup and Delivery. Conveyors usually don’t pick up and drop off
loads, as in the case of lift trucks, pallet jacks, or hand carrying, but rather loads
are typically placed onto and removed from conveyors. Putting a load onto an
input spur of a branching conveyor and identifying the load to the system so it can
enter the conveyor system is commonly referred to as load induction (the spur is
called an induction conveyor).
har01307_ch12_323-344.indd 327 12/17/10 9:48 AM

12.4.3 Modeling Conveyor Systems

Depending on the nature of the conveyor and its operation, modeling a conveyor
can range from fairly straightforward to extremely complex. For single conveyor
sections, modeling is very simple. Branching and recirculating conveyors, on the
other hand, give rise to several complexities (recirculation, merging, and so on)
that make it nearly impossible to predict how the system will perform. These
types of conveyors make especially good problems for simulation.
From a modeling standpoint, a conveyor may be classified as being either
accumulation or nonaccumulation (also referred to as accumulating or nonaccu-
mulating), with either fixed or random load spacing (see Table 12.1).
Accumulation Conveyors. Accumulation conveyors are a class of conveyors

that provide queuing capability so that if the lead load stops, succeeding loads
still advance along the conveyor and form a queue behind the stopped part. Ac-
cumulation conveyors permit independent or nonsynchronous part movement and
therefore maximize part carrying capacity.
Nonaccumulation Conveyors. Nonaccumulation or simple transport conveyors

are usually powered conveyors having a single drive such that all movement on
the conveyor occurs in unison, or synchronously. If a load must be brought to a
stop, the entire conveyor and every load on it is also brought to a stop. When the
conveyor starts again, all load movement resumes.
Fixed Spacing. Conveyors with fixed load spacing, sometimes referred to as

segmented conveyors, require that loads be introduced to the conveyor at fixed
intervals. Tow conveyors and power-and-free conveyors, which have fixed spaced
dogs that engage and disengage a part carrying cart or trolley, are examples of
accumulation conveyors with fixed load spacing. Trolley conveyors with hanging
carriers at fixed intervals or tray conveyors with permanently attached trays at
fixed intervals are examples of nonaccumulation conveyors with fixed spacing.
Random Load Spacing. Random load spacing permits parts to be placed at

any distance from another load on the same conveyor. A powered roller conveyor
is a common example of an accumulation conveyor with random load spacing.
Powered conveyors typically provide noncontact accumulation, while gravity
conveyors provide contact accumulation (that is, the loads accumulate against
TABLE 12.1 Types of Conveyors for Modeling Purposes

Accumulation Nonaccumulation
Fixed Spacing Power-and-free, towline Trolley, sortation

Random Spacing Roller, monorail Belt, chain
har01307_ch12_323-344.indd 328 12/17/10 9:48 AM

each other). A powered belt conveyor is a typical example of a nonaccumulation

conveyor with random spacing. Parts may be interjected anywhere on the belt,
and when the belt stops, all parts stop.
Performance Measures and Decision Variables. Throughput capacity, delivery

time, and queue lengths (for accumulation conveyors) are performance measures
in conveyor system simulation. Several issues that are addressed in conveyor anal-
ysis include conveyor speed, accumulation capacity, and the number of carriers.
Questions to Be Answered. Common questions that simulation can help answer

in designing and operating a conveyor system are
• What is the minimum conveyor speed that still meets throughput
requirements?
• What is the throughput capacity of the conveyor?
• What is the load delivery time for different activity levels?
• How much queuing is needed on accumulation conveyors?
• How many carriers are needed on a trolley or power-and-free conveyor?
• What is the optimum number of pallets that maximizes productivity in a
recirculating conveyor?
12.4.4 Modeling Single-Section Conveyors

Conveyors used for simple queuing or buffering often consist of a single stretch
or section of conveyor. Loads enter at one end and are removed at the other end.
Load spacing is generally random. These types of conveyors are quite easy to
model, especially using simulation products that provide conveyor constructs. In
ProModel, a conveyor is simply a type of location. The modeler merely defines
the length and speed and states whether the conveyor is an accumulation conveyor
or nonaccumulation conveyor. Since a conveyor is just a location, routings can be
defined to and from a conveyor. Operations can be defined at either the beginning
or end of a conveyor.
If the simulation product being used does not have a conveyor construct, a
simple accumulation conveyor can be modeled as a queue with a delay time equal
to the time it takes to move from the beginning to the end. This can also be easily
done in ProModel. For nonaccumulation conveyors the approach is more difficult.
Henriksen and Schriber (1986) present a simple yet accurate approach to model-
ing single sections of accumulation and nonaccumulation conveyors with random
load spacing and no branching or merging.
While these simplified approaches may be adequate for rough modeling of
single conveyor sections in which parts enter at one end and exit at the other, they
are inadequate for a conveyor system in which parts are continuously branching
and merging onto conveyor sections. This type of system requires continuous
tracking, which, in many discrete-oriented simulation languages, gets very com-
plex or tends to be impractical.
har01307_ch12_323-344.indd 329 12/17/10 9:48 AM

12.4.5 Modeling Conveyor Networks

A conveyor network consists of two or more sections that are connected together
to enable part merging, diverting, and recirculating. In such instances, a conveyor
may have one or more entry points and one or more exit points. Furthermore,
part entry and exit may not always be located at extreme opposite ends of the
conveyor.
Conveyor networks are modeled in ProModel by placing conveyors next to
each other and defining routings from one conveyor to the next. Branching is
limited in ProModel in that entities can enter a conveyor only at the beginning
and exit at the end. This may require breaking a conveyor into multiple sections
to accommodate branching in the middle of a conveyor.
Conveyor networks may provide alternative routes to various destinations.
Part movement time on a conveyor often is determinable only by modeling
the combined effect of the conveying speed, the length of the conveyor, the
load spacing requirements, the connecting conveyor logic, and any conveyor
transfer times. The traffic and conveyor dynamics ultimately determine the
move time.
12.5 Industrial Vehicles

Industrial vehicles include all push or powered carts and vehicles that generally
have free movement. Powered vehicles such as lift trucks are usually utilized
for medium-distance movement of batched parts in a container or pallet. For
short moves, manual or semipowered carts are useful. Single-load transporters
are capable of moving only one load at a time from one location to another. Such
devices are fairly straightforward to model because they involve only a single
source and a single destination for each move. Multiple-load transporters, on
the other hand, can move more than one load at a time from one or more sources
to one or more destinations. Defining the capacity and operation for multiple-
load transporters can be extremely difficult because there are special rules
defining when to retrieve additional loads and when to deliver the loads already
on board. One example of a multiple-load transporter is a picking transporter,
which simply may be a person pushing a cart through a part storage area load-
ing several different part types onto the cart. These parts may then be delivered
to a single destination, such as when filling a customer order, or to several
destinations, such as the replenishment of stock along a production line. Towing
vehicles are another example of this type of transporter. A towing vehicle pulls
a train of carts behind it that may haul several loads at once. Towing vehicles
may have a single pickup and drop point for each move similar to a unit-load
transporter; however, they are also capable of making several stops, picking
up multiple loads, and then delivering them to several remote locations. Such
an operation can become extremely complex to incorporate into a generalized
simulation language.
har01307_ch12_323-344.indd 330 12/17/10 9:48 AM

12.5.1 Modeling Industrial Vehicles

Modeling an industrial vehicle involves modeling a resource that moves along a
path network. Paths are typically open aisles in which bidirectional movement is
possible and passing is permitted. Some powered vehicles operate on a unidirec-
tional track. Deployment strategies (work searches, idle vehicle parking, and so
on) must be capable of being incorporated into the model, similar to that of an
automated guided vehicle system (AGVS), described later.
Because industrial vehicles are often human-operated, they may take breaks
and be available only during certain shifts.
Common performance measures that are of interest when modeling industrial
vehicles include
• Vehicle utilization.
• Response time.
• Move rate capability.
Decision variables that are tested for industrial vehicles include
• Number of vehicles.
• Empty vehicle positioning.
Typical questions to be answered regarding industrial vehicles are
• What is the required number of vehicles to handle the required activity?
• What is the best deployment of vehicles to maximize utilization?
• What is the best deployment of empty vehicles to minimize response time?
12.6 Automated Storage/Retrieval Systems

An automated storage/retrieval system (AS/RS) consists of one or more automated
storage/retrieval machines that store and retrieve material to and from a storage
facility such as a storage rack. Because it is automated, the system operates under
computer control. The goal of an AS/RS is to provide random, high-density stor-
age with quick load access, all under computer control.
An AS/RS is characterized by one or more aisles of rack storage having one
or more storage/retrieval (S/R) machines (sometimes called vertical or stacker
cranes) that store and retrieve material into and out of a rack storage system.
Material is picked up for storage by the S/R machine at an input (pickup) station.
Retrieved material is delivered by the S/R machine to an output (deposit) station
to be taken away. Usually there is one S/R machine per aisle; however, there may
also be one S/R machine assigned to two or more aisles or perhaps even two S/R
machines assigned to a single aisle. For higher storage volume, some AS/RSs
utilize double-deep storage, which may require load shuffling to access loads.
Where even higher storage density is required and longer-term storage is needed,
har01307_ch12_323-344.indd 331 12/17/10 9:48 AM

multidepth or deep lane storage systems are used. AS/RSs that use pallet storage
racks are usually referred to as unit-load systems, while bin storage rack systems
are known as miniload systems.
The throughput capacity of an AS/RS is a function of the rack configuration
and speed of the S/R machine. Throughput is measured in terms of how many
single or dual cycles can be performed per hour. A single cycle is measured as the
average time required to pick up a load at the pickup station, store the load in a
rack location, and return to the pickup station. A dual cycle is the time required
to pick up a load at the input station, store the load in a rack location, retrieve a
load from another rack location, and deliver the load to the output station. Obvi-
ously, there is considerable variation in cycle times due to the number of different
possible rack locations that can be accessed. If the AS/RS is a stand-alone system
with no critical interface with another system, average times are adequate for
designing the system. If, however, the system interfaces with other systems such
as front-end or remote-picking stations, it is important to take into account the
variability in cycle times.
12.6.1 Configuring an AS/RS

Configuring an AS/RS depends primarily on the storage/retrieval or throughput
activity that is required and the amount of storage space needed. Other consider-
ations are building constraints in length, width, and height (especially if it is an
existing facility); crane height limitations; and cost factors (a square building is
less costly than a rectangular one).
Configuring an AS/RS is essentially an iterative process that converges on
a solution that meets both total storage requirements and total throughput re-
quirements. Once the number of rack locations is calculated based on storage
requirements, an estimate of the number of aisles is made based on throughput re-
quirements. This usually assumes a typical throughput rate of, say, 20 dual cycles
per hour and assumes that one S/R machine is assigned to an aisle. If, for example,
the required throughput is 100 loads in and 100 loads out per hour, an initial con-
figuration would probably be made showing five aisles. If, however, it was shown
using an AS/RS simulator or other analytical technique that the throughput capac-
ity was only 15 cycles per hour due to the large number of rack openings required,
the number of aisles may be increased to seven or eight and analyzed again. This
process is repeated until a solution is found.
Occasionally it is discovered that the throughput requirement is extremely
low relative to the number of storage locations in the system. In such situations,
rather than have enormously long aisles, a person-aboard system will be used in
which operator-driven vehicles will be shared among several aisles, or perhaps
double-deep storage racks will be used.
Computerized Cycle Time Calculations

It is often thought necessary to simulate an AS/RS in order to determine through-
put capacity. This is because the S/R machine accesses random bins or rack
har01307_ch12_323-344.indd 332 12/17/10 9:48 AM

locations, making the cycle time extremely difficult to calculate. The problem is
further compounded by mixed single (pickup and store or retrieve and deposit)
and dual (pickup, storage, retrieval, and deposit) cycles. Activity zoning, in which
items are stored in assigned zones based on frequency of use, also complicates
cycle time calculations. The easiest way to accurately determine the cycle time
for an AS/RS is by using a computer to enumerate the possible movements of the
S/R machine from the pickup and delivery (P&D) stand to every rack or bin loca-
tion. This produces an empirical distribution for the single cycle time. For dual
cycle times, an intermediate cycle time—the time to go from any location to any
other location—must be determined. For a rack 10 tiers high and 40 bays long,
this can be 400 400, or 160,000 calculations! Because of the many calculations,
sometimes a large sample size is used to develop the distribution. Most suppliers
of automated storage/retrieval systems have computer programs for calculating
cycle times that can be generated based on a defined configuration in a matter
of minutes.
Analytical Cycle Time Calculations

Analytical solutions have been derived for calculating system throughput based
on a given aisle configuration. Such solutions often rely on simplified assumptions
about the operation of the system. Bozer and White (1984), for example, derive
an equation for estimating single and dual cycle times assuming (1) randomized
storage, (2) equal rack opening sizes, (3) P&D location at the base level on one
end, (4) constant horizontal and vertical speeds, and (5) simultaneous horizontal
and vertical rack movement. In actual practice, rack openings are seldom of equal
size, and horizontal and vertical accelerations can have a significant influence on
throughput capacity.
While analytical solutions to throughput estimation may provide a rough
approximation for simple configurations, other configurations become extremely
difficult to estimate. In addition, there are control strategies that may improve
throughput rate, such as retrieving the load on the order list that is closest to the
load just stored or storing a load in an opening that is closest to the next load to
be retrieved. Finally, analytical solutions provide not a distribution of cycle times
but merely a single expected time, which is inadequate for analyzing AS/RSs that
interface with other systems.
AS/RS with Picking Operations

Whereas some AS/RSs (especially unit-load systems) have loads that are not
captive to the system, many systems (particularly miniload systems) deliver
bins or pallets either to the end of the aisle or to a remote area where material is
picked from the bin or pallet, which is then returned for storage. Remote picking
is usually achieved by linking a conveyor system to the AS/RS where loads are
delivered to remote picking stations. In this way, containers stored in any aisle
can be delivered to any workstation. This permits entire orders to be picked at a
single station and eliminates the two-step process of picking followed by order
consolidation.
har01307_ch12_323-344.indd 333 12/17/10 9:48 AM

Where picking takes place is an important issue achieving the highest pro-
ductivity from both AS/RS and picker. Both are expensive resources, and it is
undesirable to have either one waiting on the other.
12.6.2 Modeling AS/RSs

Simulation of AS/RSs has been a popular area of application for simulation. Ford
developed a simulator called GENAWS as early as 1972. IBM developed one in
1973 called the IBM WAREHOUSE SIMULATOR. Currently, most simulation of
AS/RSs is performed by vendors of the systems. Typical inputs for precision model-
ing of single-deep AS/RSs with each aisle having a captive S/R machine include
• Number of aisles.
• Number of S/R machines.
• Rack configuration (bays and tiers).
• Bay or column width.
• Tier or row height.
• Input point(s).
• Output point(s).
• Zone boundaries and activity profile if activity zoning is utilized.
• S/R machine speed and acceleration/deceleration.
• Pickup and deposit times.
• Downtime and repair time characteristics.
At a simple level, an AS/RS move time may be modeled by taking a time from
a probability distribution that approximates the time to store or retrieve a load.
More precise modeling incorporates the actual crane (horizontal) and lift (verti-
cal) speeds. Each movement usually has a different speed and distance to travel,
which means that movement along one axis is complete before movement along
the other axis. From a modeling standpoint, it is usually necessary to calculate and
model only the longest move time.
In modeling an AS/RS, the storage capacity is usually not a consideration
and the actual inventory of the system is not modeled. It would require lots of
overhead to model the complete inventory in a rack with 60,000 pallet locations.
Because only the activity is of interest in the simulation, actual inventory is ig-
nored. In fact, it is usually not even necessary to model specific stock keeping
units (SKUs) being stored or retrieved, but only to distinguish between load types
insofar as they affect routing and subsequent operations.
Common performance measures that are of interest in simulating an AS/RS
include
• S/R machine utilization.
• Response time.
• Throughput capability.
har01307_ch12_323-344.indd 334 12/17/10 9:48 AM

Decision variables on which the modeler conducts experiments include

• Rack configuration and number of aisles.
• Storage and retrieval sequence and priorities.
• First-in, first-out or closest item retrieval.
• Empty S/R machine positioning.
• Aisle selection (random, round robin, or another method).
Typical questions addressed in simulation of an AS/RS include
• What is the required number of aisles to handle the required activity?
• Should storage activity be performed at separate times from retrieval
activity, or should they be intermixed?
• How can dual cycling (combining a store with a retrieve) be maximized?
• What is the best stationing of empty S/R machines to minimize response
time?
• How can activity zoning (that is, storing fast movers closer than slow
movers) improve throughput?
• How is response time affected by peak periods?
12.7 Carousels
One class of storage and retrieval systems is a carousel. Carousels are essentially
moving racks that bring the material to the retriever (operator or robot) rather than
sending a retriever to the rack location.
12.7.1 Carousel Configurations

A carousel storage system consists of a collection of bins that revolve in either
a horizontal or vertical direction. The typical front-end activity on a carousel is
order picking. If carousels are used for WIP storage, bins might enter and leave
the carousel. The average time to access a bin in a carousel is usually equal to
one-half of the revolution time. If bidirectional movement is implemented, aver-
age access time is reduced to one-fourth of the revolution time. These times can
be further reduced if special storing or ordering schemes are used to minimize
access time.
A variation of the carousel is a rotary rack consisting of independently operat-
ing tiers. Rotary racks provide the added advantage of being able to position another
bin into place while the operator is picking out of a bin from a different tier.
12.7.2 Modeling Carousels

Carousels are easily modeled by defining an appropriate response time represent-
ing the time for the carousel to bring a bin into position for picking. In addition to
har01307_ch12_323-344.indd 335 12/17/10 9:48 AM

response times, carousels may have capacity considerations. The current contents
may even affect response time, especially if the carousel is used to store multiple
bins of the same item such as WIP storage. Unlike large AS/RSs, storage capacity
may be an important issue in modeling carousel systems.
12.8 Automatic Guided Vehicle Systems

An automatic guided vehicle system (AGVS) is a path network along which
computer-controlled, driverless vehicles transport loads. AGV systems are usu-
ally used for medium activity over medium distances. If parts are large, they are
moved individually; otherwise parts are typically consolidated into a container
or onto a pallet. Operationally, an AGVS is more flexible than conveyors, which
provide fixed-path, fixed-point pickup and delivery. However, they cannot handle
the high activity rates of a conveyor. On the other extreme, an AGVS is not as
flexible as industrial vehicles, which provide open-path, any-point pickup and de-
livery. However, an AGVS can handle higher activity rates and eliminate the need
for operators. AGVSs provide a flexible yet defined path of movement and allow
pickup and delivery between any one of many fixed points, except for where cer-
tain points have been specifically designated as pickup or drop-off points.
AGVs are controlled by either a central computer, onboard computers, or
a combination of the two. AGVs with onboard microprocessors are capable of
having maps of the system in their memories. They can direct their own routing
and blocking and can communicate with other vehicles if necessary. Some AGVs
travel along an inductive guide path, which is a wire buried a few inches beneath
the floor. Newer systems are non-wire guided. Travel is generally unidirectional
to avoid traffic jams, although occasionally a path might be bidirectional.
The AGVS control system must make such decisions as what vehicle to dis-
patch to pick up a load, what route to take to get to the pickup point, what to do
if the vehicle gets blocked by another vehicle, what to do with idle vehicles, and
how to avoid collisions at intersections.
AGVs have acceleration and deceleration, and speeds often vary depending
on whether vehicles are full or empty and whether they are on a turn or a stretch.
There is also a time associated with loading and unloading parts. Furthermore,
loaded vehicles may stop at points along the path for operations to be performed
while the load is on the vehicle. One or more vehicles may be in a system, and
more than one system may share common path segments. AGVs stop periodically
for battery recharging and may experience unscheduled failures.
One of the modeling requirements of an AGVS is to accurately describe the
method for controlling traffic and providing anticollision control. This is usually ac-
complished in one of two ways: (1) onboard vehicle sensing or (2) zone blocking.
Onboard vehicle sensing works by having a sensor detect the presence of a
vehicle ahead and stop until it detects that the vehicle ahead has moved. This is
usually accomplished using a light beam and receiver on the front of each vehicle
with a reflective target on the rear of each.
har01307_ch12_323-344.indd 336 12/17/10 9:48 AM

In zone blocking, the guide path is divided into various zones (segments) that
allow no more than one vehicle at a time. Zones can be set up using a variety of
different sensing and communication techniques. When one vehicle enters a zone,
other vehicles are prevented from entering until the current vehicle occupying the
zone leaves. Once the vehicle leaves, any vehicle waiting for access to the freed
zone can resume travel.
12.8.1 Designing an AGVS

When designing an AGVS, the first step is to identify all of the pickup and drop-
off points. Next the path should be laid out. In laying out the guide path, several
principles of good design should be followed (Askin and Standridge 1993):
• Travel should be unidirectional in the form of one or more recirculating
loops unless traffic is very light. This avoids deadlock in traffic.
• Pickup stations should be downstream of drop-off stations. This enables
an AGV to pick up a load, if available, whenever a load is dropped off.
• Pickup and deposit stations should generally be placed on spurs unless
traffic is extremely light. This prevents AGVs from being blocked.
• Crossover paths should be used occasionally if measurable reductions in
travel times can be achieved.
Once the path has been configured, a rough estimate is made of the number of
AGVs required. This is easily done by looking at the total travel time required per
hour, dividing it by the number of minutes in an hour (60), and multiplying it by
1 plus a traffic allowance factor:
Total AGV minutes (1 Traffic allowance)
AGV ________________
60 minutes/hour
The traffic allowance is a percentage factored into the vehicle requirements
to reflect delays caused by waiting at intersections, rerouting due to congestion,
or some other delay. This factor is a function of both poor path layout and quantity
of vehicles and may vary anywhere from 0 in the case of a single-vehicle system
to around 15 percent in congested layouts with numerous vehicles (Fitzgerald
1985).
The total vehicle minutes are determined by looking at the average required
number of pickups and deposits per hour and converting this into the number of
minutes required for one vehicle to perform these pickups and drop-offs. It is
equal to the sum of the total empty move time, the total full move time, and the
total pickup and deposit time for every delivery. Assuming I pickup points and
J drop-off points, the total vehicle time can be estimated by

I J

Distanceij
Total vehicle minutes 2 __________ Pickupi Depositj
i1 j1 Avg. speedij
Deliveriesij
har01307_ch12_323-344.indd 337 12/17/10 9:48 AM

The expression
Distanceij
2 __________
Avg. speedij
implies that the time required to travel to the point of pickup (empty travel time)
is the same as the time required to travel to the point of deposit (full travel time).
This assumption provides only an estimate of the time required to travel to a point
of pickup because it is uncertain where a vehicle will be coming from. In most
cases, this should be a conservative estimate because (1) vehicles usually follow a
work search routine in which the closest loads are picked up first and (2) vehicles
frequently travel faster when empty than when full. A more accurate way of cal-
culating empty load travel for complex systems is to use a compound weighted-
averaging technique that considers all possible empty moves together with their
probabilities (Fitzgerald 1985).
12.8.2 Controlling an AGVS

A multivehicle AGVS can be quite complex and expensive, so it is imperative
that the system be well managed to achieve the utilization and level of produc-
tivity desired. Part of the management problem is determining the best deploy-
ment strategy or dispatching rules used. Some of these operating strategies are
discussed next.
AGV Selection Rules. When a load is ready to be moved by an AGV and more
than one AGV is available for executing the move, a decision must be made as
to which vehicle to use. The most common selection rule is a “closest rule” in
which the nearest available vehicle is selected. This rule is intended to minimize
empty vehicle travel time as well as to minimize the part waiting time. Other
less frequently used vehicle selection rules include longest idle vehicle and least
utilized vehicle.
Work Search Rules. If a vehicle becomes available and two or more parts are
waiting to be moved, a decision must be made as to which part should be moved
first. A number of rules are used to make this decision, each of which can be effec-
tive depending on the production objective. Common rules used for dispatching
an available vehicle include
• Longest-waiting load.
• Closest waiting load.
• Highest-priority load.
• Most loads waiting at a location.
Vehicle Parking Rules. If a transporter delivers a part and no other parts are
waiting for pickup, a decision must be made relative to the deployment of the
transporter. For example, the vehicle can remain where it is, or it can be sent to
a more strategic location where it is likely to be needed next. If several transport
har01307_ch12_323-344.indd 338 12/17/10 9:48 AM

vehicles are idle, it may be desirable to have a prioritized list for a vehicle to
follow for alternative parking preferences.
Work Zoning. In some cases it may be desirable to keep a vehicle captive to a

particular area of production and not allow it to leave this area unless it has work
to deliver elsewhere. In this case, the transporter must be given a zone boundary
within which it is capable of operating.
12.8.3 Modeling an AGVS

Modeling an AGVS is very similar to modeling an industrial vehicle (which it is
in a sense) except that the operation is more controlled with less freedom of move-
ment. Paths are generally unidirectional, and no vehicle passing is allowed.
One of the challenges in modeling an AGVS in the past has been finding
the shortest routes between any two stops in a system. Current state-of-the-art
simulation software, like ProModel, provides built-in capability to automatically
determine the shortest routes between points in a complex network. Alternatively,
the modeler may want to have the capability to explicitly define a set of paths for
getting from one location to another.
Common performance measures that are of interest when modeling an AGVS
include
• Load throughput rate.
• Response time.
• Vehicle congestion.
In addition to the obvious purpose of simulating an AGVS to find out if the num-
ber of vehicles is sufficient or excessive, simulation can also be used to determine
the following:
• Number of vehicles.
• Work search rules.
• Park search rules.
• Placement of crossover and bypass paths.
Typical questions that simulation can help answer include
• What is the best path layout to minimize travel time?
• Where are the potential bottleneck areas?
• How many vehicles are needed to meet activity requirements?
• What are the best scheduled maintenance/recharging strategies?
• Which task assignment rules maximize vehicle utilization?
• What is the best idle vehicle deployment that minimizes response time?
• Is there any possibility of deadlocks?
• Is there any possibility of collisions?
har01307_ch12_323-344.indd 339 12/17/10 9:48 AM

12.9 Cranes and Hoists

Cranes are floor-, ceiling-, or wall-mounted mechanical devices generally used
for short- to medium-distance, discrete movement of material. Cranes utilize a
hoist mechanism for attaching and positioning loads. Aside from bridge and gan-
try cranes, most other types of cranes and hoists can be modeled as an industrial
vehicle. Since a gantry crane is just a simple type of bridge crane that is floor-
mounted, our discussion here will be limited to bridge cranes. A bridge crane
consists of a beam that bridges a bay (wall to wall). This beam or bridge moves
on two tracks mounted on either wall. A retracting hoist unit travels back and forth
under the bridge, providing a total of three axes of motion (gantry cranes have
similar movement, only the runway is floor-mounted).
12.9.1 Crane Management

Managing crane movement requires an understanding of the priority of loads to be
moved as well as the move characteristics of the crane. One must find the optimum
balance of providing adequate response time to high-priority moves while maxi-
mizing the utilization of the crane. Crane utilization is maximized when drop-offs
are always combined with a nearby pickup. Crane management becomes more
complicated when (as is often the case) two or more cranes operate in the same
bay or envelope of space. Pritsker (1986) identifies typical rules derived from
practical experience for managing bridge cranes sharing the same runway:
1. Cranes moving to drop points have priority over cranes moving to
pickup points.
2. If two cranes are performing the same type of function, the crane that
was assigned its task earliest is given priority.
3. To break ties on rules 1 and 2, the crane closest to its drop-off point is
given priority.
4. Idle cranes are moved out of the way of cranes that have been given
assignments.
These rules tend to minimize the waiting time of loaded cranes and maximize
crane throughput.
12.9.2 Modeling Bridge Cranes

For single cranes a simple resource can be defined that moves along a path that
is the resultant of the x and y movement of the bridge and hoist. For more precise
modeling or for managing multiple cranes in a runway, it is preferable to utilize
built-in crane constructs. ProModel provides powerful bridge crane modeling ca-
pabilities that greatly simplify bridge crane modeling.
Common performance measures when simulating a crane include
• Crane utilization.
• Load movement rate.
har01307_ch12_323-344.indd 340 12/17/10 9:48 AM

• Response time.
• Percentage of time blocked by another crane.
Decision variables that become the basis for experimentation include
• Work search rules.
• Park search rules.
• Multiple-crane priority rules.
Typical questions addressed in a simulation model involving cranes include
• Which task assignment rules maximize crane utilization?
• What idle crane parking strategy minimizes response time?
• How much time are cranes blocked in a multicrane system?
12.10 Robots
Robots are programmable, multifunctional manipulators used for handling material
or manipulating a tool such as a welder to process material. Robots are often clas-
sified based on the type of coordinate system on which they are based: cylindrical,
cartesian, or revolute. The choice of robot coordinate system depends on the ap-
plication. Cylindrical or polar coordinate robots are generally more appropriate for
machine loading. Cartesian coordinate robots are easier to equip with tactile sen-
sors for assembly work. Revolute or anthropomorphic coordinate robots have the
most degrees of freedom (freedom of movement) and are especially suited for use
as a processing tool, such as for welding or painting. Because cartesian or gantry
robots can be modeled easily as cranes, our discussion will focus on cylindrical
and revolute robots. When used for handling, cylindrical or revolute robots are
generally used to handle a medium level of movement activity over very short dis-
tances, usually to perform pick-and-place or load/unload functions. Robots gener-
ally move parts individually rather than in a consolidated load.
One of the applications of simulation is in designing the cell control logic for
a robotic work cell. A robotic work cell may be a machining, assembly, inspec-
tion, or a combination cell. Robotic cells are characterized by a robot with 3 to
5 degrees of freedom surrounded by workstations. The workstation is fed parts
by an input conveyor or other accumulation device, and parts exit from the cell
on a similar device. Each workstation usually has one or more buffer positions to
which parts are brought if the workstation is busy. Like all cellular manufacturing,
a robotic cell usually handles more than one part type, and each part type may not
have to be routed through the same sequence of workstations. In addition to part
handling, the robot may be required to handle tooling or fixtures.
12.10.1 Robot Control

Robotic cells are controlled by algorithms that determine the sequence of activi-
ties for the robot to perform. This control is developed offline using algorithms
har01307_ch12_323-344.indd 341 12/17/10 9:48 AM

for selecting what activity to perform. The sequence of activities performed by

the robot can be either a fixed sequence of moves that is continually repeated
or a dynamically determined sequence of moves, depending on the status of the
cell. For fixed-sequence logic, the robot control software consists of a sequence
of commands that are executed sequentially. Dynamic sequence selection is pro-
grammed using if–then statements followed by branches dependent on the truth
value of the conditional statement. System variables must also be included in the
system to keep track of system status, or else the robot polls the sensors, and so
on, of the system each time a status check is needed.
In developing an algorithm for a cell controller to make these decisions, the
analyst must first decide if the control logic is going to control the robot, the part,
or the workstation. In a part-oriented system, the events associated with the part
determine when decisions are made. When a part finishes an operation, a routing
selection is made and then the robot is requested. In a robot-oriented cell, the
robot drives the decisions and actions so that when a part completes an operation,
it merely sends a signal or sets a status indicator showing that it is finished. Then
when the robot becomes available, it sees that the part is waiting and, at that mo-
ment, the routing decision for the part is made.
12.10.2 Modeling Robots

In modeling robots, it is sometimes presumed that the kinematics of the robot
need to be simulated. Kinematic simulation is a special type of robot simulation
used for cycle time analysis and offline programming. For discrete event simula-
tion, it is necessary to know only the move times from every pickup point to every
deposit point, which usually aren’t very many. This makes modeling a robot no
more difficult than modeling a simple resource. The advantage to having a spe-
cific robot construct in a simulation product is primarily for providing the graphic
animation of the robot. ProModel doesn’t have a built-in construct specifically for
robots, although submodels have been defined that can be used in other models.
Common performance measures when modeling robots include
• Robot utilization.
• Response time.
• Throughput.
Decision variables that come into play when simulating robots include
• Pickup sequence.
• Idle robot positioning.
Typical questions that simulation can help answer are
• What priority of tasks results in the highest productivity?
• Where is the best position for the robot after completing each particular
drop-off?
har01307_ch12_323-344.indd 342 12/17/10 9:48 AM

12.11 Summary
Material handling systems can be one of the most difficult elements to model
using simulation simply because of the sheer complexity. At the same time, the
material handling system is often the critical element in the smooth operation of
a manufacturing or warehousing system. The material handling system should be
designed first using estimates of resource requirements and operating parameters
(speed, move capacity, and so forth). Simulation can then help verify design deci-
sions and fine-tune the design.
In modeling material handling systems, it is advisable to simplify wherever
possible so models don’t become too complex. A major challenge in modeling
conveyor systems comes when multiple merging and branching occur. A chal-
lenging issue in modeling discrete part movement devices, such as AGVs, is how
to manage their deployment in order to get maximum utilization and meet produc-
tion goals.

1. What impact can the material handling system have on the performance
of a manufacturing system?
2. What is the essential difference between a conveyor and other transport
devices from a modeling standpoint?
3. In material handling simulation, should we be interested in modeling
individual parts or unit loads? Explain.
4. How is the capacity of a conveyor determined?
5. What is the difference between an accumulation and a
nonaccumulation conveyor?
6. What is the difference between fixed spacing and random spacing on a
conveyor?
7. Under what conditions could you model an accumulation conveyor as a
simple queue or buffer?
8. What are typical conveyor design decisions that simulation can help
make?
9. What are typical simulation objectives when modeling an AS/RS?
10. When modeling an AS/RS, is there a benefit in modeling the actual
storage positions? Explain.
11. Conceptually, how would you model the movement of a carousel?
12. Draw a logic diagram of what happens from the time an AGV deposits
a load to the time it retrieves another load.
13. When modeling a robot in system simulation, is it necessary to model
the actual kinematics of the robot? Explain.
14. What are the unique characteristics of a bridge crane from a modeling
perspective?
har01307_ch12_323-344.indd 343 12/17/10 9:48 AM

References
Apple, J. M. Plant Layout and Material Handling. 3rd ed. N.Y.: Ronald Press, 1977.
Askin, Ronald G., and C. R. Standridge. Modeling and Analysis of Manufacturing Systems.
New York: John Wiley & Sons, 1993.
“Automatic Monorail Systems.” Material Handling Engineering, May 1988, p. 95.
Bozer, Y. A., and J. A. White. “Travel-Time Models for Automated Storage/Retrieval
Systems.” IIE Transactions 16, no. 4 (1984), pp. 329–38.
Fitzgerald, K. R. “How to Estimate the Number of AGVs You Need.” Modern Materials
Handling, October 1985, p. 79.
Henriksen, J., and T. Schriber. “Simplified Approaches to Modeling Accumulation and
Non-Accumulating Conveyor Systems.” In Proceedings of the 1986 Winter Simula-
tion Conference, ed. J. Wilson, J. Henricksen, and S. Roberts. Piscataway, NJ: Insti-
tute of Electrical and Electronics Engineers, 1986.
Matthew, Stephens, and Fred E. Meyers. Manufacturing Facilities Design and Material
Handling. 4th ed. Englewood Cliffs, NJ: Prentice Hall, 2009.
Meyers, Fred E. Plant Layout and Material Handling. Englewood Cliffs, NJ: Regents/
Prentice Hall, 1993.
Pritsker, A. A. B. Introduction to Simulation and Slam II. West Lafayette, IN: Systems
Publishing Corporation, 1986, p. 600.
Tompkins, J. A.; J. A. White; Y. A. Bozer; E. H. Frazelle; J. M. A. Tanchoco; and J. Trevino.
Facilities Planning. 2nd ed. New York: John Wiley & Sons, 1996.
Zisk, B. “Flexibility Is Key to Automated Material Transport System for Manufacturing
Cells.” Industrial Engineering, November 1983, p. 60.
har01307_ch12_323-344.indd 344 12/17/10 9:48 AM

C H A P T E R
13 MODELING SERVICE
SYSTEMS
“No matter which line you move to, the other line always moves faster.”
—Unknown
13.1 Introduction
A service system is a processing system in which one or more services are pro-
vided to customers. Entities (customers, patients, paperwork) are routed through
a series of processing areas (check-in, order, service, payment) where resources
(service agents, doctors, cashiers) provide some service. Service systems exhibit
unique characteristics that are not found in manufacturing systems. Sasser, Olsen,
and Wyckoff (1978) identify four distinct characteristics of services that distin-
guish them from products that are manufactured:
1. Services are intangible; they are not things.
2. Services are perishable; they cannot be inventoried.
3. Services provide heterogeneous output; output is varied.
4. Services involve simultaneous production and consumption; the service
is produced and used at the same time.
These characteristics pose great challenges for service system design and man-
agement, particularly in the areas of process design and staffing. Having dis-
cussed general modeling procedures common to both manufacturing and service
system simulation in Chapter 6, and specific modeling procedures unique to
manufacturing systems in Chapter 11, in this chapter we discuss design and
operating considerations that are more specific to service systems. A description
is given of major classes of service systems. To provide an idea of how simula-
tion might be performed in a service industry, a call center simulation example
is presented.
345
har01307_ch13_345-362.indd 345 12/17/10 9:48 AM

13.2 Characteristics of Service Systems

Service systems represent a class of processing systems where entities (custom-
ers, orders, work, and so forth) are routed through a series of service stations
and waiting areas. Although certain characteristics of service systems are simi-
lar to manufacturing systems, service systems have some unique characteris-
tics. The aspects of service systems that involve work-flow processing (orders,
paperwork, records, and the like) and product delivery are nearly identical to
manufacturing and will not be repeated here. Those aspects of service systems
that are most different from manufacturing systems are those involving customer
processing. Many of the differences stem from the fact that in service systems,
often both the entity being served and the resource performing the service are
human. Humans have much more complex and unpredictable behavior than parts
and machines. These special characteristics and their implications for modeling
are described here:
• Entities are capricious. System conditions cause humans to change

their minds about a particular decision once it has been made. Customer
reactions to dissatisfactory circumstances include balking, jockeying, and
reneging. Balking occurs when a customer attempts to enter a queue, sees
that it is full, and leaves. Jockeying is where a customer moves to another
queue that is shorter in hopes of being served sooner. Reneging, also
called abandonment, is where a customer enters a waiting line or area,
gets tired of waiting, and leaves. Modeling these types of situations can
become complex and requires special modeling constructs or the use of
programming logic to describe the behavior.
• Entity arrivals are random and fluctuate over time. Rather than scheduling
arrivals as in production, customers enter most service systems randomly
according to a Poisson process (interarrival times are exponentially
distributed). Additionally, the rate of arrivals often changes depending on
the time of day or the day of the week. The fluctuating pattern of arrivals
usually repeats itself daily, weekly, or sometimes monthly. Accurate
modeling of these arrival patterns and cycles is essential to accurate
analysis. Having a fluctuating arrival rate makes most service simulations
terminating simulations with focus on peak activities and other changes of
interest as opposed to steady-state behavior.
• Resource decisions are complex. Typically resource allocation and task
selection decisions are made according to some general rule (like first-
come, first-served). In service systems, however, resources are intelligent
and often make decisions based on more state-related criteria. An increase
in the size of a waiting line at a cashier, for example, may prompt a new
checkout lane to open. A change in the state of an entity (a patient in
recovery requiring immediate assistance) may cause a resource (a nurse
or doctor) to interrupt its current task to service the entity. The flexibility
to respond to state changes is made possible because the resources are
har01307_ch13_345-362.indd 346 12/17/10 9:48 AM

Chapter 13 Modeling Service Systems 347
humans and therefore capable of making more complex decisions than

machines. Because service personnel are often cross-trained and activity
boundaries are flexible, resources can fill in for each other when needed.
Modeling the complex behavior of human resources often requires the use
of if–then logic to define decision rules.
• Resource work pace fluctuates. Another characteristic of human resources
is that work pace tends to vary with time of day or work conditions.
A change in the state of a queue (like the number of customers in
line) may cause a resource (a cashier) to work faster, thereby reducing
processing time. A change in the state of the entity (such as the length
of waiting time) may cause a resource to work faster to complete the
professional service. A change in the state of the resource (like fatigue
or learning curve) may change the service time (slower in the case
of fatigue, faster in the case of a learning curve). To model this variable
behavior, tests must be continually made on state variables in the system
so that resource behavior can be linked to the system state.
• Processing times are highly variable. Service processes vary considerably
due to the nature of the process as well as the fact that the entity and the
server are both human. Consequently, processing times tend to be highly
variable. From a modeling standpoint, processing times usually need to
be expressed using some probability distribution such as a normal or beta
distribution.
• Services have both front-room and back-room activities. In front-room
activities, customer service representatives meet with customers to take
orders for a good or service. In the back room, the activities are carried
out for producing the service or good. Once the service or good is
produced, it is either brought to the customer who is waiting in the front
room or, if the customer has gone, it is delivered to a remote site.
13.3 Performance Measures

The ultimate objectives of a service organization might include maximizing prof-
its or maximizing customer satisfaction. However, these measures for success are
considered external performance criteria because they are not completely de-
termined by any single activity. Simulation modeling and analysis help evaluate
those measures referred to as internal performance criteria—measures tied to
a specific activity. Simulation performance measures are both quantitative and
time-based and measure the efficiency and the effectiveness of a system con-
figuration and operating logic. Examples of internal performance measures are
waiting times, hours to process an application, cost per transaction, and percent-
age of time spent correcting transaction errors. Collier (1994) uses the term in-
terlinking to define the process of establishing quantitative, causal relationships
such as these to external performance measures. He argues that interlinking can
har01307_ch13_345-362.indd 347 12/17/10 9:48 AM

be a powerful strategic and competitive weapon. Here are some typical internal
performance measures that can be evaluated using simulation:
• Service time.
• Waiting time.
• Queue lengths.
• Service level (the percentage of customers who can be promptly serviced,
without any waiting).
• Abandonment rate (the percentage of impatient customers who leave the
system).
13.4 Use of Simulation in Service Systems

The use of simulation in service industries has been relatively limited in the past,
despite the many areas of application where simulation has proven beneficial:
health care services (hospitals, clinics), food services (restaurants, cafeterias),
and financial services (banks, credit unions) to name a few. In health care ser-
vices, Zilm et al. (1983) studied the impact of staffing on utilization and cost; and
Hancock et al. (1978) employed simulation to determine the optimal number of
beds a facility would need to meet patient demand. In food services, Aran and
Kang (1987) designed a model to determine the optimal seating configuration for
a fast-food restaurant, and Kharwat (1991) used simulation to examine restaurant
and delivery operations relative to staffing levels, equipment layout, work flow,
customer service, and capacity. The use of simulation in the service sector con-
tinues to expand.
Even within manufacturing industries there are business or support activities
similar to those found in traditional service industries. Edward J. Kane of IBM
observed (Harrington 1991),
Just taking a customer order, moving it through the plant, distributing these require-
ments out to the manufacturing floor—that activity alone has thirty sub-process steps
to it. Accounts receivable has over twenty process steps. Information processing is a
whole discipline in itself, with many challenging processes integrated into a single
total activity. Obviously, we do manage some very complex processes separate from
the manufacturing floor itself.
This entire realm of support processes presents a major area of potential ap-
plication for simulation. Similar to the problem of dealing with excess inven-
tory in manufacturing systems, customers, paperwork, and information often
sit idle in service systems while waiting to be processed. In fact, the total wait-
ing time for entities in service processes often exceeds 90 percent of the total
flow time.
The types of questions that simulation helps answer in service systems
can be categorized as being either design related or management related. Here
har01307_ch13_345-362.indd 348 12/17/10 9:48 AM

are some typical design and management decisions that can be addressed by
simulation:
Design Decisions
1. How much capacity should be provided in service and waiting areas?
2. What is the maximum throughput capability of the service system?
3. What are the equipment requirements to meet service demand?
4. How long does it take to service a customer?
5. How long do customers have to wait before being serviced?
6. Where should the service and waiting areas be located?
7. How can work flow and customer flow be streamlined?
8. What effect would information technology have on reducing non-
value-added time?
Management Decisions
1. What is the best way to schedule personnel?
2. What is the best way to schedule appointments for customers?
3. What is the best way to schedule carriers or vehicles in transportation
systems?
4. How should specific jobs be prioritized?
5. What is the best way to deal with emergency situations when a needed
resource is unavailable?
Because service systems are nearly always in a state of transition, going from
one activity level to another during different periods of the day or week, they
rarely reach a steady state. Consequently, we are frequently interested in analyz-
ing the transient behavior of service systems. Questions such as “How long is the
transient cycle?” or “How many replications should be run?” become very impor-
tant. Overall performance measures may not be as useful as the performance for
each particular period of the transient cycle. An example where this is true is in
the analysis of resource utilization statistics. In the types of service systems where
arrival patterns and staff schedules fluctuate over the activity cycle (such as by
day or week), the average utilization for the entire cycle is almost meaningless. It
is much more informative to look at the resource utilization for each period of the
activity cycle.
Most design and management decisions in service systems involve answer-
ing questions based on transient system conditions, so it is important that the
results of the simulation measure transient behavior. Multiple replications should
be run, with statistics gathered and analyzed for different periods of the transient
cycle. In an attempt to simplify a simulation model, sometimes there is a tempta-
tion to model only the peak period, which is often the period of greatest interest.
What is overlooked is the fact that the state of the system prior to each period and
the length of each period significantly impact the performance measures for any
particular period, including the peak period.
har01307_ch13_345-362.indd 349 12/17/10 9:48 AM

ProModel provides a periodic reporting option that allows statistics to be

gathered by periods. This provides a more complete picture of system activity
over the entire simulation.
13.5 Applications of Simulation in Service Industries

Some of the primary applications of simulation in service industries are process
design, method selection, system layout, staff planning, and flow control.
13.5.1 Process Design

Because of the number of options in customer or paper processing as well as
the flexibility of largely human resources, service systems can be quite fluid,
allowing them to be quickly and easily reconfigured. Simulation helps identify
configurations that are more efficient. Consider a process as simple as a car wash
service that provides the following activities:
• Payment.
• Wash.
• Rinse.
• Dry.
• Vacuum.
• Interior cleaning.
How many ways can cars be processed through the car wash? Each activity could
be done at separate places, or any and even all of them could be combined at a sta-
tion. This is because of the flexibility of the resources that perform each operation
and the imparticularity of the space requirements for most of the activities. Many of
the activities could also be performed in any sequence. Payment, vacuuming, and
interior cleaning could be done in almost any order. The only order that possibly
could not change easily is washing, rinsing, and drying. The other consideration
with many service processes is that not all entities receive the same services. A car
wash customer, for example, may forgo getting vacuum service or interior cleaning.
Thus it is apparent that the mix of activities in service processes can vary greatly.
Simulation helps in process design by allowing different processing sequences
and combinations to be tried to find the best process flow. Modeling process flow is
relatively simple in most service industries. It is only when shifts, resource pools,
and preemptive tasks get involved that it starts to become more challenging.
13.5.2 Method Selection

Method selection usually accompanies process design. When determining what
activities will be performed and how the process should be broken up, alternative
methods for performing the activities are usually taken into consideration. Some
har01307_ch13_345-362.indd 350 12/17/10 9:48 AM

of the greatest changes in the methods used in service industries have come about
as a result of advances in information technology. Routine banking is being re-
placed by automated teller machines (ATMs). Shopping is being conducted over
the Internet. Electronic document management systems are replacing paper docu-
ment processing. Processing of orders and other documents that previously took
weeks now takes only days or hours.
Automation of service processes presents challenges similar to the automa-
tion of manufacturing processes. If automating a service process speeds up a
particular activity but does not minimize the overall processing time, it is not
effective and may even be creating waste (such as large pools of waiting entities).
13.5.3 System Layout

An important consideration in designing a service system is the layout of the facil-
ity, especially if the building construction is going to be based on the system layout.
For many service operations, however, the layout is easily changed as it may in-
volve only rearranging desks and changing offices. Human resources and desktop
computers are much easier to move around than an anchored 200-ton press used in
sheet-metal processing. Still, a good layout can help provide a more streamlined
flow and minimize the amount of movement for both customers and personnel.
The way in which work areas are laid out can have a significant impact on
customer satisfaction and processing efficiency. In some systems, for example,
multiple servers have individual queues for customers to wait before being
served. This can cause customers to be served out of order of arrival and result in
jockeying and customer discontent. Other systems provide a single input queue
that feeds multiple servers (queuing for a bank teller is usually designed this way).
This ensures customers will be serviced in the order of arrival. It may cause some
reneging, however, if grouping all customers in a single queue creates the percep-
tion to an incoming customer that the waiting time is long.
13.5.4 Staff Planning

A major decision in nearly every service operation pertains to the level and type
of staffing to meet customer demand. Understaffing can lead to excessive waiting
times and lost or dissatisfied customers. Overstaffing can result in needless costs
for resources that are inadequately utilized.
Related to staffing levels is the staff type. In service operations, personnel
are frequently cross-trained so that alternative resources may perform any given
task. One objective in this regard is to make the best use of resources by allowing
higher-skilled resources to fill in occasionally when lower-skilled workers are
unavailable.
Modeling staffing requirements is done by defining the pattern of incoming
customers, specifying the servicing policies and procedures, and setting a trial
staffing level. If more than one skill level is capable of performing a task, the list
and priority for using these alternative resources need to be specified.
har01307_ch13_345-362.indd 351 12/17/10 9:48 AM

After running the simulation, waiting times, abandonment counts, and resource
utilization rates can be evaluated to determine if the optimum conditions have been
achieved. If results are unacceptable, either the incoming pattern, the service force, or
the servicing policies and procedures can be modified to run additional experiments.
13.5.5 Flow Control

Flow control is to service systems what production control is to manufacturing
systems. Service system operations planners must decide how to allow customers,
documents, and so on to flow through the system. As in manufacturing systems,
customers and information may be pushed or pulled through the system. By limit-
ing queue capacities, a pull system can reduce the total number of customers or
items waiting in the system. It also reduces the average waiting time of customers
or items and results in greater overall efficiency.
Fast-food restaurants practice pull methods when they keep two or three food
items (like burgers and fries) queued in anticipation of upcoming orders. When the
queue limit is reached, further replenishment of that item is suspended. When a
server withdraws an item, that is the pull signal to the cook to replenish it with
another one. The lower the inventory of prepared food items, the more tightly
linked the pull system is to customer demand. Excessive inventory results in dete-
riorating quality (cold food) and waste at the end of the day.
13.6 Types of Service Systems

Service systems cover a much broader range than manufacturing systems. They
range from a pure service system to almost a pure production system. Customers
may be directly involved in the service process or they may not be involved at all.
The classification of service systems presented here is based on operational charac-
teristics and is adapted, in part, from the classification given by Schmenner (1994):
• Service factory.
• Pure service shop.
• Retail service store.
• Professional service.
• Telephonic service.
• Delivery service.
• Transportation service.
A description of each in terms of its operating characteristics is given next.
13.6.1 Service Factory

Service factories are systems in which customers are provided services using
equipment and facilities requiring low labor involvement. Consequently, labor
har01307_ch13_345-362.indd 352 12/17/10 9:48 AM

costs are low while equipment and facility costs are high. Service factories usu-
ally have both front-room and back-room activities with total service being pro-
vided in a matter of minutes. Customization is done by selecting from a menu of
options previously defined by the provider. Waiting time and service time are two
primary factors in selecting the provider. Convenience of location is another im-
portant consideration. Customer commitment to the provider is low because there
are usually alternative providers just as conveniently located.
Examples include banks (branch operations), restaurants, copy centers, bar-
bers, check-in counters of airlines, hotels, and car rental agencies.
13.6.2 Pure Service Shop

In a pure service shop, service times are longer than for a service factory. Service
customization is also greater. Customer needs must be identified before service can
be provided. Customers may leave the location and return later for pickup, check-
ing on an order, payment, or additional service. Price is often determined after
the service is provided. Although front-room activity times may be short, back-
room activity times may be long, typically measured in hours or days. The pri-
mary consideration is quality of service. Delivery time and price are of secondary
importance. The customer’s ability to describe the symptoms and possible service
requirements are helpful in minimizing service and waiting times.
When customers arrive, they usually all go through some type of check-in
activity. At this time, a record (paperwork or computer file) is generated for the
customer and a sequence of service or care is prescribed. The duration of the ser-
vice or the type of resources required may change during the process of providing
service because of a change in the status of the entity. After the service is pro-
vided, tests may be performed to make sure that the service is acceptable before
releasing the entity from the facility. If the results are acceptable, the customer
and the record are matched and the customer leaves the system.
Examples include hospitals, repair shops (automobiles), equipment rental
shops, banking (loan processing), Department of Motor Vehicles, Social Security
offices, courtrooms, and prisons.
13.6.3 Retail Service Store

In retail services, the size of the facility is large in order to accommodate many
customers at the same time. Customers have many product options from which
to choose. Retail services require a high degree of labor intensity but a low de-
gree of customization or interaction with the customer. Customers are influenced
by price more than service quality or delivery time. Customers are interested in
convenient location, assistance with finding the products in the store, and quick
checkout. Total service time is usually measured in minutes.
When customers arrive in a retail shop, they often get a cart and use that cart as
a carrier throughout the purchasing process. Customers may need assistance from
customer service representatives during the shopping process. Once the customer
har01307_ch13_345-362.indd 353 12/17/10 9:48 AM

has obtained the merchandise, he or she must get in line for the checkout process.
For large items such as furniture or appliances, the customer may have to order and
pay for the merchandise first. The delivery of the product may take place later.
Examples include department stores, grocery stores, hardware stores, and
convenience stores.
13.6.4 Professional Service

Professional services are usually provided by a single person or a small group
of experts in a particular field. The service is highly customized and provided
by expensive resources. Duration of the service is long, extremely variable, and
difficult to predict because customer involvement during the process is highly
variable.
Processing may be performed by a single resource or multiple resources.
When the customer arrives, the first process is diagnostic. Usually an expert
resource evaluates the service needed by the customer and determines the type
of service, the estimated service time, and the cost. This diagnosis then dictates
what resources are used to process the order. The duration of the service or the
type of resources required may change during the process of providing service.
This is usually a result of the customer’s review of the work. After the service is
provided, a final review with the customer may be done to make sure that the
service is acceptable. If the results are acceptable, the customer and the record are
matched and the customer leaves the system.
Examples include auditing services, tax preparation, legal services, architec-
tural services, construction services, and tailor services.
13.6.5 Telephonic Service

Telephonic services or teleservicing are services provided over the telephone.
They are unique in that the service is provided without face-to-face contact with
the customer. The service may be making reservations, ordering from a catalog,
or providing customer support. In a telephonic service system, there are a number
of issues to address, including
• Overflow calls—the caller receives a busy signal.
• Reneges—the customer gets in but hangs up after a certain amount of
time if no assistance is received.
• Redials—a customer who hangs up or fails to get through calls again.
The most important criterion for measuring effectiveness is delivery time. The cus-
tomer is interested in getting the service or ordering the product as quickly as pos-
sible. The customer’s ability to communicate the need is critical to the service time.
Calls usually arrive in the incoming call queue and are serviced based on the
FIFO rule. Some advanced telephone systems allow routing of calls into mul-
tiple queues for quicker service. Processing of a call is done by a single resource.
Service duration depends on the nature of the service. If the service is an ordering
har01307_ch13_345-362.indd 354 12/17/10 9:48 AM

process, then the service time is short. If the service is a technical support process,
then the service time may be long or the call may require a callback after some
research.
Examples include technical support services (hotlines) for software or hard-
ware, mail-order services, and airline and hotel reservations.
13.6.6 Delivery Service

Delivery services involve the ordering, shipping, and delivery of goods, raw
materials, or finished products to points of use or sale. Customers may accept
deliveries only within certain time schedules. In practice, there are often other
constraints besides time windows. Certain sequences in deliveries may be inflex-
ible. Customers are interested in convenient, fast delivery. If the products that are
delivered are perishable or fragile goods, the quality of the products delivered is
also important to the customer.
Deliveries begin with the preparation of the product and loading of the product
on the delivery resources. Determination of the best routing decisions for drivers
may depend on the number or proximity of customers waiting for the product.
13.6.7 Transportation Service

Transportation services involve the movement of people from one place to an-
other. A fundamental difference between transportation and delivery systems is
that people are being transported rather than goods. Another important difference
is that the routes in transportation services tend to be fixed, whereas the routes in
delivery services are somewhat flexible. Customers are interested in convenient,
fast transportation. Cost of transportation plays a significant role in the selection
of the service. Because set schedules and routes are used in transportation, cus-
tomers expect reliable service.
Two types of systems are used in transportation: (1) multiple pickup and
drop-off points and (2) single pickup and drop-off points. In multiple pickup and
drop-off point systems, customers enter and leave the transportation vehicle inde-
pendently. In single pickup and drop-off transportation, customers all enter at one
place and are dropped off at the same destination.
Examples include airlines, railroads, cruise lines, mass transit systems, and
limousine services.
13.7 Simulation Example: A Help Desk Operation

This section, based on an article by Singer and Gasparatos (1994), presents an
example of a simulation project performed to improve a help desk operation. This
project illustrates the flexibility of processing and how variable some of the data
can be. Even for existing service systems, radical changes can be made with little
disruption that can dramatically improve performance.
har01307_ch13_345-362.indd 355 12/17/10 9:48 AM

13.7.1 Background
Society Bank’s Information Technology and Operations (ITO) group offers
a “help desk” service to customers of the ITO function. This service is offered
to both internal and external customers, handling over 12,000 calls per month.
The client services help desk provides technical support and information on a
variety of technical topics including resetting passwords, ordering PC equipment,
requesting phone installation, ordering extra copies of internal reports, and report-
ing mainframe and network problems. The help desk acts as the primary source of
communication between ITO and its customers. It interacts with authority groups
within ITO by providing work and support when requested by a customer.
The old client services help desk process consisted of (1) a mainframe help
desk, (2) a phone/local area network help desk, and (3) a PC help desk. Each of
the three operated separately with separate phone numbers, operators, and facili-
ties. All calls were received by their respective help desk operators, who manually
logged all information about the problem and the customer, and then proceeded to
pass the problem on to an authority group or expert for resolution.
Because of acquisitions, the increased use of information technologies, and the
passing of time, Society’s help desk process had become fragmented and layered
with bureaucracy. This made the help desk a good candidate for a process redesign.
It was determined that the current operation did not have a set of clearly defined
goals, other than to provide a help desk service. The organizational boundaries of
the current process were often obscured by the fact that much of the different help
desks’ work overlapped and was consistently being handed off. There were no pro-
cess performance measures in the old process, only measures of call volume. A pro-
posal was made to consolidate the help desk functions. The proposal also called for
the introduction of automation to enhance the speed and accuracy of the services.
13.7.2 Model Description

The situation suggested a process redesign to be supported by simulation of the
business processes to select and validate different operational alternatives.
Detailed historical information was gathered on call frequencies, call arrival
patterns, and length of calls to the help desk. This information was obtained from
the help desk’s database, ASIM. Table 13.1 summarizes the call breakdown by
TABLE 13.1 Types of Calls for Help Desk

Percent Percent Percent
Time Period Password Reset Device Reset Inquiries Level 1 Level 1A Level 2
7 A.M.–11 A.M. 11.7% 25.7% 8.2% 45.6% 4.6% 47.3%

11 A.M.–2 P.M. 8.8 29.0 10.9 48.7 3.6 44.3
2 P.M.–5 P.M. 7.7 27.8 11.1 46.6 4.4 45.8
5 P.M.–8 P.M. 8.6 36.5 17.8 62.9 3.7 32.2
Average 9.9% 27.5% 9.9% 47.3% 4.3% 48.4%
har01307_ch13_345-362.indd 356 12/17/10 9:48 AM

call level. Level 1 calls are resolved immediately by the help desk, Level 1A calls
are resolved later by the help desk, and Level 2 calls are handed off to an authority
group for resolution.
Historically, calls averaged 2.5 minutes, lasting anywhere from 30 seconds
to 25 minutes. Periodically, follow-up work is required after calls that ranges
from 1 to 10 minutes. Overall, the help desk service abandonment rate was 4 to
12 percent (as measured by the percentage of calls abandoned), depending on
staffing levels.
The help desk process was broken down into its individual work steps and
owners of each work step were identified. Then a flowchart that described the
process was developed (Figure 13.1). From the flowchart, a computer simulation
model was developed of the old operation, which was validated by comparing ac-
tual performance of the help desk with that of the simulation’s output. During the
10-day test period, the simulation model produced results consistent with those
of the actual performance. The user of the model was able to define such model
parameters as daily call volume and staffing levels through the use of the model’s
Interact Box, which provided sensitivity analysis.
Joint requirements planning (JRP) sessions allowed the project team to col-
lect information about likes, dislikes, needs, and improvement suggestions from
users, customers, and executives. This information clarified the target goals of
the process along with its operational scope. Suggestions were collected and
prioritized from the JRP sessions for improving the help desk process. Internal
benchmarking was also performed using Society’s customer service help desk as
a reference for performance and operational ideas.
FIGURE 13.1
Customer problem, Client notification
Flow diagram of query, or change and escalation
client services.
Automated problem Help desk Automated warning

recognition recognition
Customer self-service
Automated resolution
Qualified support assistance
Expert support assistance
har01307_ch13_345-362.indd 357 12/17/10 9:48 AM

A target process was defined as providing a single-source help desk (“one-

stop shopping” approach) for ITO customers with performance targets of
• 90 percent of calls to be answered by five rings.
• Less than 2 percent abandonment rate.
Other goals of the target process were to enhance the user’s perception of the help
desk and to significantly reduce the time required to resolve a customer’s request.
A combination of radical redesign ideas (reengineering) and incremental change
ideas (TQM) formed the nucleus of a target help desk process. The redesigned
process implemented the following changes:
• Consolidate three help desks into one central help desk.
• Create a consistent means of problem/request logging and resolution.
• Introduce automation for receiving, queuing, and categorizing calls for
resetting terminals.
• Capture information pertaining to the call once at the source, and, if the
call is handed off, have the information passed along also.
• Use existing technologies to create a hierarchy of problem resolution
where approximately 60 percent of problems can be resolved immediately
without using the operators and approximately 15 percent of the calls can
be resolved immediately by the operators.
• Create an automated warning and problem recognition system that detects
and corrects mainframe problems before they occur.
The original simulation model was revisited to better understand the current
customer service level and what potential impact software changes, automation,
and consolidation would have on the staffing and equipment needs and opera-
tional capacity. Simulation results could also be used to manage the expectations
for potential outcomes of the target process implementation.
Immediate benefit was gained from the use of this application of simulation to
better understand the old operational interrelationships between staffing, call vol-
ume, and customer service. Figure 13.2 shows how much the abandonment rate
FIGURE 13.2 0.20

700 calls per day
Abandonment rates 0.18
650 calls per day
for alternative staffing 0.16
Calls abandoned
levels. 0.14 600 calls per day

0.12 550 calls per day
0.04
0.02
0
5 6 7 8 9 10
Operators staffed
har01307_ch13_345-362.indd 358 12/17/10 9:48 AM

will change when the average daily call volume or the number of operators varies.
The importance of this graph is realized when one notices that it becomes in-
creasingly harder to lower the abandonment rate once the number of operators
increases above seven. Above this point, the help desk can easily handle substan-
tial increases in average daily call volume while maintaining approximately the
same abandonment rate.
After modeling and analyzing the current process, the project team evaluated
the following operational alternatives using the simulation model:
• The option to select from a number of different shift schedules so that
staffing can easily be varied from current levels.
• The introduction of the automated voice response unit and its ability to
both receive and place calls automatically.
• The ability of the automated voice response unit to handle device resets,
password resets, and system inquiries.
• The incorporation of the PC and LAN help desks so that clients with PC-
related problems can have their calls routed directly to an available expert
via the automated voice response unit.
• The ability to change the response time of the ASIM problem-logging
system.
Additionally, two alternative staffing schedules were proposed. The alterna-
tive schedules attempt to better match the time at which operators are available
for answering calls to the time the calls are arriving. The two alternative schedules
reduce effort hours by up to 8 percent while maintaining current service levels.
Additional results related to the Alternative Operations simulation model were
• The automated voice response unit will permit approximately 75 percent
of PC-related calls to be answered immediately by a PC expert directly.
• Using Figure 13.2, the automated voice response unit’s ability to aid in
reducing the abandonment rate can be ascertained simply by estimating
the reduction in the number of calls routed to help desk operators and
finding the appropriate point on the chart for a given number of operators.
• Improving the response time of ASIM will noticeably affect the operation
when staffing levels are low and call volumes are high. For example,
with five operators on staff and average call volume of 650 calls per day,
a 25 percent improvement in the response time of ASIM resulted in a
reduction in the abandonment rate of approximately 2 percent.
13.7.3 Results
The nonlinear relationship between the abandonment rate and the number of op-
erators on duty (see Figure 13.2) indicates the difficulty in greatly improving per-
formance once the abandonment rate drops below 5 percent. Results generated
from the validated simulation model compare the impact of the proposed staffing
changes with that of the current staffing levels. In addition, the analysis of the
har01307_ch13_345-362.indd 359 12/17/10 9:48 AM

effect of the automated voice response unit can be predicted before implementa-
tion so that the best alternative can be identified.
The introduction of simulation to help desk operations has shown that it can
be a powerful and effective management tool that should be utilized to better
achieve operational goals and to understand the impact of changes. As the auto-
mation project continues to be implemented, the simulation model can greatly
aid management and the project team members by allowing them to intelligently
predict how each new phase will affect the help desk.
13.8 Summary
Service systems provide a unique challenge in simulation modeling, largely
due to the human element involved. Service systems have a high human con-
tent in the process. The customer is often involved in the process and, in many
cases, is the actual entity being processed. In this chapter we discussed the aspects
that should be considered when modeling service systems and suggested ways in
which different situations might be modeled. We also discussed the different types
of service systems and addressed the modeling issues associated with each. The
example case study showed how fluid service systems can be.

1. What characteristics of service systems make them different from
manufacturing systems?
2. What internal performance measures are linked to customer
satisfaction?
3. Why are service systems more reconfigurable than manufacturing
systems?
4. Identify a service system where the customer is not part of the process.
5. What is abandonment or reneging in simulation terminology?
6. Explain how simulation was used to create the abandonment rate chart
shown in Figure 13.2.
7. What are three questions that simulation might help answer in designing
bus routes for a city bus line?
8. Is a transportation system entity-driven or resource-driven? Explain.
9. Describe conceptually how you would model jockeying in a simulation
model.
References
Aran, M. M., and K. Kang. “Design of a Fast Food Restaurant Simulation Model.” Simula-
tion. Norcross, GA: Industrial Engineering and Management Press, 1987.
Collier, D. A. The Service/Quality Solution. Milwaukee: ASQC Quality Press, 1994.
har01307_ch13_345-362.indd 360 12/17/10 9:48 AM

Hancock, W.; D. C. Magerlein; R. H. Stores; and J. B. Martin. “Parameters Affecting

Hospital Occupancy and Implications for Facility Sizing.” Health Services Research
13, 1978: pp. 276–89.
Harrington, J. H. Business Process Improvement: The Breakthrough Strategy for Total
Quality, Productivity, and Competitiveness. New York: McGraw-Hill, 1991.
Harrington, J. H. Business Process Improvement. New York: McGraw-Hill, 1994.
Kharwat, A. K. “Computer Simulation: An Important Tool in the Fast Food Industry.”
Proceedings of the 1991 Winter Simulation Conference, eds. B. Nelson, W. D.
Kelton, and G. M. Clark. Piscataway, NJ: Institute of Electrical and Electronics
Engineers, 1991.
Sasser, W. E.; R. P. Olsen; and D. D. Wyckoff. Management of Service Operations. Boston:
Allyn & Bacon, 1978, pp. 8–21.
Schmenner, Roger W. Plant and Service Tours in Operations Management. 4th ed. New
York: MacMillan, 1994.
Singer, R., and A. Gasparatos. “Help Desks Hear Voice.” Software Magazine, February
1994, pp. 24–26.
Zilm, F., and R. Hollis. “An Application of Simulation Modeling to Surgical Intensive Care
Bed Needs Analysis.” Hospital and Health Services Administration, Sept/Oct. 1983,
pp. 82–101.
har01307_ch13_345-362.indd 361 12/17/10 9:48 AM

har01307_ch13_345-362.indd 362 12/17/10 9:48 AM
P A R T
II LABS
Lab 1 Introduction to ProModel 365

Lab 2 Building Your First Model 379
Lab 3 ProModel’s Output Viewer 405
Lab 4 Basic Modeling Concepts 419
Lab 5 Fitting Statistical Distributions to Input Data 465
Lab 6 Intermediate Model Building 475
Lab 7 Model Verification and Validation 517
Lab 8 Simulation Output Analysis 527
Lab 9 Comparing Alternative Systems 549
Lab 10 Simulation Optimization with SimRunner 569
Lab 11 Modeling Manufacturing Systems 595
Lab 12 Material Handling Concepts 617
Lab 13 Modeling Service Systems 639
har01307_LB01_363-378.indd 363 12/17/10 9:51 AM

har01307_LB01_363-378.indd 364 12/17/10 9:51 AM
L A B
1 INTRODUCTION TO
PROMODEL
Imagination is the beginning of creation. You imagine what you desire, you will
what you imagine and at last you create what you will.
—George Bernard Shaw
ProModel (Production Modeler) by ProModel Corporation is a simulation tool

for modeling all types of manufacturing systems, service systems, and business
processes. Manufacturing systems such as mass production, assembly lines, job
shops, manufacturing cells, flexible manufacturing systems, conveyors, lift trucks,
cranes, just-in-time systems, kanban systems and others can be modeled by Pro-
Model. Service systems such as hospitals, banks, entertainment parks, emergency
rooms, call centers, warehouse operations, sorting facilities, airport operations,
grocery/retail stores, and so forth, can be quickly and efficiently modeled with
ProModel. Business processes like supply chains, logistic systems, customer ser-
vice management, and information systems can also be modeled by ProModel.
ProModel is a powerful tool in the hands of engineers, process analysts,
and managers to test and evaluate alternative designs, ideas, and process maps
before actual implementation. Resource utilization, production capacity, man-
power needs, productivity, inventory levels, bottlenecks, throughput times, and
other performance measures are used in evaluating alternatives. Improvements
in existing systems or the design of new systems can be modeled and tested be-
fore committing any money, time, or other resources. Various operating strategies
and control alternatives can be compared and analyzed. Simulation enables one
to accurately predict and improve system performance by modeling the actual
or proposed physical location of the manufacturing/service system (such as a
plant floor, a bank lobby, or an emergency room) or an abstract business process.
Through testing various what-if scenarios, one can determine the best (optimum)
way to conduct business.
365
har01307_LB01_363-378.indd 365 12/17/10 9:51 AM

366 Part II Labs
FIGURE L1.1
The ProModel
Shortcut Panel
(student package).
ProModel is a discrete-event simulator and is therefore intended to model

systems where events happen at definite points in time, such as what would occur
in discrete systems. It does, however, also have the capability of modeling the
continuous flow of fluids in and out of tanks. The time resolution is controllable
and ranges from 0.01 hours to 0.00001 seconds.
ProModel uses a graphical user interface (GUI). It is a true Windows–based
simulation tool and utilizes all the Windows features such as standard user
interface, multitasking, built-in printer drivers, and “point-and-click” operation.
ProModel has an online help, index, and search system.
L 1.1 ProModel Opening Screen

ProModel’s opening screen displays a Shortcut Panel with the six items shown in
Figure L 1.1. Please note that depending on the version of the ProModel software
you are using, the actual opening screen and other screens might look slightly dif-
ferent. However, the basic use of ProModel will be essentially the same.
The student version of ProModel is limited to 20 locations, 8 entity types,
8 resource types, 5 attributes, and 15 RTI parameters. Stat::Fit is also limited to
three probability distributions and 100 data points. The lab version of ProModel
has no limitations.
L1.2 ProModel Menu Bar

If you are ready to build a model, you can ignore the Shortcut Panel and begin
accessing the menu bar options for building a model. All of the tools necessary
to build and run a model and view the corresponding output are accessed through
har01307_LB01_363-378.indd 366 12/17/10 9:51 AM

the menu bar (ProModel User’s Guide Chapter 4). The menu bar is located right
below the caption bar and contains the menu items described below.
1. File Menu allows the user to open a new model, open an existing model,
merge two or more submodels into one, save the current model, and
import models created in earlier versions of ProModel.
2. Edit Menu contains selections for editing the contents of edit tables
and logic windows. Importantly, data elements created when building a
model can only be deleted through the edit menu.
3. View Menu contains selections for setting up the model viewing
environment.
4. Build Menu consists of all of the modules for creating and editing a
model, which include the following basic and optional modules:
Basic Modules: Locations, Entities, Processing, and Arrivals.
Optional Modules: Path Networks, Resources, Shifts, Cost, Attributes,
Variables, Arrays, Macros, Subroutines Arrival Cycles, Table Functions,
User Distributions, External Files, and Streams.
5. Simulation Menu controls the execution of a simulation and contains
options for running a model, defining model parameters, and defining
and running scenarios.
6. Output Menu starts the ProModel Output Viewer for viewing model
statistics.
7. Tools Menu contains various utilities as follows:
a. Expression Search
b. Graphic Editor
c. Stat::Fit
d. 3D Animator
e. License Manager
f. Options
g. Customize
8. Window Menu allows you to arrange the windows (or iconized
window) that are currently displayed on the screen so that all windows
are visible at once.
9. Help Menu accesses the ProModel Online Help System.
For more detailed information on the items in the ProModel menu bar, please
refer to the ProModel User’s Guide.
L1.3 ProModel Tool Bars

The ProModel toolbars (ProModel User’s Guide Chapter 4) provide quick
access to many of the options found in the menu bar. Up to nine toolbars
har01307_LB01_363-378.indd 367 12/17/10 9:51 AM

368 Part II Labs
are available in the toolbar space. These toolbars are listed below with short
descriptions for each.
1. File starts a new model, opens a previously saved model, saves an open
model, creates and installs model package.
2. Layout is used to show grids the layout, show hidden path networks, and
make routing paths visible.
3. View is used to define and access specific areas of the layout and
zooming the layout. It is also useful for resetting the window positions
back to a neatly tiled arrangement if moving them around has disheveled
them.
4. Build Basic is used to open basic build modules.
5. Build Advanced is used to open advanced build modules.
6. Simulation opens simulation options and scenarios dialog, plays, pauses,
and stops simulation, toggles animation on/off, and opens output viewer.
7. Simulation Information is used to access location, variable, and array
information. It is also used to define and view dynamic plots.
8. Debug allows the user to pause simulation, trace of filter the steps, and
open the debug dialog box.
9. Tools open the graphic editor, SimRunner, Stat::Fit, and 3D Animator.
For more detailed information on the items in the ProModel tool bar, please refer
to the ProModel User’s Guide.
L1.4 Run-Time Menus and Controls

Once a simulation begins, a new menu bar appears at the top of the screen with
selections for controlling and interacting with the simulation. As shown in
Figure L1.2 (Run-time Menus and Controls), these menu items appear above the
animation speed control bar and simulation clock. The items in the run-time menu
are listed below with short descriptions for each.
1. File menu has only one selection View Text that bring up a window with
a text listing of the current model.
2. Simulation menu is used to either run or pause the simulation.
3. Options menu allows you to interact with the simulation while the model
is running. Various options available are turning the animation on/off,
zooming in or out, selecting a specific view of the model, tracing the
model execution, bringing up the debugger, pausing simulation using a
calendar date and/or clock time.
4. Information menu allows you to see the status of locations in two
different ways, show the current state of all variables and array elements,
and graphically monitor the performance of model elements during run
har01307_LB01_363-378.indd 368 12/17/10 9:51 AM

FIGURE L1.2
Run-time Menus and Controls.
time and store statistical data in an Excel spreadsheet using dynamic

plots.
5. Window menu allows you to rearrange windows and icons and select the
active window.
6. Interact menu displays the Interact dialog box and allows you to execute
interactive subroutines during run time.
7. Help menu contains selections for accessing the ProModel online help
system
For more detailed information on the items in the ProModel tool bar, please refer
to the ProModel User’s Guide.
L1.5 Simulation in Decision Making

ProModel is an excellent decision support tool and is used to help plan and make
improvements in many areas of manufacturing and service industries. Click on
Run demo model from the Shortcut Panel and select any model that looks inter-
esting to run. You can also use the main menu to select File Open and then
choose a model from the Demos subdirectory. To run a model, select Simula-
tion Run (optionally you can press F10 or click on the play button ). To stop a
har01307_LB01_363-378.indd 369 12/17/10 9:51 AM

370 Part II Labs
simulation permanently, select Simulation End Simulation or click on the stop

button . When a simulation ends, you are given the option to view the output
results or return to the model editor.
In this section you will gain some experience running ProModel simulation
before building your own simulation model using ProModel in Lab 2. We present
two example simulation models here—California Cellular and ATM system.
L1.5.1 California Cellular

Problem Statement
At a call center for California Cellular, customer service associates are employed
to respond to customer calls and complaints. The time between two calls is expo-
nentially distributed with a mean of six minutes (10 customers call per hour, on
average). Responding to each call takes a time that varies uniformly from a low
of 2 minutes to a high of 8 minutes, with a mean of 5 minutes. If the company had
a policy that
a. The average time to respond to a customer call should not be any more
than six minutes, how many customer service associates should be
employed by the company?
b. The maximum number of calls waiting should be no more than seven,
how many customer service associates should be employed by the
company?
Average Waiting Time

First we define a model with one customer service associate (CSA) answering
all customer calls (Figure L1.3). The average time a customer has to wait before
reaching the CSA is 18.9 minutes (Figure L1.4). This is a much longer wait time
than is the company policy of six minutes (average).
FIGURE L1.3
California Cellular
with one customer
service agent.
FIGURE L1.4
Customer waiting
time statistics with
one customer service
associate.
har01307_LB01_363-378.indd 370 12/17/10 9:51 AM

FIGURE L1.5
California Cellular
with two customer
service associates.
FIGURE L1.6
Customer waiting
time statistics with
two customer service
associates.
FIGURE L1.7
Number of calls waiting with one customer service agent.
If we have two CSAs (Figure L1.5), we would like to figure out what the av-
erage wait would be. The average wait drops down to 5.9 minutes (Figure L1.6),
which is clearly a much more acceptable result. Hence the decision recommended
to management will be to hire two CSAs.
Maximum Queue Length

Let us now evaluate the policy of no more than five calls waiting for response
at any time. First we look at the results (Figure L1.7) of having one CSA. The
maximum number of customers waiting for response is 18. This is clearly not an
acceptable situation.
har01307_LB01_363-378.indd 371 12/17/10 9:51 AM

372 Part II Labs
With two CSAs working together, the results obtained are shown in Fig-
ure L1.8. The maximum number of customers waiting is 5. When we graph
the number of incoming calls waiting for the duration of the simulation run of
500 hours (Figure L1.9), there were only three occasions of as many as five
calls waiting. Hence, we recommend to management to have two CSAs working
together.
FIGURE L1.8
Number of calls waiting with two customer service agents.
FIGURE L1.9
Graph of number of customers waiting during the simulation run time.
har01307_LB01_363-378.indd 372 12/17/10 9:51 AM

The ProModel model of the California Cellular system is in the Student

subdirectory of ProModel (generally installed under My Documents). Load the
simulation model using the File option on the ProModel menu bar across the
top of the screen. Click on File Open, and then select the Lab 1.4.1_California
Cellular.
L1.5.2 ATM System

In a moment you will load and run a ProModel simulation of an automatic teller
machine (ATM) system described below.
Problem Statement
Customers arrive to use an ATM at an average interarrival time of 3.0 minutes
exponentially distributed. When customers arrive to the system, they join a queue
to wait for their turn on the ATM. The queue has the capacity to hold an infinite
number of customers. Customers spend an average of 2.4 minutes, exponentially
distributed, at the ATM to complete their transactions, which is called the service
time of the ATM.
The objective is to simulate the system to determine the expected waiting
time for customers in the queue (the average time customers wait in line for the
ATM) and the expected time in the system (the average time customers wait in
the queue plus the average time it takes them to complete their transaction at the
ATM).
This is the same ATM system simulated by spreadsheet in Chapter 3
but with a different objective. In Chapter 3 the objective is to simulate only
the first 25 customers arriving to the system. Here no such restriction is ap-
plied. This new objective provides an opportunity for comparing the simu-
lated results with those computed using queuing theory, which is presented in
Chapter 2.
Queuing theory allows us to compute the exact values for the expected time
that customers wait in the queue and in the system. Given that queuing theory can
be used to get exact answers, why are we using simulation to estimate the two
expected values? There are two parts to the answer. First, it gives us an opportu-
nity to verify the simulation model by comparing the simulation output with the
exact results produced using queuing theory. Second, most systems of interest are
too complex to be modeled with the mathematical equations of queuing theory.
In those cases, good estimates from simulation are valuable commodities when
faced with expensive decisions.
Queuing Theory’s Answer to the ATM System

As explained in Chapter 2, queuing theory can be used to calculate the steady
state or long-term expected values of particular performance measures. A system
reaches steady state after it has been running awhile with no changes in the condi-
tions influencing the system. After a stream of customers arrives to the ATM over
har01307_LB01_363-378.indd 373 12/17/10 9:51 AM

374 Part II Labs
time, the system will begin to reach steady state, which means that the statistical
distributions that describe the output of the system stabilize. The steady state topic
is covered in detail in Chapter 8.
Using the queuing theory equations from Chapter 2, the expected time cus-
tomers wait in queue is 9.6 minutes and the expected amount of time customers
spend in the system is 12.0 minutes. Now, let’s use ProModel simulation to esti-
mate these values.
ProModel’s Answer to the ATM System

The ProModel model of the ATM system is in the Student subdirectory. Load
the ATM system simulation model using the File option on the ProModel menu
bar across the top of the screen. Click on File Open, and then select the
Lab 1.4.2_ATM System.
Begin the simulation by pressing the F10 function key. The animation
(Figure L1.10) is very informative and is more interesting to watch than the Excel
spreadsheet simulation in Chapter 3.
The ProModel simulation is programmed to pause the simulation and display
progress messages. To continue the simulation after a message, press the OK but-
ton at the bottom of the message (Figures L1.11 and L1.12).
The simulation results of an average 9.47 minutes in the queue and an aver-
age of 11.86 minutes in the system (Figure L1.13) are close to the steady state
FIGURE L1.10
ATM system simulation
model.
har01307_LB01_363-378.indd 374 12/17/10 9:51 AM

FIGURE L1.11
Displaying a message after the 25th customer has been processed.
FIGURE L1.12
Displaying a message after the 150,000th customer has been processed.
expected values of 9.6 minutes in the queue and 12.0 minutes in the system
computed using queuing theory. You can get closer to the exact values, if needed.
Lab 8 discusses how to achieve a desired level of accuracy.
ProModel displays a prompt at the end of the simulation asking if you want
to collect statistics; select “No” for now (Figure L1.14). ProModel automatically
collects a multitude of statistics over the course of the simulation. You will learn
this feature in Lab 3.
The simulation required only a minute of your time to process 19,496
customers. In Lab 2, you will begin to see how easy it is to build models using
ProModel as compared to building them with spreadsheets.
har01307_LB01_363-378.indd 375 12/17/10 9:51 AM

376 Part II Labs
FIGURE L1.13
Displaying the simulation results.
FIGURE L1.14
ProModel prompt at
the end of a simulation
run.
L1.6 Exercises
1. How do you open an existing simulation model?
2. What is SimRunner? How can you use it in your simulation analysis?
3. What does the Stat::Fit package do? Do you need it when building a
simulation model?
4. At the most, how many locations, entities, and types of resources can be
modeled using the student version of ProModel?
har01307_LB01_363-378.indd 376 12/17/10 9:51 AM

5. Open the Manufacturing Cost model from the Demos subdirectory and
run the model three different times to find out whether one, two, or
three operators are optimal for minimizing the cost per part (the
cost per part is displayed on the scoreboard during the simulation).
Selecting Model Parameters, you can change the number of operators
from the Simulation menu by double-clicking on the first parameter
(number of operators) and entering 1, 2, or 3. Then select Run from the
Model Parameters dialog. Each simulation will run for 15 hours.
6. Without knowing how the model was constructed, can you give a
rational explanation for the number of operators that resulted in the
least cost?
7. Go to the ProModel Web site on the Internet (www.promodel.com).
What are some of the successful real-world applications where
ProModel was used? What other applications are ProModel products
used for besides manufacturing?
8. Identify the ProModel menu where you will find the following items:
a. Save As
b. Delete
c. View Trace
d. Shifts
e. Index
f. General Information
g. Options
h. Printer Setup
i. Processing
j. Scenarios
k. Tile
l. Zoom
9. Which of the following is not a valid ProModel menu or submenu
item?
a. AutoBuild
b. What’s This?
c. Merge
d. Merge Documents
e. Snap to Grid
f. Normal
g. Paste
h. Print Preview
i. View Text
10. Which of the following are valid modeling elements found under the
Build menu?
a. Activities
b. Locations
c. Conveyors
har01307_LB01_363-378.indd 377 12/17/10 9:51 AM

378 Part II Labs
d. Queues
e. Shifts
f. Subroutines
g. Station
h. Server
i. Schedules
j. Arrivals
k. Expressions
l. Variables
m. Create
n. Function
11. What are some of the valid logic statements in the ProModel logic
builder?
12. Describe the functions of the following items in the ProModel Edit
menu:
a. Delete
b. Insert
c. Append
d. Move
e. Move To
f. Copy Record
g. Paste Record
har01307_LB01_363-378.indd 378 12/17/10 9:51 AM

L A B
2 BUILDING YOUR
FIRST MODEL
As far as the laws of mathematics refer to reality, they are not certain; and as far
as they are certain, they do not refer to reality.
—Albert Einstein
In this lab we learn to start building simulation models using ProModel. In section
L2.1 we build our first model of a neighborhood barbershop. In section L2.2 we
build an automatic teller machine system model. In section L2.3 we build a model
for a small furniture factory using multiple locations and multiple entities. In sec-
tion L2.4 we add another location to the previous model. Section L2.6 introduces
the concept of blocking an entity. In section L2.7 we discuss the effects of traffic
intensity on the performance of the system.
L2.1 Building Your First Simulation Model

In this section we describe how you can build your very first simulation model
using ProModel software. We introduce the concepts of locations, entities, entity
arrivals, processes, and routings.
Customers visit the neighborhood barbershop Fantastic Dan for a haircut.
The customer interarrival time is exponentially distributed with an average of 10
minutes. Dan (the barber) takes anywhere from 8 to 10 minutes (depending on the
amount of gossip he has to share with that customer), uniformly distributed (mean
and half-width of 9 and 1 minute respectively) for each haircut. This time also in-
cludes the initial greetings and the transaction of money at the end of the haircut.
Run the simulation model for one day (8 hours or 480 minutes). Find these answers:
a. About how many customers does Dan process per day?
b. What is the average number of customers waiting to get a haircut? What
is the maximum?
379
har01307_LB02_379-404.indd 379 20/12/10 2:02 PM

380 Part II Labs
c. What is the average time spent by a customer in the salon?

d. What is the utilization of Barber Dan?
From the menu bar select File New. In the General Information panel
(Figure L2.1) fill in the title of the simulation model as “Fantastic Dan.” Fill in
some of the other general information about the model like the time and distance
units. Click OK to proceed to define the locations.
From the menu bar select Build Locations. Define two locations—Waiting_
for_Barber and BarberDan (Figure L2.2). Following is a description of rules and
conventions for naming various model elements (ProModel Help System).
FIGURE L2.1
General Information
for the Fantastic Dan
simulation model.
FIGURE L2.2
Defining locations Waiting_for_Dan and BarberDan.
har01307_LB02_379-404.indd 380 20/12/10 2:02 PM

A name or identifier is any combination (up to eighty characters long) of letters, num-
bers, and underscores (“_”), used to identify model elements such as locations, enti-
ties, variables, and functions. Although any valid name can refer to any object, it
is best to use names which describe the object they identify (e.g., using “ClientA”
to describe an entity that represents a client). Names, like all words in ProModel,
are case insensitive, meaning that ProModel sees “PARKING_A,” “Parking_A,” and
“PaRkInG_a” as identical.
Names must use the following conventions:
• Names may contain the letters A through Z (upper or lower case), digits (0–9)
and the underscore “_”. Names may not contain spaces. When opening an older
model, ProModel flags improper use of these restricted characters.
• Do not use a digit as the first character of a name. After the first character, any
character may be a digit. For example, “2Var” is an invalid name, but “V2”, is a
valid name.
• Names may not be keywords, but names may contain keywords. For example,
“For” is an invalid name, but “Forklift” is a valid name. Similarly, “Queue2” and
“OrderQty” are both valid.
• No matter what they identify, every name in a model must be unique. For
example, you cannot name both a location and an entity “Server.” You may,
however, name one “Server1” or “Server_Location” and the other “Server2” or
“Server_Resource.”
Note that the first location is actually a region (it is the icon that looks like a square
box). A region is a boundary used to represent a location’s area. When the model
is running the region is not visible. Also, note that we changed the capacity of this
location to “infinite.” The icon selected for the second location is actually called
operator. We changed the name to BarberDan in the Name column of the Locations
table.
Uncheck the New checkbox on the Graphics panel (Figure L2.3) and click
the text button marked Aa. Click on the location icon in the Layout panel. The
name of the location (BarberDan) appears on the location icon.
Define the entity (Figure L2.4) and change its name to Customer. Define the
processes and the routings the customers go through at the barbershop. All custom-
ers arrive and wait at the location Waiting_for_Barber. Then they are routed to the
location BarberDan. At this location the barber performs the haircut, which takes an
amount of time uniformly distributed between 8 and 10 minutes or Uniform(9,1).
To create the processing and routing tables graphically we need to do the following:
a. Left-click on the entity name in the toolbar and then left-click on the
desired beginning location.
b. Left-click on the destination location. A processing record is created
(Figure L2.5).
c. To add multiple routing lines to the same record, left-click on the Add
Routing button in the toolbox.
d. To route the part to exit, left-click on the Route to Exit button in the
toolbox.
har01307_LB02_379-404.indd 381 20/12/10 2:02 PM

382 Part II Labs
FIGURE L2.3
The Graphics panel.
FIGURE L2.4
Define the entity—Customer.
FIGURE L2.5
Creating the processing and routing tables.
har01307_LB02_379-404.indd 382 20/12/10 2:02 PM

FIGURE L2.6
Completed process and routing tables for Fantastic Dan.
FIGURE L2.7
Processing and routing tables for Fantastic Dan in text format.
FIGURE L2.8
The Logic Builder
menu.
The completed processing and routing tables are shown in Figure L2.6. A text
version of the same processing and routing tables (Figure L2.7) can be viewed by
going to File View Text.
To define the haircut time, click Operation in the Process table. Click the button
with the hammer symbol (Build). A new window named Logic Builder opens up. Se-
lect the command Wait. The ProModel expression Wait causes the customer (entity)
to be delayed for a specified amount time. This is how processing times are modeled.
Click Build Expression. In the Logic window, select Distribution Function
(Figure L2.8). In the Distribution Function window, select Uniform distribution.
har01307_LB02_379-404.indd 383 20/12/10 2:02 PM

384 Part II Labs
Click Mean and select 9. Click Half-Range and select 1. Click Return. Click
Paste. Close the Logic Builder window. Close the Operation window.
Finally the customers leave the barbershop. They are routed to a default loca-
tion called EXIT in ProModel. When entities (or customers) are routed to the EXIT
location, they are, in effect, disposed from the system. All the information associated
with the disposed entity is deleted from the computer’s memory to conserve space.
The distribution functions are built into ProModel and generate random val-
ues based on the specified distribution. Some of the commonly used distribution
functions are shown in Table L2.1.
Now we will define the location and frequency at which customers enter
the model. Open the Arrivals table and click on the location in the layout called
Waiting_for_Barber. This creates an arrival record as shown in Figure L2.9. In
the frequency column, enter E(10), which is every 10 minutes exponentially
distributed.
Next we will define some of the simulation options—that is, run time,
number of replications, warm-up time, unit of time, and clock precision
(Figure L2.10). Let us select Options from the Simulation menu. The run time
is the number of hours the simulation model will run. The number of replica-
tions refers to the number of times the simulation model will be run using dif-
ferent seed values for the random number generator (each time the model will
run for an amount of time specified by run hours). The warm-up time refers
to the amount of time to let the simulation model run to achieve steady-state
behavior. Statistics are collected after the warm-up period is over. The run time
begins at the end of the warm-up period. The unit of time used in the model can
be seconds, minutes, or hours. The clock precision refers to the precision used
when either calculating or randomly generating times for all simulation event
timings.
TABLE L2.1 Commonly Used Distribution Functions
Distribution ProModel Expression
Uniform U(mean, half range)

Triangular T(minimum, mode, maximum)
Exponential E(mean)
Normal N(mean, standard deviation)
FIGURE L2.9
Customer arrival table.
har01307_LB02_379-404.indd 384 20/12/10 2:02 PM

FIGURE L2.10
Definition of
simulation options.
FIGURE L2.11
Screen shot at run time.
Figure L2.11 shows a screen shot during run time. The button in the middle
of the scroll bar at the top controls the speed of the simulation run. Drag it right to
increase the speed and left to decrease the speed.
har01307_LB02_379-404.indd 385 20/12/10 2:02 PM

386 Part II Labs
After the simulation runs to its completion, the user is prompted, “Do you
want to see the results?” (Figure L2.12). Click Yes. Figures L2.13 and L2.14 are
part of the results that are automatically generated by ProModel in the Output
Viewer. To view these tables and charts in a report, simply click on the desired
table or chart in the menu ribbon and drag them into the report space on top of one
of the position spots that appear.
Note that the average time a customer spends waiting for Barber Dan
is 22.95 minutes. The average time spent by a customer in the barbershop is
32.28 minutes. The utilization of the Barber is 89.15 percent. The number of cus-
tomers served in 480 minutes (8 hours) is 47. On average 5.875 customers are
served per hour. The maximum number of customers waiting for a haircut is 8,
although the average number of customers waiting is only 2.3.
FIGURE L2.12
Simulation complete
prompt.
FIGURE L2.13
The Output Viewer
for the Fantastic Dan
model.
har01307_LB02_379-404.indd 386 20/12/10 2:02 PM

FIGURE L2.14
Time Plot of the
number of customers
waiting for Fantastic
Dan.
L2.2 Building the Bank of USA ATM Model

In this section we describe how you can build a simplified version of the auto-
matic teller machine (ATM) system model described in Lab 1, section L1.4.2,
using ProModel software. We also introduce the concept of a queue.
Customers arrive to use a Bank of USA ATM. The average customer
interarrival time is 3.0 minutes exponentially distributed. When customers
arrive to the system they join a queue to wait for their turn on the ATM. The
queue has the capacity to hold an infinite number of customers. Customers
spend an average of 2.4 minutes exponentially distributed at the ATM to com-
plete their transactions, which is called the service time at the ATM. Build
a simulation model of the Bank of USA ATM. Run the simulation model for
980 hours.
a. About how many customers are served per hour?
b. What is the average time spent by a customer in the system?
c. What is the utilization of the ATM?
d. What are the maximum and average numbers of customers waiting in the
ATM queue?
From the menu bar select File New. In the General Information panel
(Figure L2.15) fill in the title of the simulation model as “Bank of USA ATM.”
Fill in some of the other general information about the model like the time and
distance units. Click OK to proceed to define the locations.
har01307_LB02_379-404.indd 387 20/12/10 2:02 PM

388 Part II Labs
FIGURE L2.15
General Information
for the Bank of USA
ATM simulation
model.
FIGURE L2.16
Defining locations ATM_Queue and ATM.
From the menu bar select Build Locations. For this model we will define
two locations—ATM_Queue and ATM (Figure L2.16). The icon for the first lo-
cation (a queue) is selected from the graphics panel. The icon (third from top)
originally looks like a “ladder on its side.” To place it in our model layout, first
left-click on the layout where you want the queue to start. Then move the mouse
pointer to the end of the queue and right-click (or double-click). (Hint: to make
the queue exactly horizontal or vertical, hold the shift key down before right-
clicking.) Change the name of this queue location from Loc1 to ATM_Queue.
Now double-click the ATM_Queue icon on the layout. This opens another win-
dow as shown in Figure L2.17. Make sure to click on the Queue option in the
Conveyor/Queue option window. This causes entities to move at their own speed
har01307_LB02_379-404.indd 388 20/12/10 2:02 PM

FIGURE L2.17
Click on the Queue
option in the
Conveyor/Queue
options window.
FIGURE L2.18
Define the entity—Bank of USA ATM_Customer.
FIGURE L2.19
Processing and routing tables for Bank of USA ATM model.
instead of the speed of the conveyor. However, let’s change the length of the
queue to zero feet as we do not want to model the time it takes to travel the length
of the queue (it is insignificant, anyway).
The icon selected for the second location is actually called brake. We changed
its name to ATM in the name column of the Locations table.
Display the names of each location in the Layout window (Figure L2.16),
using the procedure described in section L2.1.
Define the entity (Figure L2.18) and change its name to ATM_Customer.
Define the processes and the routings (Figures L2.19 and L2.20) the customers go
through at the ATM system. All customers arrive and wait at the location ATM_
Queue. Then they are routed to the location ATM. At this location the customers
deposit or withdraw money or check their balances, which take an average of
2.4 minutes exponentially distributed. Use the step-by-step procedure detailed in
section L2.1 to create the process and routing tables graphically.
har01307_LB02_379-404.indd 389 20/12/10 2:02 PM

390 Part II Labs
FIGURE L2.20
Process and Routing tables for Bank of USA ATM model in text format.
FIGURE L2.21
Customer arrival table.
FIGURE L2.22
Definition of
simulation options.
After using the ATM, the customers route to EXIT causing them to leave
the Bank of USA ATM. All the information associated with the disposed entity is
deleted from the computer’s memory to conserve space. Now we will define the
entity arrival process, as shown in Figure L2.21.
Next we will define some of the simulation options following the procedure
outlined in section L2.1. We will set run time to 980 hours (Figure L2.22). The
har01307_LB02_379-404.indd 390 20/12/10 2:02 PM

FIGURE L2.23
Screen shot at run time.
FIGURE L2.24
The Output Viewer for
the Bank of USA ATM
model.
rest of the options are left at their default values. Figure L2.23 shows a screenshot
of the simulation model during run time.
Figures L2.24 and L2.25 are part of the results that are automatically gener-
ated by ProModel in the Output Viewer.
Note that the average time a customer spends waiting in the ATM Queue
is 9.47 minutes. The average time spent by a customer in the ATM system is
11.86 minutes. The utilization of the ATM is 79.35 percent. Also, 19,496 custom-
ers are served in 980 hours or 19.9 customers per hour. The maximum number
of customers waiting in the ATM Queue is 30, although the average number of
customers waiting is only 3.14. This model is an enhancement of the ATM model
in Lab 1, section L1.4.2. Results will not match exactly as some realism has been
added to the model that cannot be addressed in queuing theory.
har01307_LB02_379-404.indd 391 20/12/10 2:02 PM

392 Part II Labs
FIGURE L2.25
Fluctuation in the
number of customers
waiting in the queue
the Bank of USA ATM
simulation model.
L2.3 Locations, Entities, Processing, and Arrivals

In sections L2.1 and L2.2 we have used the basic elements of ProModel. In this
section we incorporate locations, entities, arrivals, processing, and routing logic
into another simulation model example.
Problem Statement—Multiple Locations, Multiple Entities

Wooden logs are received at the receiving dock of the USA Furniture Fac-
tory at the rate of one every 17 minutes. Logs go to the splitter, where two
pieces are made from each log. The splitting time is Normal(3,1) minutes.
The individual pieces go to the lathe, where they are turned into round legs
[Triangular(3,6,9) minutes]. The round legs go on to a paint booth, where they
are converted to painted legs. Painting time is Exponential(5) minutes. Painted
legs go on to the assembly facility, next door. Consider a material handling
time of one minute between each process. Build a simulation model and run
it for 2,000 hours.
We will start with filling in some of the general information about the model
similar to Figures L2.1 or L2.15. Next we define the locations (Figure L2.26)—
receiving dock, splitter saw, lathe, paint booth, and painted leg store. Note that the
capacity of the receiving dock location has been changed to infinity (INF) to make
sure that all incoming entities (raw material) are allowed into the manufacturing
facility.
Define the entities—logs, piece, round leg, and painted leg (Figure L2.27).
Define the entity arrival process (Figure L2.28)—logs arriving to the location
receiving dock (one every 17 minutes).
Define the processes and the routings (Figure L2.29). All logs arrive at the
receiving dock. From there they go to the splitter. At the splitter each log is cut
har01307_LB02_379-404.indd 392 20/12/10 2:02 PM

FIGURE L2.26
Locations in the USA Furniture Factory.
FIGURE L2.27
Entities in the USA
Furniture Factory.
FIGURE L2.28
Entity arrivals in the USA Furniture Factory.
FIGURE L2.29
Processing and routing tables in the USA Furniture Factory.
har01307_LB02_379-404.indd 393 20/12/10 2:02 PM

394 Part II Labs
FIGURE L2.30
Routing Rule in
the USA Furniture
Factory.
into two pieces. To be able to model this, click on the Rule button in the routing
table. Change the quantity to two (Figure L2.30). This models the process of one
log going into the location splitter saw and two pieces coming out. Pieces are
routed to the lathe, where they are turned into round legs. Round legs are routed
to the paint booth, where they become painted legs. Painted legs are sent to the
painted legs store. Finally, the painted legs are sent to the default location EXIT
for disposal.
The time to move the material between processes is modeled in the Move
Logic field of the Routing table. Four choices of constructs are available in the
Move Logic field:
• MOVE FOR—to move the entity to the next location or in operation
logic, to the end of queue or conveyor, in a specific time.
• MOVE ON—to move the entity to the next location using a specific path
network.
• MOVE WITH—to move the entity to the next location using a specified
resource (forklift, crane). The resource may be freed after the move.
Define some of the simulation options: the simulation run time, the number
of replications, the warm-up time, and the clock precision (Figure L2.31).
Now we go on to the Simulation menu. Select Save & Run. This will save the
model we have built so far, compile it, and also run it. When the simulation model
finishes running, we will be asked if we would like to view the results. Select Yes.
A sample of the results is shown in Figure L2.32.
har01307_LB02_379-404.indd 394 20/12/10 2:02 PM

FIGURE L2.31
Simulation options
in the USA Furniture
Factory.
FIGURE L2.32
Sample of the results
of the simulation run
for the USA Furniture
Factory.
har01307_LB02_379-404.indd 395 20/12/10 2:02 PM

396 Part II Labs
L2.4 Add Location

Now let us add a location to an existing model. For the USA Furniture Factory
example in section L2.3, we will add an oven after the painting booth for drying
the painted legs individually. The drying takes a time that is normally distributed
with a mean of 20 minutes and standard deviation of 2 minutes. After drying, the
painted legs go on to the painted legs store.
To add an additional location to the model, select Locations from the Build
menu. With the mouse, first select any appropriate icon from the Graphics toolbar
(left-click), then left-click again in the Layout window. Change the name of the
location (in the Locations table) from Loc 1 to Oven (Figure L2.33).
Locations can also be added from the Edit menu. Select Append from the
Edit menu. A new location called Loc1 is appended to the end of the Locations
table. Change the name to Oven. Deselect the New option in the Graphics toolbar.
Select an appropriate icon from the Graphics toolbar. Click in the Layout window.
Display the names of each location in the Layout window, using the proce-
dure described in section L2.1. Reposition the names and change their font, color,
and other properties, if desired.
The complete simulation model layout is shown in Figure L2.33. Note that
the capacity of the oven as well as the paint booth has been changed to 10. The
process and routings are created using the step-by-step procedure detailed in sec-
tion L2.1 and are shown in Figure L2.34. At this point it is worthwhile to note the
differences between the capacity and the units of a location.
• Capacity—This is the number of units of entities that the location can
hold simultaneously. The default capacity of a location is one.
FIGURE L2.33
Locations and layout of the USA Furniture Factory.
har01307_LB02_379-404.indd 396 20/12/10 2:02 PM

FIGURE L2.34
Processing and routing tables at the USA Furniture Factory.
• Units—A multiunit location consists of several identical units that are

referenced as a single location for processing and routing purposes. A
multiunit location eliminates the need to create multiple locations and
multiple processes for locations that do the same thing. The default
number of units of a location is one.
The contents of a location can be displayed in one of the following two alter-
native ways:
a. To show the contents of a location (Oven) as a counter, first select a
location (left-click). Next deselect the New option from the Graphics
toolbar. Left-click on the command button 00 (counter icon) in the
Graphics toolbar (left column, top). Finally, left-click on the location
selected. The location counter will appear in the Layout window next to
the location Oven (Figure L2.33).
b. To show the contents of a location (Paint Booth) as a gauge, first select
a location (left-click). Next deselect the New option from the Graphics
toolbar. Left-click on the second command button (gauge icon) from the
top in the left column in the Graphics toolbar. The gauge icon will appear
in the layout window next to the location Paint Booth (Figure L2.33).
The fill color, fill direction, scale and other properties of the gauge can
now be changed, if needed.
L2.5 Effect of Variability on Model Performance

Variability in the data makes a big impact on the performance of any system. Let
us take the example of Fantastic Dan, the barbershop in section L2.1. Assume that
one customer arrives for getting a haircut exactly every 10 minutes. Also, Dan
takes exactly nine minutes for each haircut.
Let us modify the model of section L2.1 to reflect these changes. The arrival
and the processing tables are changed as shown in Figures L2.35 and L2.36, re-
spectively. A summary of the results of these two models (with and without vari-
ability) is shown in Table L2.2. This analysis clearly shows the adverse effects of
variability in the system on its performance.
har01307_LB02_379-404.indd 397 20/12/10 2:02 PM

398 Part II Labs
FIGURE L2.35
Customer arrival for haircut at Fantastic Dan.
FIGURE L2.36
Processing of customers at Fantastic Dan.
TABLE L2.2 Comparison of the Fantastic Dan Barbershop

with and without Variability
With Variability Without Variability

Statistics on System Performance (Section L2.1) (Section L2.5)
Average customer time at the barbershop 32.28 min. (Figure L2.13) 9 min.
Average waiting in line for the barber 22.95 min. (Figure L2.13) 0 min.
Average size of the waiting line 2.3 customers 0 customers
Maximum size of the waiting line 8 customers 1 customer
Average number of customers served per hour 5.875 customers/hr. 6 customers/hr.
L2.6 Blocking
With respect to the way statistics are gathered, here are the rules that are used in
ProModel (ProModel Users Manual, Report Data).
1. Average Time in System. The average total time the entity spends in the
system.
2. Average Time in Move Logic. The average time the entity spent
traveling between locations, including any delays incurred in move logic.
3. Average Time Waiting. The average time the entity spent waiting for
a resource or another entity (to join or combine). Also includes time
waiting in queue behind a blocked entity.
4. Average Time in Operation. The average time the entity spent
processing (i.e., WAIT or USE statements) at a location or traveling on a
conveyor/queue.
5. Average Time Blocked. The average time the entity spent waiting for
a destination location to have available capacity. Any entities held up
behind another blocked entity are actually waiting on the blocked entity,
so they are reported as “time waiting for resource, etc.”
har01307_LB02_379-404.indd 398 20/12/10 2:02 PM

FIGURE L2.37
Layout of the Socal
Machine Shop.
FIGURE L2.38
Processing and Routing tables for the Socal Machine Shop.
Example
At the Socal Machine Shop (Figure L2.37) gear blanks arriving to the shop wait in
a queue (Incoming_Q) for processing on a turning center and a mill, in that order.
A total of 100 gear blanks arrive at the rate of one every eight minutes. The pro-
cessing times on the turning center and mill are eight minutes and nine minutes,
respectively. Develop a simulation model and run it.
To figure out the time the “gear blanks” are blocked in the machine shop,
waiting for a processing location, we have entered “Move for 0” in the operation
logic (Figure L2.38) of the location Incoming_Q. Also, the decision rule for the
queue has been changed to “No Queuing” in place of FIFO (Figure L2.39). This
way all the entities waiting in the queue for the turning center to be freed up are
reported as blocked. When you specify FIFO as the queuing rule for a location,
only the lead entity is ever blocked (other entities in the location are waiting for
the lead entity and are reported as “wait for resource, etc.”).
From the entity activity statistics (Figure L4.40) in the output report we can
see the entities spend on average 66.5 minutes in the system, of which 49.5 min-
utes are blocked (waiting for another process location) and 17 minutes are spent in
operation (eight minutes at the turning center and nine minutes at the mill). Block-
ing as a percentage of the average time in system is 74.4 percent. The utilization
of the turning center and the mill are about 98 percent each.
har01307_LB02_379-404.indd 399 20/12/10 2:02 PM

400 Part II Labs
FIGURE L2.39
Decision rules for
Incoming_Q.
FIGURE L2.40
Entity activity statistics at the Socal Machine Shop.
FIGURE L2.41
Entity activity statistics at the Socal Machine Shop with two mills.
In general, the blocking time as a percentage of the time in system increases

as the utilization of the processing locations increases. Also, as the utilization
of the processing locations increases the average time in system as well as the
congestion (number of customers) in the system increases. This topic is explored
in more details in the following section L2.7. To reduce blocking in the machine
shop, let us install a second mill. In the location table, change the number of
units of mill to 2. The entity activity statistics from the resulting output report are
shown in Figure L2.41. As expected, the blocking time has been reduced to zero.
har01307_LB02_379-404.indd 400 20/12/10 2:02 PM

L2.7 Effect of Traffic Intensity on System Performance

Traffic intensity (␳) is defined as the ratio of the average number of customer
arrival (␭) and average number of customer served (␮) per unit time. We would
like to demonstrate the effects of an increase in the traffic intensity (more custom-
ers arriving relative to the number of customers being served per unit time) on
the performance of the system. Here we have used the average time a customer
spends in the system and also the congestion (number of customers present—
either waiting for service or being served) in the system as metrics for evaluating
system performance.
Problem Statement
Atithi Restaurant has one drive-through window. The average number of cus-
tomers arriving per hour (exponentially distributed) is variable. The service time
is normally distributed with an average of 5 minutes and a standard deviation of
1 minute. We would like to investigate the effects of customer arrival rate on the
utilization of the server, the time a customer spends in the drive-through facility
of the restaurant (waiting for service and getting served) and also on the conges-
tion of customers in the drive-through facility. Figure L2.42 shows the processing
and routing tables for the simulation model of the drive-through facility of the
Atithi Restaurant. Figure L2.43 shows a snapshot of the simulation model of the
FIGURE L2.42
Processing and Routing tables for Atithi Restaurant model.
FIGURE L2.43
Snapshot of the
Simulation Model
of the drive-through
facility of the Atithi
Restaurant.
har01307_LB02_379-404.indd 401 20/12/10 2:02 PM

402 Part II Labs
TABLE L2.3 Effects of Customer Arrival Rate on Server Utilization, Time

in System, and Number of Customers in the System
Arrival Service Average

Rate Rate Server Average Time Number
(customers/ (customers/ Traffic Utilization in System of Customers
hour) hour) Intensity (%) (min) in System
2 12 0.167 17.61 5.54 0.2

3 12 0.25 22.56 5.71 0.27
4 12 0.333 31.02 6.14 0.38
6 12 0.5 45.69 7.53 0.69
8 12 0.667 65.17 11.56 1.53
10 12 0.833 91.3 16.84 2.98
11 12 0.917 95.83 21.3 4.11
12 12 1 99.25 28.91 5.68
drive-through facility of the Atithi Restaurant. Table L2.3 shows the effects of
customer arrival rate on server utilization, time in system, and number of custom-
ers in the system.
As can be seen, as the traffic intensity (␳) exceeds 0.8 (80%) server utiliza-
tion flattens out while the average time in system and the number of customers
in the system grows very rapidly to unacceptable levels. For server utilizations
higher than about 80 percent the time a customer spends in the system grows
to very high (and unacceptable) levels. All this analysis is meant to show that
a very high traffic intensity (or very high server utilization) is not necessarily a
good thing.
L2.8 Exercises
1. For the Atithi Restaurant example in section L2.7, make plots (in
EXCEL) to show the effects of:
a. Increasing traffic intensity (increasing arrival rates for a fixed service
rate) on the server utilization.
b. Increasing traffic intensity on the average time a customer spends in
the drive-through facility.
c. Increasing traffic intensity on the average number of customers in the
drive-through facility.
d. Increasing server utilization on the average time a customer spends in
the system.
What happens when the traffic intensity (␳) exceeds 0.8 (80%)?
har01307_LB02_379-404.indd 402 20/12/10 2:02 PM

2. For the Fantastic Dan example in section L2.1, run the simulation model
for a whole year (250 days, 8 hours each day) and answer the following
questions:
a. On average how many customers does Dan serve each day?
b. What is the average number of customers waiting to get a haircut?
What is the maximum?
c. What is the average time spent by a customer in the salon?
d. What is the utilization of Barber Dan? Is it satisfactory?
e. How many chairs should Dan have for customers waiting for a
haircut? Why?
3. If Dan could take exactly 9 minutes for each haircut will it improve the
situation? Rework question 2 and answer parts a through e.
4. For the SoCal Machine Shop example in section L2.6, run the simulation
model for a whole year of operations (2,500 hours) and respond to the
following:
a. What is the average and maximum number of gear blanks in the
Incoming_Q? Is it satisfactory? Why or why not?
b. What are the utilizations of the Turning Center and the Mill? Are they
satisfactory? Why or why not?
c. What is the average time a gear blank stays in the shop?
d. What is the average time the gear blank is blocked (waiting for
another process location)?
e. How will all these results change when we decide to install a
second mill?
5. In the Arrivals element (Build menu), there are two items—occurrences
and frequency. What are the differences between these two items? What
is the significance of the Qty each column? What is the significance of
the First Time column? When and how can we use it in a simulation
model?
6. What are the various time units that you can use while developing a
model? Where do you provide this information?
7. The Processing element (Build menu) has two tables that need to be
edited. What are their names and their functions? How do we enter
information in those tables? How does the Processing editor work?
8. When an entity has completed all the processing in the system, what
should we do with it? Where should it be routed? Why?
9. Differentiate between the following:
a. Entity vs. Locations
b. Locations vs. Resources
c. Attributes vs. Variables
d. Subroutines vs. Macros
e. Table Functions vs. Arrays
f. Save vs. Save As
har01307_LB02_379-404.indd 403 20/12/10 2:02 PM

404 Part II Labs
10. What are some of the differences between the following logic
statements?
a. Wait vs. Wait Until
b. Move vs. Move For
c. Pause vs. Stop
d. Accum vs. Group
e. End vs. Stop
f. Do Until vs. Do While
g. Inc vs. Int
11. Describe the differences between the following items in the ProModel
View menu:
a. Zoom vs. Zoom to Fit Layout
b. Show Grid vs. Snap to Grid
har01307_LB02_379-404.indd 404 20/12/10 2:02 PM

L A B
3 PROMODEL’S OUTPUT
VIEWER
I am not discouraged, because every wrong attempt discarded is another step

forward.
—Thomas Edison
In this lab we introduce you to the Output Viewer in ProModel. Section L3.1 de-
scribes the various menus in the Output Viewer. In sections L3.2 and L3.3 we de-
scribe the various tables and charts for comparing table information for particular
columns. In sections L3.4 and L3.5 we describe utilization and state charts. Sec-
tion L3.6 presents the different time-series charts available in the Output Viewer.
Section L3.7 is an introduction to the dynamic plots that can be displayed and
updated dynamically during simulation run time.
L3.1 The Output Viewer

The ProModel Output Viewer can be run either from ProModel by choosing View
Statistics from the Output menu (Figure L3.1) or from the Windows Program
Manager by choosing Output Viewer (Figure L3.2). This allows us to view Pro-
Model output files at any time, with or without starting ProModel. (Choose Output
Viewer 3DR only if you prefer running the older Output Viewer.) Some examples
of output views were shown in Lab 2, Figures L2.13, L2.14, L2.24, L2.25, and
L2.32. The Output Viewer simplifies the process of generating output reports and
analyzing the output data. We encourage you to use the ProModel’s online help to
learn more about the Output Viewer.
The Output menu in ProModel (Figure L3.1) has the following options:
• View Statistics. Allows the user to view the statistics generated from
running a simulation model. Selecting this option loads the Output Viewer.
405
har01307_LB03_405-418.indd 405 12/17/10 9:53 AM

406 Part II Labs
FIGURE L3.1
Output menu options
in ProModel.
FIGURE L3.2
Output Viewer in
Windows Program
Manager.
FIGURE L3.3
File menu in the
Output Viewer.
• Minitab. Exports the output to Minitab to perform a six-sigma analysis.

• View Trace. Allows the user to view the trace file generated from
running a simulation model. Sending a trace listing to a text file during
runtime generates a trace. Please refer to Lab 7, section L7.2, for a more
complete discussion of tracing a simulation model.
The File menu (Figure L3.3) allows the user to open one or more output databases
for review, analysis, and comparison. Output can be viewed for individual replica-
tions or for average values across multiple replications. Statistics from different
scenarios and even completely different models can be displayed together on the
same chart for comparison.
Output reports can be printed or copied using Windows’ screen capture func-
tionality or by exporting the data to Excel using the Export tab and printing it as
an Excel table or chart.
har01307_LB03_405-418.indd 406 12/17/10 9:53 AM

FIGURE L3.4
Charts menu in the
Output Viewer.
The Charts menu (Figure L3.4) allows the user to select the tables and charts to
display in a report. To place a table or chart in a report, select the desired table or
chart from the Chart menu and drag it to one of the position indicators that appear
on the report. You can create as many custom reports as you wish. Just click on a
new tab to create a new report. To move charts within a report, just click and drag
on the chart caption bar. To move charts to another report, drag it over the desired
report tab until the report opens and then drag it to the desired position in the report.
The filter that appears on the left side of the Output Viewer is used to cus-
tomize the content of a table or chart. In a Location Summary table, for example,
you can select which locations to display. You can also select which columns of
information to display for the table. The filter also allows you to select which
replications and scenarios to display. For multiple replications, you can display
average, minimum and maximum values, or a confidence interval value range.
Using the Format and Style tabs at the top of the Output Viewer can change
the appearance of tables and charts. These tabs only appear when a table or chart
is selected in a report. You can also right-click on a chart to display visual options.
Once you have a report or set of reports looking just the way you want, you
can save them using the Save or Save As options under the File menu. The next
time you run the model, the exact same report will be displayed as before, only
updated to reflect any model changes that have been made.
The Export menu allows data and charts to be copied to the clipboard to paste into
other applications. It also allows you to export reports directly into Microsoft Excel.
L3.2 Report Tables

A variety of tables can be displayed in Output Viewer. Some of the more com-
monly used tables include the following:
a. Entity Summary
b. Entity States
c. Entity Costs
d. Entity Failed Arrivals
e. Location Summary
f. Location States
g. Location Costs
h. Resource Summary
i. Resource States
har01307_LB03_405-418.indd 407 12/17/10 9:53 AM

408 Part II Labs
j. Resource Costs
k. Variable Summary
l. Simulation Information
m. Log Information
n. Node Entries
See the online help system for an explanation of the columns displayed in each
of these tables.
L3.3 Column Charts

A Column chart (Figures L3.5 and L3.6) is essentially a bar chart of a selected
column from one of the data tables described in section L3.2. It is useful for visu-
ally comparing the performance of different entities, locations, or resources in a
FIGURE L3.5
A column chart
showing average
contents for different
locations.
FIGURE L3.6
A column chart
showing average
time in the system for
different entities.
har01307_LB03_405-418.indd 408 12/17/10 9:53 AM

model. For example, you might want to compare the average contents of different
locations (Figure L3.5) or the average time in the system for different entities
(Figure L3.6). The examples given in these and the other figures that follow are
taken from the mfg_cost.mod model in the Demo subdirectory. This model was
run for 15 hours and five replications, so all of the charts represent values aver-
aged across all five replications.
L3.4 Utilization Charts

In ProModel the utilization chart shows the percentage of utilization for either lo-
cations or resources. Figure L3.7 shows an example of a location utilization chart.
Location utilization represents the percentage of the location’s capacity that was
used during the simulation. Resource utilization represents the percentage of time
a resource was in use. A resource is utilized when it is transporting or processing
an entity or servicing a location or other resource.
L3.5 State Charts

State charts show the percentage of time that entities, locations, and resources were
in a particular state during the simulation (e.g., in operation, waiting, blocked,
down, etc.). Entities, locations, and resources have slightly different states they
can be in, and even single-capacity locations differ somewhat from multicapac-
ity locations. State charts can be displayed as either stacked-bar charts or as pie
charts. The advantage of displaying them as stacked-bar charts is that you can see
a better side-by-side comparison.
• Entity States: The following information is given for each entity
type: % in Move Logic, % Waiting for Resources, % in Operation,
and % Blocked. Figure L3.8 shows an example of an entity state
chart.
FIGURE L3.7
Utilization chart for
locations.
har01307_LB03_405-418.indd 409 12/17/10 9:53 AM

410 Part II Labs
FIGURE L3.8
Entity state chart.
FIGURE L3.9
Entity state pie chart.
Figure L3.9 shows an entity state pie chart for the entity Cog that was shown in
Figure L3.8.
• Single-Capacity Location States: Single-capacity location state graphs
show the percentage of time that each location in the system was either
in operation, being set up, idle, waiting for resources, blocked, or down
(Figure L3.10).
• Multicapacity Location States: This chart shows the percentage of
time that each multicapacity location in the system was empty, partially
occupied, or full (Figure L3.11).
• Resource States: This chart shows the percentage of time a resource
was in use, traveling to the point of use, traveling to park, idle, and down
(Figure L3.12).
har01307_LB03_405-418.indd 410 12/17/10 9:53 AM

FIGURE L3.10
Single-capacity
location state chart.
FIGURE L3.11
Multicapacity location
state chart.
FIGURE L3.12
Resource state chart.
har01307_LB03_405-418.indd 411 12/17/10 9:53 AM

412 Part II Labs
L3.6 Time-Series Charts

Time-series statistics can be collected for locations, entities, resources, and vari-
ables. Available time-series charts include time plots, histograms, entity counts,
and location states. Figure L3.13 shows an example of a time plot for the contents
of a queue. Since it is based on five replications, the average contents per minute
averaged across all replications was selected for display.
A histogram indicates the percentage of time (if time weighted) or the per-
centage of times (if observation based) that the value of a variable or parameter
fell within different ranges. Figure L3.14 shows a histogram of the time-weighted
contents of a queue. Using the filter, additional statistics have also been selected to
display on the histogram including the mean and the 95th percentile. The percen-
tile shows that 95 percent of the time the contents never exceeded 58.4 averaged
across all replications.
Another example of a time-series chart is the entity or throughput count
that charts the number of entity exits per time period during the simulation
(Figure L3.15).
FIGURE L3.13
Time plot of the
contents of a queue.
FIGURE L3.14
Histogram of contents
of a queue.
har01307_LB03_405-418.indd 412 12/17/10 9:53 AM

FIGURE L3.15
An hourly count of
entities processed.
L3.7 Dynamic Plots

Dynamic plots allow you to select certain metrics for various model elements and
observe value changes for those metrics dynamically as the model runs. Configu-
rations of one or more plot windows can be saved and later retrieved to quickly
view a customized set of graphs.
Features include:
• Up to six elements may be graphed on the same chart at the same time.
• More than one chart may be active at a time.
• Charts can be resized.
• Chart settings may be saved to use over from one simulation run to the
next.
• Dynamic plots work with multiple replications and scenarios.
The Dynamic Plot dialog is only accessible during simulation run time. If you will
be creating saved chart views, you may wish to pause the simulation, define Chart
views, then resume the simulation. Dynamic plots are set up using the Informa-
tion | Dynamic Plots menu (Figure L3.16) during simulation. Under Dynamic
Plots you will find two sub-items, New and Configurations.
New opens a window similar to the one in Figure L3.17. The tree view rep-
resents a hierarchical list of all plottable items in the model. Expanding and col-
lapsing of tree items works in a manner consistent with standard Windows tree
FIGURE L3.16
Dynamic Plots Menu.
har01307_LB03_405-418.indd 413 12/17/10 9:53 AM

414 Part II Labs
FIGURE L3.17
Dynamic Plots Menu.
controls, including, but not necessarily limited to, double-clicking on a parent

item, clicking on plus or minus graphic next to a parent item, and using the plus
and minus keys on the number pad of most keyboards. When a subtree is fully
expanded, individual plottable items are shown (Figure L3.17).
Each plottable item displays a checkbox to its left. When checked, the item is
added to the list of currently plotted items below the tree view and the chart view
begins to plot the value of the item. When unchecked, the item is removed from
the list of currently plotted items, and the item is removed from the chart view.
Items can also be added or removed from the chart view by double-clicking the
item. If the item was not checked, this will check the box next to the item, add it
to the current item list and the chart view, and switch immediately to the chart dis-
play. If the item was checked, this will uncheck the box next to the item, remove it
from the current item list and the chart view, and switch immediately to the chart
display.
The maximum number of statistics that can be plotted on a single chart is six.
When six items are being charted and you attempt to add an additional item by
either clicking the box next to the item or by double-clicking the item, nothing
will happen and no changes to the chart will take place.
The chart view is where graphical representations of changing model ele-
ments are displayed. It consists of a chart area and, optionally, any combination
of the following: toolbar, legend, x- and y-axis scales, x- and y-axis labels, verti-
cal and horizontal gridlines, and a chart title. An example of a chart view of the
current contents of the location Bearing Queue (Mfg_Cost.mod from the Demo
subdirectory) is shown in Figure L3.18.
har01307_LB03_405-418.indd 414 12/17/10 9:53 AM

FIGURE L3.18
Chart View in
Dynamic Plots—
Current Contents of
the Bearing Queue.
L3.8 Exercises
1. Customers arrive at the Lake Gardens post office for buying stamps,
mailing letters and packages, and so forth. The interarrival time is
exponentially distributed with a mean of two minutes. The time to
process each customer is normally distributed with a mean of 10 minutes
and a standard deviation of two minutes.
a. Make a time-series plot of the number of customers waiting in line at
the post office in a typical eight-hour day.
b. How many postal clerks are needed at the counter so that there are
no more than 15 customers waiting in line at the post office at any
time? There is only one line serving all the postal clerks. Change the
number of postal clerks until you find the optimum number.
2. The Lake Gardens postmaster in Exercise 1 wants to serve her customers
well. She would like to see that the average time spent by a postal
customer at the post office is no more than 15 minutes. How many postal
clerks should she hire?
3. For the USA Furniture Factory example in Lab 2, section L2.3,
a. Make a state graph and a pie graph for the splitter and the lathe.
b. Find the percentage of time the splitter and the lathe are idle.
har01307_LB03_405-418.indd 415 12/17/10 9:53 AM

416 Part II Labs
4. For the USA Furniture Factory example in Lab 2, section L2.3,

a. Make histograms of the contents of the oven and the paint booth.
Make sure the bar width is set equal to one. What information can
you gather from these histograms?
b. Make a dynamic plot of the contents of the oven and the paint booth
as the simulation model is running.
c. Plot a pie chart for the various states of the entity Painted_Logs. What
percentage of time the Painted_Logs are in operation?
d. Make a time-series plot of the oven and the paint booth contents.
How would you interpret these plots?
5. For the Bank of USA ATM example in Lab 2, section L2.2,
a. Plot a pie chart for the various states of the ATM customer.
b. What is the percentage of time the ATM customer is in process (using
the ATM)?
6. Run the Tube Distribution Supply Chain example model (logistics.
mod from the Demo subdirectory) for 40 hours. What are the various
entities modeled in this example? What are the various operations and
processes modeled in this example? Look at the results and find:
a. The percentage utilization of the locations Mill and the Process
Grades Threads.
b. The capacities of Inventory and Inventory 2-6. The maximum
contents of Inventory and Inventory 2-6.
c. The idle time percent of the location Mill.
d. The percent utilization of the resources Transport 1-5 and Fork Lift 1-6.
e. What are the various states (in percentages) of the location Process
Grades Threads?
7. Run the Warehouse model (deaerco.mod from the Demo subdirectory)
for 2,000 hours. Run only Scenario 1. Go into the Simulation Options
menu and change the Run hours from 10 to 2,000.
a. What are the average values of inventory of Inventory Aisle 1-12?
b. What is the average Time to Fill, Box, and Check?
c. What is the Average Time in System?
d. What are the cost of customer order, receiving order, and the sum
total of cost per order?
e. What are the percentage idle times for Checker 1 and Checker 2?
f. What are the utilizations of Boxers 1-4 and the average utilization of
all the Boxers?
g. Make a dynamic plot of the current contents of the Order Table location.
8. Run the Receiving Dock model (receive.mod from the Demo subdirectory)
for one year. Go to the Simulation Options menu and select Calendar
Date as the Run Length option. Then select simulation begin and end dates
that are one year apart. Answer the following questions:
a. What are the utilizations of the Driver Door, Box Handling, and the
three Docks?
har01307_LB03_405-418.indd 416 12/17/10 9:53 AM

b. Make a time-weighted plot of the Box Staging Contents. Comment on

the graph.
c. Make a histogram of the Traffic Desk Contents. Comment on the
graph.
d. What is the percentage utilization of the Clerk and the Supervisor?
e. What is the average utilization of the three forks?
f. What is the pallet load throughput?
g. What are the average and maximum cycle times for semitruck, small
truck, pallet loads, and items?
h. What is average number of pallets put away per hour?
i. What is the average number of items put away per hour?
j. What is the average number of items per load?
har01307_LB03_405-418.indd 417 12/17/10 9:53 AM

har01307_LB03_405-418.indd 418 12/17/10 9:53 AM
L A B
4 BASIC MODELING
CONCEPTS
Imagination is more important than knowledge.

—Albert Einstein
In this chapter we continue to describe other basic features and modeling con-
cepts of ProModel. In section L4.1 we show an application with multiple lo-
cations and multiple entity types. Section L4.2 describes modeling of multiple
parallel locations. In section L4.3 we introduce you to the concepts of resources.
Section L4.4 shows various routing rules. In section L4.5 we introduce the con-
cept of variables. Section L4.6 introduces the inspection process, tracking of
defects and rework. In section L4.7 we show how to permanently or temporar-
ily batch entities of the same type or different types. In section L4.8 we show
the process of making permanent and temporary attachments of one entity to
another. Section L4.9 describes how similar entities can be accumulated before
processing. Section L4.10 shows the splitting of one entity into multiple entities.
In section L4.11 we introduce various decision statements with appropriate ex-
amples. Finally, in section L4.12 we show you how to model a system that shuts
down periodically.
L4.1 Multiple Locations, Multiple Entity Types

Problem Statement
In one department at Pomona Electronics, three different printed circuit boards
are assembled. Each board is routed through three assembly areas. The routing
order is different for each of the boards. Further, the time to assemble a board
depends on the board type and the operation. The simulation model is intended to
determine the time to complete 500 boards of each type. The assembly time for
419
har01307_LB04_419-464.indd 419 20/12/10 1:53 PM

420 Part II Labs
TABLE L4.1 Routing and Assembly Times for the Printed Circuit Boards
Printed Circuit Board 1 Printed Circuit Board 2 Printed Circuit Board 3
Area Mean Time Area Mean Time Area Mean Time
1 10 2 5 3 12
2 12 1 6 2 14
3 15 3 8 1 15
FIGURE L4.1
Three processing locations and the PCB Storage.
FIGURE L4.2
Layout of Pomona
Electronics.
each board is exponentially distributed with the mean times (in minutes) shown
in Table L4.1
Define three locations (area1, area2, and area3) where assembly work is
done and another location, PCB_Store, where all the printed circuit boards are
stored for future assembly (Figure L4.1). Assume each of the assembly areas
has infinite capacity. The layout of Pomona Electronics is shown in Figure L4.2.
Note that we used Background Graphics Behind Grid, from the Build menu,
to add the Pomona Electronics logo on the simulation model layout. Add the
robot graphics (or something appropriate). Define three entities PCB1, PCB2,
and PCB3 (Figure L4.3).
Define the arrival process (Figure L4.4). Assume all 1,500 boards are in stock
when the assembly operations begin. The process and routing tables are devel-
oped as in Figure L4.5.
har01307_LB04_419-464.indd 420 20/12/10 1:53 PM

FIGURE L4.3
Three types of circuit boards.
FIGURE L4.4
Arrival process for all circuit boards.
FIGURE L4.5
Processes and routings for Pomona Electronics.
Run the simulation model. Note that the whole batch of 1,500 printed circuit
boards (500 of each) takes a total of 2.46 hours to be processed.
L4.2 Multiple Parallel Identical Locations

Location: Capacity versus Units
The differences between the capacity and the units of a location were briefly dis-
cussed in Lab 2, section L2.4. Each location in a model has associated with it
some finite capacity. The capacity of a location refers to the maximum number of
har01307_LB04_419-464.indd 421 20/12/10 1:53 PM

422 Part II Labs
entities that the location can hold at any time. The reserved word INF or INFINITE
sets the capacity to the maximum allowable value. The default capacity for a loca-
tion is one unless a counter, conveyor, or queue is the first graphic assigned to it.
The default capacity in that case is INFINITE.
The number of units of a location refers to the number of identical, inter-
changeable, and parallel locations referenced as a single location for routing and
processing. A multiunit location eliminates the need to create multiple locations
and multiple processes for locations that do the same thing. While routings to and
from each unit are the same, each unit may have unique downtimes.
A multiunit location is created by entering a number greater than one as the
number of units for a normal location. A corresponding number of locations will
appear below the original location with a numeric extension appended to each
copied location name designating the unit number of that location. The original
location record becomes the prototype for the unit records. Each unit will inherit the
prototype’s characteristics unless the individual unit’s characteristics are changed.
Downtimes, graphic symbols, and in the case of multicapacity locations, capacities
may be assigned to each unit. As the number of units of a location is changed, the
individual unit locations are automatically created or destroyed automatically.
The following three situations describe the advantages and disadvantages of
modeling multicapacity versus multiunit locations.
• Multicapacity locations. Locations are modeled as a single unit with
multicapacity (Figure L4.6). All elements of the location perform
identical operations. When one element of the location is unavailable due
to downtime, all elements are unavailable. Only clock-based downtimes
are allowed.
• Multiunit locations. Locations are modeled as single-capacity but
multiple units (Figure L4.7). These are locations consisting of two or
FIGURE L4.6
A single unit location of multicapacity.
FIGURE L4.7
A multiunit location with single-capacity.
har01307_LB04_419-464.indd 422 20/12/10 1:53 PM

FIGURE L4.8
Multiple single-capacity locations.
more parallel and interchangeable processing units. Each unit shares the
same sources of input and the same destinations for output, but each may
have independent operating characteristics. This method provides more
flexibility in the assignment of downtimes and in selecting an individual
unit to process a particular entity.
• Multiple, single-capacity locations. Locations are modeled as individual
and single capacity (Figure L4.8). Usually noninterchangeable locations
are modeled as such. By modeling as individual locations, we gain the
flexibility of modeling separate downtimes for each element. In addition,
we now have complete flexibility to determine which element will process
a particular entity.
Problem Statement
At San Dimas Electronics, jobs arrive at three identical inspection machines ac-
cording to an exponential distribution with a mean interarrival time of 12 minutes.
The first available machine is selected. Processing on any of the parallel machines
is normally distributed with a mean of 10 minutes and a standard deviation of
3 minutes. Upon completion, all jobs are sent to a fourth machine, where they
queue up for date stamping and packing for shipment; this takes five minutes nor-
mally distributed with a standard deviation of two minutes. Completed jobs then
leave the system. Run the simulation for one month (20 days, eight hours each).
Calculate the average utilization of the four machines. Also, how many jobs are
processed by each of the four machines?
Define a location called Inspect. Change its units to 3. Three identical parallel
locations—that is, Inspect.1, Inspect.2, and Inspect.3—are thus created. Also, de-
fine a location for all the raw material to arrive (Material_Receiving). Change the
capacity of this location to infinite. Define a location for Packing (Figure L4.9).
Select Background Graphics from the Build menu. Make up a label “San Dimas
Electronics.” Add a rectangular border. Change the font and color appropriately.
Define an entity called PCB. Define the frequency of arrival of the entity PCB
as exponential with a mean interarrival time of 12 minutes (Figure L4.10). Define
the process and routing at San Dimas Electronics as shown in Figure L4.11.
In the Simulation menu select Options. Enter 160 in the Run Hours box. Run
the simulation model. The average utilization and the number of jobs processed at
the four locations are given in Table L4.2.
har01307_LB04_419-464.indd 423 20/12/10 1:53 PM

424 Part II Labs
FIGURE L4.9
The locations and the layout of San Dimas Electronics.
FIGURE L4.10
Arrivals of PCB at San Dimas Electronics.
FIGURE L4.11
Process and routing tables at San Dimas Electronics.
har01307_LB04_419-464.indd 424 20/12/10 1:53 PM

TABLE L4.2 Summary of Results of the Simulation Run
Average Utilization Number of Jobs Processed
Inspector 1 49.5 437

Packing 41.5 798
L4.3 Resources
A resource is a person, piece of equipment (viz. lift truck, pallet truck, etc.)
or some other discrete device that is used for one or more of the following
functions: transporting entities, assisting in performing a process on entities at
locations, performing maintenance on locations, or performing maintenance on
other resources. Resources consist of one or more units with common charac-
teristics, such as a pool of service technicians or a fleet of lift trucks. Resources
may be dynamic, meaning that they move along a path network, or static, in
which no movement occurs. Resources may also have downtimes. The follow-
ing is an example of the use of a static resource. We discuss dynamic resources
in Lab 12, section L12.2. The downtimes of resources are discussed in Lab 6,
section L6.4.
Static resources are resources that do not visibly move. A static resource may
be needed to perform an operation at only one location, such as an inspection
operator, and will appear during the entire simulation in the same place it was
defined graphically. Although a static resource is stationary and does not visibly
move between locations, it may be used at more than one location or to move
entities between locations. In the following example, an operator is used as an
example of a static resource.
At Ghosh’s Gear Shop, a small automotive OEM facility, there are two
NC lathes, a degreaser, and an inspection equipment. Blanks are received at the
receiving area. The blanks are machined either on NC Lathe 1 or NC Lathe 2
(machines are assigned by turn for successive blanks). After machining, the cogs
go to the degreaser for washing. Then the cogs are inspected. Ghosh (a static
resource) moves the parts between machines. The time to move parts between
machines is normally distributed (mean of 3 minutes and a standard deviation
of 1 minute). All the machines run on automated cycles. Figure L4.12 shows
the resource table. The process and routing tables for this simulation model are
shown in Figure L4.13. Figure L4.14 shows a snapshot of the simulation model
of Ghosh’s Gear Shop. Please note that there are no status lights for a static
resource. However, the resource graphic changes color to green when in use and
red when down.
har01307_LB04_419-464.indd 425 20/12/10 1:53 PM

426 Part II Labs
FIGURE L4.12
Ghosh defined as a static resource.
FIGURE L4.13
Process and routing tables for Ghosh’s Gear Shop.
FIGURE L4.14
A snapshot of the
simulation model for
Ghosh’s Gear Shop.
L4.4 Routing Rules

Routing rules are used in selecting among alternative downstream locations
for routing an entity. Sometimes customers prefer one downstream location to
another. The selection among several alternative locations may follow one of
these rules.
• FIRST available: Select the first location that has available capacity.
• MOST available: Select the location that has the most available capacity.
• By TURN: Rotate the selection among two or more available locations.
• RANDOM: Select randomly among two or more available locations.
har01307_LB04_419-464.indd 426 20/12/10 1:53 PM

• If JOIN request: Select the location that satisfies a JOIN request (see the
JOIN statement).
• If LOAD request: Select the location that satisfies a LOAD request (see the
LOAD statement).
• If SEND request: Select the location that satisfies a SEND request (see the
SEND statement).
• Longest: Select the location that has been unoccupied the longest.
• Unoccupied until FULL: Continue to select the same location until it is full.
• If EMPTY: Select the location only when empty and continue to select until
it is full.
• Probabilistic: Select based on the probability specified (such as 0.75).
• User condition: Select the location that satisfies the Boolean condition
specified by the user (such as AT2>5). Conditions may include any
numeric expression except for location attributes, resource-specific
functions, and downtime-specific functions.
• CONTINUE: Continue at the same location for additional operations.
CONTINUE is allowed only for blocks with a single routing.
• As ALTERNATE to: Select as an alternate if available and if none of the
above rules can be satisfied.
• As BACKUP: Select as a backup if the location of the first preference is
down.
• DEPENDENT: Select only if the immediately preceding routing was
processed.
If only one routing is defined for a routing block, use the first available, join, load,
send, if empty, or continue rule. The most available, by turn, random, longest
unoccupied, until full, probabilistic, and user condition rules are generally only
useful when a block has multiple routings.
Problem Statement
Amar, Akbar, and Anthony are three tellers in the local branch of Bank of India.
Figure L4.15 shows the layout of the bank. Assume that customers arrive at the bank
according to a uniform distribution (mean of five minutes and half-width of four min-
utes). All the tellers service the customers according to another uniform distribution
(mean of 10 minutes and half-width of 6 minutes). However, the customers prefer
Amar to Akbar, and Akbar over Anthony. If the teller of choice is busy, the custom-
ers choose the first available teller. Simulate the system for 200 customer service
completions. Estimate the teller’s utilization (percentage of time busy).
The locations are defined as Akbar, Anthony, Amar, Teller_Q, and Enter as
shown in Figure L4.16. The Teller_Q is exactly 100 feet long. Note that we have
checked the queue option in the Conveyor/Queue dialog box (Figure L4.17). The
customer arrival process is shown in Figure L4.18. The processes and routings
are shown in Figure L4.19. Note that the customers go to the tellers Amar, Akbar,
and Anthony in the order they are specified in the routing table.
har01307_LB04_419-464.indd 427 20/12/10 1:53 PM

428 Part II Labs
FIGURE L4.15
Layout of the Bank of
India.
FIGURE L4.16
Locations at the Bank of India.
FIGURE L4.17
Conveyor/Queue
dialog box.
har01307_LB04_419-464.indd 428 20/12/10 1:53 PM

FIGURE L4.18
Customer arrival at the Bank of India.
FIGURE L4.19
Process and routing tables at the Bank of India.
TABLE L4.3 Utilization of Tellers at the Bank of India
% Utilization
Selection in Order of
Tellers Preference Selection by Turn
Amar 79 63.9
Akbar 64.7 65.1
Anthony 46.9 61.5
The results of the simulation model are shown in Table L4.3. Note that Amar
being the favorite teller is busier than Akbar and Anthony.
If the customers were routed to the three tellers in turn (selected in rotation)
the process and routing tables would be as in Figure L4.17. Note that By Turn was
selected from the Rule menu in the routing table. These results of the simulation
model are also shown in Table L4.3. Note that Amar, Akbar, and Anthony are now
utilized almost equally.
L4.5 Variables
A variable is something that is capable of changing or varying. Variables are
placeholders for either real or integer numbers that may change during simulation.
Variables are typically used for making decisions or for gathering data. Variables
har01307_LB04_419-464.indd 429 20/12/10 1:53 PM

430 Part II Labs
can be defined to track statistics and monitor other activities during a simulation
run. This is useful when the built-in statistics don’t capture a particular perfor-
mance metric of interest. Variables might be defined to track:
• The number of customers waiting in multiple queues.
• Customer waiting time during a specific time period.
• Customer time in the store or bank or office.
• Work-in-process inventory.
• Production quantity.
In ProModel two types of variables are used—local variables and global variables.
• Global variables are accessible from anywhere in the model and at any
time. Global variables are defined through the Variables (global) editor
in the Build menu. The value of a global variable may be displayed
dynamically during the simulation. It can also be changed interactively.
Global variables can be referenced anywhere a numeric expression is
valid.
• Local variables are temporary variables that are used for quick convenience
when a variable is needed only within a particular operation (in the Process
table), move logic (in the Routing table), logic (in the Arrivals, Resources,
or Subroutine tables), the initialization or termination logic (in the General
Information dialog box), and so forth. Local variables are available only
within the logic in which they are declared and are not defined in the
Variables edit table. They are created for each entity, downtime occurrence,
or the like executing a particular section of logic. A new local variable is
created for each entity that encounters an INT or REAL statement. It exists
only while the entity processes the logic that declared the local variable.
Local variables may be passed to subroutines as parameters and are
available to macros.
A local variable must be declared before it is used. To declare a local variable,
use the following syntax:
INT or REAL <name1> {= expression}, <name2> {= expression}
Examples:
INT HourOfDay, WIP
REAL const1 ⫽ 2.5, const2 ⫽ 5.0
INT Init_Inventory ⫽ 170
In section L6.1.2 we show the use of a local variable in the simulation model
logic.
Global Variable—Tracking Work in Process Inventory and Production

In the SoCal Casting Inc, machine shop, raw castings arrive in batches of four
every hour. From the raw material store they are sent to the mill, where their
processing time takes an average of three minutes, with a standard deviation of
har01307_LB04_419-464.indd 430 20/12/10 1:53 PM

FIGURE L4.20
Layout of SoCal
Casting Inc.
FIGURE L4.21
Locations at SoCal Casting Inc.
FIGURE L4.22
Arrival of castings at SoCal Casting Inc.
one minute (normally distributed). The milled castings go to the grinder, where
they are ground for a duration that is uniformly distributed (four to six minutes)
or U(5,1). After grinding, the ground pieces go to the finished parts store. Run the
simulation for 100 hours. Track the work-in-process inventory and the production
quantity.
The complete simulation model layout is shown in Figure L4.20. The loca-
tions are defined as Receiving_Dock, Mill, Grinder, and Finish_Parts_Store (Fig-
ure L4.21). Castings (entity) are defined to arrive in batches of four (Qty each)
every 60 minutes (Frequency) as shown in Figure L4.22. The processing and rout-
ing tables are shown in Figure L4.23.
Define a variable in your model to track the work-in-process inventory (WIP)
of parts in the machine shop. Also, define another variable to track the production
(PROD_QTY) of finished parts (Figure L4.24). Note that both of these are integer
type variables.
har01307_LB04_419-464.indd 431 20/12/10 1:53 PM

432 Part II Labs
FIGURE L4.23
Processing and routing tables for the SoCal Casting Inc.
FIGURE L4.24
Variables for the SoCal Casting Inc. model.
In the processing table, add the following operation statement in the Receiv-
ing location (Figure L4.23).
WIP ⫽ WIP ⫹ 1
Add the following operation statements in the outgoing Finish_Parts_Store loca-

tion (Figure L4.23):
WIP ⫽ WIP ⫺ 1
PROD_QTY ⫽ PROD_QTY ⫹ 1
Alternatively, these three statements could be written as INC WIP, DEC WIP, and
INC PROD_QTY.
L4.6 Uncertainty in Routing—Track Defects and Rework

Sometimes an entity needs to be routed to a destination from a list of avail-
able locations on a probabilistic basis. The location to which a specific entity
is routed remains uncertain. A certain percentage of entities will be routed
to a specific destination. The total of these percentages, however, should be
100 percent.
After processing is completed, a workpiece is frequently inspected and
checked for correctness. Good workpieces move on to the next operation. Defec-
tive ones are either scrapped or reworked. For rework, the workpiece is sent to
one of the earlier operations. The inspection process is carried out either on all
workpieces or on a sample of workpieces.
har01307_LB04_419-464.indd 432 20/12/10 1:53 PM

Problem Statement
SoCal Casting Inc. in section L4.5 decides to add an inspection process after the
grinding operation. After inspection, 30 percent of the parts are sent back to the
mill and 10 percent are sent back to the grinder for rework. 5 percent parts are
scrapped. The balance (55 percent) passes inspection and goes on to the finished
parts store. The inspection takes a time that has a triangular distribution (min. 4,
mode 5, max. 6 minutes). The process times for rework are the same as those
for new jobs. Track the amount of rework at the mill and the grinder. Also, track
the amount of scrapped parts and finished production. Run the simulation for
100 hours.
The locations are defined as mill, grinder, inspection, finish parts store, receiv-
ing dock, scrap parts, mill rework, and grind rework as shown in Figure L4.25.
The last four locations are defined with infinite capacities. The arrivals of
castings are defined in batches of four every hour. Next we define five variables
(Figure L4.26) to track work-in-process, production, mill rework, grind rework,
and scrap quantities. The processes and routings are defined as in Figure L4.27.
Note the use of the probabilistic routing rule for Casting at the Inspector loca-
tion. The probabilistic routing rule selects a location listed in a block of routings
based on a probability. Several probability routings should be used together and
the sum of all probabilities must equal one. The entity will wait to be routed until
the selected location has available capacity.
FIGURE L4.25
Simulation model
layout for SoCal
Casting Inc. with
inspection.
FIGURE L4.26
Variables for the SoCal Casting Inc. with inspection model.
har01307_LB04_419-464.indd 433 20/12/10 1:53 PM

434 Part II Labs
FIGURE L4.27
Processing and routing tables for the SoCal Casting Inc. model with inspection.
L4.7 Batching Multiple Entities of Similar Type

L4.7.1 Temporary Batching
Frequently we encounter a situation where work parts (or customers) are batched
and processed together. After processing is over, the parts are unbatched again.
Following are some examples:
a. An autoclave/oven baking a batch of parts, cookies, and so on.
b. A batch of visitors (Disneyland) going on a ride.
c. A group of students attending a class.
After baking/riding/attending class the entities are separated and they go their
individual ways. The individual entities retain their properties during the batching
and processing activities.
For such temporary batching and subsequent unbatching, we use the GROUP
and UNGROUP statements. One may group entities by individual entity type by
defining a process record for the type to group, or group them irrespective of
entity type by defining an ALL process record. ProModel maintains all of the char-
acteristics and properties of the individual entities of the grouped entities and
allows them to remain with the individual entities after an UNGROUP command.
Note that the capacity of the location where Grouping occurs must be at least as
large as the group size. The location where entities are Ungrouped need only have
a capacity of one.
Problem Statement
El Segundo Composites receive orders for aerospace parts that go through cut-
ting, lay-up, and bonding operations. The bonding operation is done in an au-
toclave in batches of five parts. The processing times for these operations are
har01307_LB04_419-464.indd 434 20/12/10 1:53 PM

uniform(20,5), uniform(30,10) and uniform(100,10) minutes. After bonding,

the parts go to the shipment clerk individually. The shipment clerk takes normal
(20,5) minutes to pack and ship the parts. The orders are received on average once
every hour, exponentially distributed. It takes an average of 15 minutes to trans-
port these parts from one machine to another. Figure out the amount of WIP in the
shop. Simulate for six months or 1,000 working hours.
The locations are defined as Ply_Cutting, LayUp, Oven, Order_Q, Ship_
Clerk, and Ship_Q. The layout of El Segundo Composites is shown in Fig-
ure L4.28. The processing and routing tables are shown in Figure L4.29 A variable
WIP is defined to keep track of the amount of work in progress in the shop. The
WIP value history over the simulated 1,000 hours is shown in Figure L4.30. Note
that the average WIP in the shop has been around 7, although on rare occasions
the WIP has been as much as 23.
Note that we need to change the name of the entity when a group is created.
This new name represents the entire group of entities. When we ungroup, we must
define a new process record at the same location for whatever entity names may
have been ungrouped.
FIGURE L4.28
Layout of El Segundo
Composites.
FIGURE L4.29
Processing and routing tables for El Segundo Composites.
har01307_LB04_419-464.indd 435 20/12/10 1:53 PM

436 Part II Labs
FIGURE L4.30
Work in process value history.
L4.7.2 Permanent Batching

In some situations, multiple entities of the same type or different types are batched
together permanently. After batching, a single batch entity is formed, alternatively
with a different name. In such permanent batching, the characteristics of the indi-
vidual entities are lost; that is, the individual entities cannot be ungrouped later.
To define a permanent batching of entities, use the COMBINE statement. The
capacity of the location where you use the COMBINE statement should be at least as
large as the combined quantity.
Problem Statement
At the Garden Reach plant of the Darjeeling Tea Company, the filling machine
fills empty cans with 50 bags of the best Darjeeling tea at the rate of one can every
1⫾0.5 seconds uniformly distributed. The tea bags arrive to the packing line with
a mean interarrival time of one second exponentially distributed. The filled cans
go to a packing machine where 20 cans are combined into one box. The packing
operation takes uniform (20⫾10) seconds. The boxes are shipped to the dealers.
This facility runs 24 hours a day. Simulate for one day.
The various locations at the Darjeeling Tea Company plant are shown in Fig-
ure L4.31. Three entities (Teabag, Can, and Box) are defined next. The processing
and routing tables are shown in Figure L4.32. The layout of the Darjeeling Tea
Company plant is shown in Figure L4.33.
har01307_LB04_419-464.indd 436 20/12/10 1:53 PM

FIGURE L4.31
Locations at the Darjeeling Tea Company.
FIGURE L4.32
Processing and routing tables at the Darjeeling Tea Company.
FIGURE L4.33
Layout of the
Darjeeling Tea
Company.
har01307_LB04_419-464.indd 437 20/12/10 1:53 PM

438 Part II Labs
L4.8 Attaching One or More Entities to Another Entity

L4.8.1 Permanent Attachment
Sometimes one or more entities are attached permanently to another entity, as
in an assembly operation. The assembly process is a permanent bonding; the as-
sembled entities lose their separate identities and properties. The individual enti-
ties that are attached cannot be separated again.
The JOIN statement is used to permanently assemble two or more individual
entities together. For every JOIN statement, there must be at least one correspond-
ing If Join Request rule. Joining is a two-step process:
1. Use the JOIN statement at the designated assembly location.
2. Use the join routing rule for all joining entities.
One of the joining entities, designated as the “base” entity, issues the JOIN state-
ment. All other joining entities must travel to the assembly location on an If Join
Request routing rule. To join multiple entity types to the base entity, enter a JOIN
statement for each entity type to be joined along with a corresponding If Join
Request routing.
Problem Statement
At Shipping Boxes Unlimited computer monitors arrive at Monitor_Q at the rate
of one every 15 minutes (exponential) and are moved to the packing table. Boxes
arrive at Box_Q, at an average rate of one every 15 minutes (exponential) and are
also moved to the packing table. At the packing table, monitors are packed into
boxes. The packing operation takes normal (5,1) minutes. Packed boxes are sent
to the inspector (Inspect_Q). The inspector checks the contents of the box and tal-
lies with the packing slip. Inspection takes normal (4,2) minutes. After inspection,
the boxes are loaded into trucks at the shipping dock (Shipping_Q). The loading
takes uniform (5,1) minutes. Simulate for 100 hours. Track the number of moni-
tors shipped and the WIP of monitors in the system.
The locations at Shipping Boxes Unlimited are defined as Monitor_Q,
Box_Q, Shipping_Q, Inspect_Q, Shipping_Dock, Packing_Table, and Inspec-
tor. All the queue locations are defined with a capacity of infinity. Three entities
(Monitor, Empty_Box, and Full_Box) are defined next. The arrivals of monitors
and emptyboxes are shown in Figure L4.34. The processing and routing tables
are shown in Figure L4.35. A snapshot of the simulation model is captured in
FIGURE L4.34
Arrival of monitors and empty boxes at Shipping Boxes Unlimited.
har01307_LB04_419-464.indd 438 20/12/10 1:53 PM

FIGURE L4.35
Processing and routing tables for Shipping Boxes Unlimited.
FIGURE L4.36
A snapshot of the
the Shipping Boxes
Unlimited.
Figure L4.36. The plot of the work-in-process inventory for the 100 hours of
simulation run is shown in Figure L4.37. Note that the work-in-process inventory
rises to as much as 12 in the beginning. However, after achieving steady state, the
WIP inventory stays mostly within the range of 0 to 3.
L4.8.2 Temporary Attachment

Sometimes one or more entities are attached to another entity temporarily, pro-
cessed or moved together, and detached later in the system. In ProModel, we
use the LOAD and UNLOAD statements to achieve that goal. The properties of the
individual attached entities are preserved.
Problem Statement
For the Shipping Boxes Unlimited problem in section L4.8.1, assume the inspec-
tor places (loads) a packed box on an empty pallet. The loading time is uniformly
distributed: U(3,1) minutes. The loaded pallet is sent to the shipping dock and
har01307_LB04_419-464.indd 439 20/12/10 1:53 PM

440 Part II Labs
FIGURE L4.37
Time-weighted plot of the WIP inventory at the Shipping Boxes Unlimited.
FIGURE L4.38
Locations at the Shipping Boxes Unlimited.
waits in the shipping queue. At the shipping dock, the packed boxes are unloaded
from the pallet. The unloading time is also uniformly distributed: U(3,1) minutes.
The boxes go onto a waiting truck. The empty pallet is returned, via the pallet
queue, to the inspector. One pallet is used and recirculated in the system. Simulate
for 100 hours. Track the number of monitors shipped and the WIP of monitors.
The locations defined in this model are Monitor_Q, Box_Q, Inspect_Q,
Shipping_Q, Pallet_Q, Packing_Table, Inspector, Shipping_Dock (Figure L4.38).
All queue locations have infinite capacity. The layout of Shipping Boxes Unlim-
ited and a snapshot of the simulation model run are shown in Figure L4.39. Five
entities are defined: Monitor, Box, Empty_Box, Empty_Pallet, and Full_Pallet
har01307_LB04_419-464.indd 440 20/12/10 1:53 PM

FIGURE L4.39
A snapshot of the
simulation model at
the Shipping Boxes
Unlimited.
FIGURE L4.40
Entities at Shipping Boxes Unlimited.
FIGURE L4.41
Arrival of monitors, empty boxes, and empty pallets at Shipping Boxes Unlimited.
as shown in Figure L4.40. The arrivals of monitors, empty boxes, and empty
pallets are shown in Figure L4.41. The processing and routing tables are shown
in Figure L4.42. Note that comments can be inserted in a line of code as follows
(Figure L4.42):
/* inspection time */
The plot of the work-in-process inventory for the 100 hours of simulation
run is presented in Figure L4.43. Note that after the initial transient period
har01307_LB04_419-464.indd 441 20/12/10 1:53 PM

442 Part II Labs
FIGURE L4.42
Processing and routing tables at Shipping Boxes Unlimited.
FIGURE L4.43
Time-weighted plot of the WIP inventory at Shipping Boxes Unlimited.
(艑20 hours), the work-in-process inventory drops and stays mostly in a range
of 0~2.
Note that like the UNGROUP statement, it is necessary to define an additional
process record after unloading for those entities that may have gotten unloaded.
har01307_LB04_419-464.indd 442 20/12/10 1:53 PM

L4.9 Accumulation of Entities

Sometimes there is a need to hold similar or different entities at a location until
a certain number accumulate. Once this critical limit is reached, the entities are
released for further processing downstream. No actual batching takes place, but
entities merely accumulate until a specified count has been reached. The capacity
of the location at which the accumulation takes place must be at least as large as
the accumulation limit or the amount of accumulation.
Problem Statement
Visitors arrive at California Adventure Park in groups that vary in size from two
to four (uniformly distributed). The average time between arrivals of two succes-
sive groups is five minutes, exponentially distributed. All visitors wait in front of
the gate until five visitors have accumulated. At that point the gate opens and al-
lows the visitors to enter the park. On average, a visitor spends U(20⫾10) minutes
in the park. Simulate for 1,000 hours. Track how many visitors are waiting outside
the gate and how many are in the park.
Three locations (Gate_In, Walk_In_Park, and Gate_Out) are defined in this
model. Visitor is defined as the entity. The processes and the layout of the adven-
ture park are shown in Figures L4.44 and L4.45, respectively.
FIGURE L4.44
Processing and routing tables for California Adventure Park.
FIGURE L4.45
Layout of California
Adventure Park.
har01307_LB04_419-464.indd 443 20/12/10 1:53 PM

444 Part II Labs
L4.10 Splitting of One Entity into Multiple Entities

When a single incoming entity is divided into multiple entities at a location and
processed individually, we can use the SPLIT construct. The SPLIT command
splits the entity into a specified number of entities, and optionally assigns them a
new name. The resulting entities will have the same attribute values as the origi-
nal entity. Use SPLIT to divide a piece of raw material into components, such as a
silicon wafer into silicon chips or six-pack of cola into individual cans.
Problem Statement
The cafeteria at San Dimas High School receives 10 cases of milk from a vendor
each day before the lunch recess. On receipt, the cases are split open and individ-
ual cartons (10 per case) are stored in the refrigerator for distribution to students
during lunchtime. The distribution of milk cartons takes triangular (.1, .15, .2)
minute per student. The time to split open the cases is U(6,1) minutes. Moving
the cases from receiving to the refrigerator area takes five minutes per case, and
moving the cartons from the refrigerator to the distribution area takes 0.2 minutes
per carton. Students wait in the lunch line to pick up one milk carton each. There
are only 100 students at this high school. Students show up for lunch with a mean
interarrival time of 1 minute (exponential). On average, how long does a carton
stay in the cafeteria before being distributed and consumed? What are the maxi-
mum and the minimum times of stay? Simulate for one lunch period.
The layout of the San Dimas High School cafeteria is shown in Figure L4.46.
Three entities—Milk_Case, Milk_Carton, and Student—are defined. Ten milk
cases arrive at the beginning of a lunch period. One hundred students show up for
lunch each day. The arrival of students and milk cases is shown in Figure L4.47.
The processing and routing tables are shown in Figure L4.48.
FIGURE L4.46
Layout of the San
Dimas High School
cafeteria.
FIGURE L4.47
Arrival of milk and students at the San Dimas High School cafeteria.
har01307_LB04_419-464.indd 444 20/12/10 1:53 PM

FIGURE L4.48
Processing and routing tables at the San Dimas High School cafeteria.
Note that, like the UNGROUP and UNLOAD statements, a new process record
needs to be defined when using the SPLIT statement.
L4.11 Decision Statements

In ProModel you can make use of several general control statements for decision
making. ProModel meets the modern standards of structured program design. Al-
though, there is no formal definition, computer scientists agree that a structured
program should have modular design and use only the following three types of
logic statements: sequences, decisions, and loops.
• Sequences: Program statements are executed one after another.
• Decisions: One of two blocks of program code is executed based on a test
for some condition.
• Loops: One or more statements are executed repeatedly as long as a
specified condition is true.
All the control statements available in ProModel are shown in Figure L4.49. In
this section we introduce you to some of these statements with examples.
L4.11.1 IF-THEN-ELSE Statement

An IF block allows a program to decide on a course of action based on whether a
certain condition is true or false. A program block of the form
IF condition THEN
action1
ELSE
action2
causes the program to take action1 if condition is true and action2 if condition is
false. Each action consists of one or more ProModel statements. After an action is
taken, execution continues with the line after the IF block.
har01307_LB04_419-464.indd 445 20/12/10 1:53 PM

446 Part II Labs
FIGURE L4.49
Control statements
available in ProModel.
TABLE L4.4 Process Times at the Bombay Restaurant
Activity Process Time
Order food Normal (5,1) minutes

Make payment Normal (7.2) minutes
Pick up food Normal (10,2) minutes
Problem Statement
The Bombay Restaurant offers only a drive-in facility. Consumers arrive at the
rate of six each hour (exponential interarrival time). They place their orders at
the first window, drive up to the next window for payment, pick up food from
the last window, and then leave. The activity times are given in Table L4.4. The
drive-in facility can accommodate 10 cars at most. However, customers typically
leave and go to the Madras Café across the street if six cars are waiting in line
when they arrive. Simulate for 100 days (8 hours each day). Estimate the number
of customers served each day. Estimate on average how many customers are lost
each day to the competition.
An additional location (Arrive) is added to the model. After the customers
arrive, they check if there are fewer than six cars at the restaurant. If yes, they join
the line and wait; if not, they leave and go across the street to the Madras Café.
An IF-THEN-ELSE statement is added to the logic window in the processing table
(Figure L4.50). A variable (Customer_Lost) is added to the model to keep track
of the number of customers lost to the competition. The layout of the Bombay
har01307_LB04_419-464.indd 446 20/12/10 1:53 PM

FIGURE L4.50
Processing and routing tables at the Bombay Restaurant.
FIGURE L4.51
Layout of the Bombay
Restaurant.
Restaurant is shown in Figure L4.51. Note that Q_1 and Q_2 are each 100 feet
long and Q_3 is 200 feet long.
The total number of customers lost is 501 in 100 days. The number of
customers served in 100 days is 4791. The average cycle time per customer is
36.6 minutes.
L4.11.2 WHILE...DO Loop

The WHILE...DO block repeats a group of statements continuously while a condi-
tion remains true. If the condition is false, the loop is bypassed.
Problem Statement
The inspector in section L4.8.2 is also the supervisor of the shop. As such, she in-
spects only when at least five full boxes are waiting for inspection in the Inspect_Q.
har01307_LB04_419-464.indd 447 20/12/10 1:53 PM

448 Part II Labs
FIGURE L4.52
An example of the WHILE . . . DO logic for Shipping Boxes Unlimited.
FIGURE L4.53
A plot of the contents of the inspection queue.
A WHILE...DO loop is used to check if the queue has five or more boxes waiting
for inspection (Figure L4.52). The loop will be executed every hour. Figure L4.53
shows a time-weighted plot of the contents of the inspection queue. Note how the
queue builds up to 5 (or more) before the inspector starts inspecting the full boxes.
har01307_LB04_419-464.indd 448 20/12/10 1:53 PM

L4.11.3 DO...WHILE Loop

The DO...WHILE block repeats a group of statements continuously while a condi-
tion remains true. This loop will be executed at least once; use it for processes that
will be executed at least once and possibly more.
Problem Statement
For the SoCal Casting Inc. example in section L4.5, we would like to assign an
upper limit to the work-in-process inventory in the shop. A DO...WHILE loop will
be used to check if the level of WIP is equal to or above five; if so, the incoming
castings will wait at the receiving dock (Figure L4.54). The loop will be executed
once every hour to check if the level of WIP has fallen below five. Figure L4.55
shows an hourly average plot of the value of WIP at SoCal Casting Inc. Note that
the level of WIP is kept at or below five.
L4.11.4 DO...UNTIL Statement

The DO...UNTIL block repeats a statement or a block of statements continuously
while a condition remains false. The DO...UNTIL loop will always be executed at
least one time and possibly more than once.
FIGURE L4.54
An example of DO . . . WHILE loop.
FIGURE L4.55
An hourly average plot
of the WIP at SoCal
Casting Inc.
har01307_LB04_419-464.indd 449 20/12/10 1:53 PM

450 Part II Labs
Problem Statement
The inspector in section L4.8.2 is also the supervisor of the shop. As such, she
waits an hour before checking if there are at least five full boxes waiting in the
Inspect_Q. A DO...UNTIL loop is used to wait 60 minutes and then finding out if
the queue has five or more boxes waiting for inspection (Figure L4.56). The loop
will be executed once every hour. Figure L4.57 shows a time-weighted plot of
the contents of the inspection queue. Note how the queue builds up to 5 (or more)
before the inspector starts inspecting the full boxes.
FIGURE L4.56
An example of DO...UNTIL loop.
FIGURE L4.57
An hourly average plot
of the contents of the
inspection queue.
har01307_LB04_419-464.indd 450 20/12/10 1:53 PM

L4.11.5 GOTO Statement

The GOTO block jumps to the statement identified by the designated label. A label
should follow the normal rules for names (Lab 2, section L2.1). A colon follows
a label.
Problem Statement
The Indian Bank has acquired the Bank of India (section L4.4). Anthony has
been given a golden handshake (that is, fired). As a result of new training pro-
vided to the two remaining tellers (Amar and Akbar) there has been a change in
their customer service time. Whenever more than three customers are waiting for
service, the service time is reduced to an average of five minutes and a standard
deviation of two minutes (normally distributed). The layout of the new Indian
Bank is shown in Figure L4.58. A GOTO statement is used to jump to the appropri-
ate wait statement depending on the number of customers waiting for service in
the Teller_Q, as shown in Figure L4.59.
FIGURE L4.58
The layout of the
Indian Bank.
FIGURE L4.59
An example of a GOTO statement.
har01307_LB04_419-464.indd 451 20/12/10 1:53 PM

452 Part II Labs
This model illustrates one of the dangers of using a GOTO statement and why
its use is discouraged in computer science textbooks. In this example, the L1 and
L2 statements ALWAYS get executed. In other words, a GOTO statement doesn’t work
like an IF-THEN-ELSE.
L4.11.6 WAIT UNTIL Statement

One of ProModel’s most powerful modeling features is to detain the processing of
entities until a (global) variable has reached, exceeded or fallen below a particular
value. This is done by using a WAIT UNTIL statement. Where a WAIT statement
delays processing for a specified time, a WAIT UNTIL statement delays processing
until a specified condition is reached. While the processing of the current logic is
delayed, the rest of the model continues to process during the delay. If the expres-
sion is initially evaluated to be false, it is only reevaluated when a location attribute,
variable, or array element in the expression changes. Multiple entities waiting on
the same condition are released one at a time. This allows a released entity to change
a variable value that will prevent the other waiting entities from being released.
Diners arrive at the rate of 30 per hour (exponentially distributed) to Bombay
Dining. The dining room has a seating capacity for 20 diners. The time for dining
is normally distributed (mean of 25 minutes and a standard deviation of 5 min-
utes). After dinner the diners wait in the cashier queue to pay the cashier. There
are two identical cashiers (chosen randomly by the customers). The cashiers take
uniform(3,1) minutes to serve a customer. The time to travel from the lobby to the
dining room and the dining room to the cashier queue is approximately one min-
ute. On arrival the diners wait in the lobby until the total number of customers in
the dining area, cashier queue, and the two cashiers is less than 20. Figure L4.60
shows a snapshot of the simulation model for the Bombay Diner. Figure L4.61
shows the process and routing tables. Note the use of the WAIT UNTIL statement to
FIGURE L4.60
A snapshot of the
Bombay Dining.
har01307_LB04_419-464.indd 452 20/12/10 1:53 PM

FIGURE L4.61
Process and routing logic for Bombay Dining.
FIGURE L4.62
Hourly average plot of the dining room contents.
check if the total number of customers in the dining room, cashier queue, and the
two cashier windows taken together is less than 20 to allow a newly arrived diner
in the dining room. The diner is open 24 hours a day, seven days a week. Simulate
for one week. Figure L4.62 shows a plot of the dining room contents.
L4.12 Periodic System Shutdown

Some service systems periodically stop admitting new customers for a given du-
ration to catch up on the backlog of work and reduce the congestion. Some auto-
mobile service centers, medical clinics, restaurants, amusement parks, and banks
use this strategy. The modulus mathematical operator in ProModel is useful for
simulating such periodic shutdowns.
har01307_LB04_419-464.indd 453 20/12/10 1:53 PM

454 Part II Labs
Problem Statement
The Bank of India in section L4.4 opens for business each day from 9 A.M. until
4 P.M. (7 hours or 420 minutes). Assume that the customers arrive to the bank
according to an exponential distribution (mean of 3 minutes). All the tellers ser-
vice the customers according to a uniform distribution U(10,6) minutes. Custom-
ers are routed to the three tellers in turn (selected in rotation).
Each day at 12 noon after 180 minutes (3 hours) of operation, the front door
of the bank is locked for a lunch recess. Any new customers arriving at the bank
are turned away. Customers already inside the bank continue to get served. The
bank reopens the front door 60 minutes (1 hour) later to new customers. At 4 P.M.
each day (after 420 minutes of bank opening) the front door is closed again. No
new customers are admitted for business. The bank eventually closes when the
last customer in the bank has been served. Simulate the system for 5 days (120
hours). Make a time-series plot of the Teller_Q to show the effect of locking the
front door of the bank.
The logic for locking the front door is shown in Figure L4.63. The simulation
clock time clock(min) is a cumulative counter of minutes elapsed since the start of
the simulation run. The current time of any given day can be determined by modu-
lus dividing the current simulation time, clock(min), by 1,440 minutes (24 hours).
If the remainder of clock(min) divided by 1,440 minutes is between 180 minutes
and 240 minutes, the arriving customers are turned away (disposed). Otherwise,
they are allowed into the bank. Also, after 4 P.M. (420 minutes after bank opening)
each day the front door is closed again to new customers. An IF-THEN-ELSE logic
block as described in section L4.11.1 is used here.
A time-series plot of the contents of the Teller_Q is shown in Figure L4.64.
This plot clearly shows how the Teller_Q (and the whole bank) builds up dur-
ing the day; then, after the front door is locked after 3 hours of operation (180
minutes) into the simulated day, customers remaining in the queue are processed
and the queue length decreases. The queue length picks back up when the bank
reopens the front door after an hour.
The histogram of the same Teller queue (Figure L4.65) shows that ap-
proximately 79 percent of the time there are 2 or less customers waiting. What
FIGURE L4.63
Process and routing logic for the Bank of India.
har01307_LB04_419-464.indd 454 20/12/10 1:53 PM

FIGURE L4.64
Hourly average plot of Teller_Q contents at the Bank of India.
FIGURE L4.65
Histogram of Teller_Q contents at the Bank of India.
is the average time a customer spends in the bank? Would you recommend
that the bank not close the door after 5 hours of operation (customers never
liked this practice anyway)? Will the average customer spend more time in the
bank?
har01307_LB04_419-464.indd 455 20/12/10 1:53 PM

456 Part II Labs
L4.13 Exercises
1. Visitors arrive at Kid’s World entertainment park according to an
exponential interarrival time distribution with mean 2.5 minutes. The
travel time from the entrance to the ticket window is normally distributed
with a mean of 3 minutes and a standard deviation of 0.5 minute. At the
ticket window, visitors wait in a single line until one of four cashiers is
available to serve them. The time for the purchase of tickets is normally
distributed with a mean of five minutes and a standard deviation of one
minute. After purchasing tickets, the visitors go to their respective gates
to enter the park. Create a simulation model of this system. Run the
simulation model for 2,000 hours to determine:
a. The average and maximum lengths of the ticketing queue.
b. The average number of customers completing ticketing per hour.
c. The average and maximum numbers of customers in the ticketing
system.
d. The average utilization of the cashiers.
e. Whether management should add more cashiers?
2. A consultant for Kid’s World recommended that four individual queues
be formed at the ticket window (one for each cashier) instead of one
common queue. Create a simulation model of this system. Run the
simulation model for 2,000 hours to determine:
a. The average and maximum lengths of the ticketing queues.
b. The average number of customers completing ticketing per hour.
c. The average and maximum numbers of customers in the ticketing
system.
d. The average utilization of the cashiers.
e. Whether you agree with the consultant’s decision. Would you
recommend a raise for the consultant or fire him/her?
3. At the Kid’s World entertainment park in Exercise 1, the operating
hours are 8 A.M. until 10 P.M. each day, seven days of the week. The
ticketing window closes at 9 P.M. each day. Any customers arriving to the
ticketing window after 9 P.M. are turned away. Simulate for a whole year
and answer questions a–e as given in Exercise 1.
4. At Southern California Airline’s traveler check-in facility, three types
of passengers arrive: passengers with e-tickets (Type E), passengers
with paper tickets (Type T), and passengers that need to purchase tickets
(Type P). The interarrival distribution and the service times for these
passengers are given in Table L4.5. Create a simulation model of this
system. Run the simulation model for 2,000 hours. If separate gate
agents serve each type of passenger, determine the following:
a. The average and maximum length of the three queues.
b. The average number of passengers of each type completing check-in
procedures per hour.
har01307_LB04_419-464.indd 456 20/12/10 1:53 PM

TABLE L4.5 Interarrival and Service Time Distributions at Southern

California Airline
Types of Passenger Interarrival Distribution Service Time Distribution
Type E Exponential (mean 5.5 min) Normal (mean 3 min., std.dev. 1 min.)
Type T Exponential (mean 10.5 min) Normal (mean 8 min., std.dev. 3 min.)
Type P Exponential (mean 15.5 min) Normal (mean 12 min., std.dev. 3 min.)
c. The average utilization of the gate agents.

d. The percentage of time the number of passengers (of each type)
waiting in line is ⱕ 2.
e. Would you recommend one single line for check-in for all three types
of passengers? Discuss the pros and cons of such a change.
5. Raja & Rani, a fancy restaurant in Santa Monica, holds a maximum of
15 diners. Customers arrive according to an exponential distribution with
a mean of 5 minutes. If a customer arrives when the restaurant is full,
they wait in the outside lobby. Customers stay in the restaurant according
to a triangular distribution with a minimum of 45 minutes, a maximum
of 60 minutes, and a mode of 50 minutes. Create a simulation model
of this system. The restaurant operating hours are 3 P.M. until 12 P.M.
Simulate for a whole year and respond to the following questions:
a. What is the average and maximum waiting time for an arriving
customer?
b. What is the average and maximum time a customer spends in the
restaurant (waiting and service)?
c. What is the average and maximum numbers of customers waiting in
the outside lobby to enter the restaurant?
d. Beginning empty, how long does it take for the restaurant to fill?
e. What is the total number of diners entering the restaurant before it
fills?
f. What is the total number of guests served per night?
g. What is the utilization of the restaurant?
6. Woodland Nursing Home has six departments—initial exam, X-ray,
operating room, cast-fitting room, recovery room, and checkout room.
The probabilities that a patient will go from one department to another
are given in Table L4.6. The time patients spend in each department
is given in Table L4.7. Patients arrive at the average rate of 5 per hour
(exponential interarrival time). The nursing home remains open 24/7.
However, every seven hours (420 minutes) the front door is
locked for an hour (60 minutes). No new patients are allowed in the
nursing home during this time. Patients already in the system continue
har01307_LB04_419-464.indd 457 20/12/10 1:53 PM

458 Part II Labs
TABLE L4.6 Probabilities of Patient Flow
From To Probability
Initial exam X-ray 0.35

Operating room 0.20
Recovery room 0.15
Checkout room 0.30
X-ray Operating room 0.15
Cast-fitting room 0.25
Recovery room 0.40
Checkout room 0.20
Operating room Cast-fitting room 0.30
Recovery room 0.65
Checkout room 0.05
Cast-fitting room Recovery room 0.55
X-ray 0.10
Checkout room 0.35
Recovery room Operating room 0.10
X-ray 0.20
Checkout room 0.70
TABLE L4.7 Patient Processing Time
Department Patient time in department
Initial exam Normal(5,2) minutes

X-ray Uniform(6 ± 1) minutes
Operating room Normal(30,5) minutes
Cast-fitting room Uniform(10 ± 4) minutes
Recovery room Normal(8,2) minutes
to get served. Simulate for a whole year and respond to the following
questions:
a. Figure out the utilization of each department.
b. What are the average and maximum numbers of patients in each
department?
c. Which is the bottleneck department?
d. What are the average and maximum times spent by a patient in the
nursing home?
7. United Electronics manufactures small custom electronic assemblies.
Parts must be processed through four stations: assembly, soldering,
painting, and inspection. Orders arrive with an exponential interarrival
distribution (mean 20 minutes). The process time distributions are shown
in Table L4.8.
The soldering operation can be performed three jobs at a time.
Painting can be done on four jobs at a time. Assembly and inspection
har01307_LB04_419-464.indd 458 20/12/10 1:53 PM

TABLE L4.8 Process Time Distributions at United Electronics
Assembly Uniform (13.5 ± 1.5) minutes

Soldering Normal (36,10) minutes
Painting Uniform (55 ± 15) minutes
Inspection Exponential (8) minutes
TABLE L4.9 Process times and probability of defects at Bengal Toy

Company
Assembly Time Inspect Time Correct Time
Standard Standard Standard

Center Mean Deviation Mean Deviation P(error) Mean Deviation
1 0.7 0.2 0.2 0.05 0.1 0.2 0.05

2 0.75 0.25 0.2 0.05 0.05 0.15 0.04
3 0.8 0.15 0.15 0.03 0.03 0.1 0.02
are performed on one job at a time. Create a simulation model of this

system. Simulate this manufacturing system for 2,000 hours. Collect
and report statistics on the utilization of each station, associated queues,
and the total number of jobs manufactured during each eight-hour shift
(average and maximum).
8. In United Electronics in Exercise 7, 10 percent of all finished
assemblies are sent back to soldering for rework after inspection, five
percent are sent back to assembly for rework after inspection, and
one percent of all assemblies fails to pass and is scrapped. Create a
simulation model of this system. Simulate this manufacturing system
for 2,000 hours. Collect and report statistics on the utilization of each
station, associated queues, total number of jobs assembled, number of
assemblies sent for rework to assembly and soldering, and the number
of assemblies scrapped during each eight-hour shift (average and
maximum).
9. Small toys are assembled in four stages (Centers 1, 2, and 3 and
Inspection) at Bengal Toy Company. After each assembly step, the
appliance is inspected or tested; if a defect is found, it must be corrected
and then checked again. The assemblies arrive at a constant rate of one
assembly every two minutes. The times to assemble, test, and correct
defects are normally distributed. The mean and standard deviations
of the times to assemble, inspect, and correct defects, as well as the
likelihood of an assembly error, are shown in Table L4.9. If an assembly
is found defective, the defect is corrected and it is inspected again. After
a defect is corrected, the likelihood of another defect being found is
har01307_LB04_419-464.indd 459 20/12/10 1:53 PM

460 Part II Labs
TABLE L4.10 Data Collected at Salt Lake City Electronics
Number Number
Job of of jobs Time between
Type batches per batch Assembly Time Soldering Time Painting Time Inspection Time Batch Arrivals
1 15 5 Triangular(5,7,10) Normal(36,10) Uniform(55±15) Exponential(8) Exponential(14)

2 25 3 Triangular(7,10,15) Uniform(35±5) Exponential(5) Exponential(10)
Note: All times are given in minutes.
FIGURE L4.66 Truck travel

Schematic of dump and unload
truck operation for
DumpOnMe.
Loader 1
Weighing
Loader queue Weighing queue scale
Loader 2
the same as during the first inspection. We assume in this model that an
assembly defect is eventually corrected and then it is passed on to the
next station. Simulate for 2,000 hours. Determine the number of good
appliances shipped in a year.
10. Salt Lake City Electronics manufactures small custom communication
equipment. Two differentjob types are to be processed within the
following manufacturing cell. The necessary data are given in
Table L4.10. Simulate the system for 200 days, eight hours each day, to
determine the average number of jobs waiting for different operations,
number of jobs of each type finished each day, average cycle time for
each type of job, and all jobs taken together.
11. Six dump trucks at the DumpOnMe facility in Riverside are used to
haul coal from the entrance of a small mine to the railroad. Figure L4.66
provides a schematic of the dump truck operation. Each truck is loaded
by one of two loaders. After loading, a truck immediately moves to
the scale to be weighed as soon as possible. Both the loaders and the
scale have a first-come, first-served waiting line (or queue) for trucks.
Travel time from a loader to the scale is considered negligible. After
being weighed, a truck begins travel time (during which time the
truck unloads), and then afterward returns to the loader queue. The
distributions of loading time, weighing time, and travel time are shown
in Table L4.11.
a. Create a simulation model of this system. Simulate for 200 days,
eight hours each day.
har01307_LB04_419-464.indd 460 20/12/10 1:53 PM

TABLE L4.11 Various Time Teasurement Data at the DumpOnMe Facility
Loading Time Uniform(7.5±2.5) min

Weighing Time Uniform(3.5±1.5) min
Travel Time Triangular(10,12,15) min
b. Collect statistics to estimate the loader and scale utilization

(percentage of time busy).
c. About how many trucks are loaded each day on average?
12. At the Pilot Pen Company, a molding machine produces pen barrels of
two different colors—red and blue—in the ratio of 3:2. The molding time
is triangular (3,4,6) minutes per barrel. The barrels go to a filling machine,
where ink of appropriate color is filled at the rate of 20 pens per hour
(exponentially distributed). Another molding machine makes caps of the
same two colors in the ratio of 3:2. The molding time is triangular (3,4,6)
minutes per cap. At the next station, caps and filled barrels of matching
colors are joined together. The joining time is exponentially distributed
with a mean of 1 minute. Simulate for 2,000 hours. Find the average
number of pens produced per hour. Collect statistics on the utilization of
the molding machines and the joining equipment.
13. Customers arrive at the NoWaitBurger hamburger stand with an
interarrival time that is exponentially distributed with a mean of one
minute. Out of 10 customers, five buy a hamburger and a drink, three
buy a hamburger, and two buy just a drink. One server handles the
hamburger while another handles the drink. A person buying both
items needs to wait in line for both servers. The time it takes to serve a
customer is N(70,10) seconds for each item. Simulate for 2,000 hours.
Collect statistics on the number of customers served each hour, size of
the two queues, and utilization of the servers. What changes (if any)
would you suggest to make the system more efficient?
14. Workers who work at the Detroit ToolNDie plant must check out tools
from a tool crib. Workers arrive according to an exponential distribution
with a mean time between arrivals of five minutes. At present, three tool
crib clerks staff the tool crib. The time to serve a worker is normally
distributed with a mean of 10 minutes and a standard deviation of two
minutes. Compare the following servicing methods. Simulate for a 2,000
hours and collect data.
a. Arriving workers form a single queue, choosing the next available
tool crib clerk.
b. Arriving workers enter the shortest queue (each clerk has his/her own
queue).
c. Arriving workers choose one of three queues at random.
d. Arriving workers choose the queues by turn.
har01307_LB04_419-464.indd 461 20/12/10 1:54 PM

462 Part II Labs
15. At the ShopNSave, a small family-oriented grocery store, there are

only four aisles: aisle 1—fruits/vegetables, aisle 2—packaged goods
(cereals, crackers, cookies, and the like), aisle 3—dairy products, and
aisle 4—meat/fish. The time between two successive customer arrivals
is exponentially distributed with a mean of five minutes. After arriving
to the store, each customer grabs a shopping cart. Twenty percent of all
customers go to aisle 1, 30 percent go to aisle 2, 50 percent go to aisle 3,
and 70 percent go to aisle 4. The number of items selected for purchase
in each aisle is uniformly distributed between two and eight. The time
spent to browse and pick up each item is N(5,2) minutes. There are three
identical checkout counters: each counter has its own checkout line. The
customer chooses the shortest line. Once a customer joins a line, he or
she is not allowed to leave or switch lines. The checkout time is given by
the following equation:
Checkout time ⫽ N(3,0.3) ⫹ (# of items) * N(0.5,0.15) minutes
The first term of the checkout time is for receiving payment (cash,
check, or credit card) from the customer, opening and closing the cash
register, and handing over the receipt and cash back to the customer.
After checking out, a customer leaves the cart at the front of the store
and leaves. Build a simulation model for the grocery store. Simulate for
2,000 hours of operation and respond to the following questions:
a. The percentages of customers visiting each aisle do not add up to 100
percent. Why?
b. What is the average and maximum amount of time a customer spends
at the grocery store?
c. How many customers check out per cashier per hour?
d. What is the average and maximum amount of time a customer spends
waiting in the checkout line?
e. What is the average utilization of the cashiers?
f. Assuming there is no limit to the number of shopping carts; determine
the average and maximum number of carts in use at any time.
g. What is the average and maximum number of customers awaiting in
line to checkout?
h. What is the average and maximum number of customers in the store?
i. If the owner adopts a customer service policy that there will never
be any more than three customers in any checkout line, how many
cashiers are needed?
16. At the ShopNSave grocery store, the store manager is considering
designating one of the checkout lines as Express, for customers checking
out with 10 or fewer items. Is that a good idea? Why or why not?
17. Realistically, at the ShopNSave grocery store, there are only 10
shopping carts. If there are no carts available when customers arrive
they wait at the front of the store for a cart to become available. Modify
the simulation model (Exercise 15) to reflect this change. Respond to
har01307_LB04_419-464.indd 462 20/12/10 1:54 PM

questions a–i as given in Exercise 15. What percentage of customers has

to wait for an empty shopping cart? How many shopping carts should
they have so that no more than five percent of customers are waiting for
shopping carts?
18. Planes arrive at the Netaji Subhash Chandra Bose International Airport,
Kolkata with interarrival times that are exponentially distributed [E(30)
minutes)]. If there is no room at the airport when a plane arrives, the
pilot flies around and comes back to land after a normally distributed
time [N(20,5) minutes]. There are two runways and three gates at this
small airport. The time from touchdown to arrival at a gate is normally
distributed [N(5,1) minutes]. A maximum of three planes can be
unloaded and loaded at the airport at any time. The times to unload and
load a plane are uniformly distributed [U(25,5) minutes]. The push off,
taxi, and takeoff times are normally distributed [N(6,1) minutes]. The
airport operates 24/7. Build a simulation model, run it for 2,000 hours,
and respond to the following questions:
a. How much will adding a new gate decrease the time that airplanes
have to circle before landing?
b. Will adding a new gate affect the turnaround time of airplanes at the
airport? How much?
c. Is it better to add another runway instead of another gate? If so why?
har01307_LB04_419-464.indd 463 20/12/10 1:54 PM

har01307_LB04_419-464.indd 464 20/12/10 1:54 PM
L A B
5 FITTING STATISTICAL
DISTRIBUTIONS
TO INPUT DATA
There are three kinds of lies: lies, damned lies, and statistics.
—Benjamin Disraeli
Input data drives our simulation models. Input data can be for interarrival times,
material handling times, setup and process times, demand rates, loading and un-
loading times, and so forth. The determination of what data to use and where
to get the appropriate data from is a complicated and time-consuming task. The
quality of data is also very important. We have all heard the cliché “garbage in,
garbage out.” In Chapter 5 we discussed various issues about input data collection
and analysis. We have also described various empirical discrete and continuous
distributions and their characteristics (Chapter 5). In this lab we describe how
ProModel helps in fitting statistical distributions to user-input data. In section
L5.1 we introduce you to the utility Stat::Fit. In section L5.2 we show the applica-
tion of Stat::Fit with an example problem. In section L5.3 we show you how to
use the Auto::Fit function in Stat::Fit to automatically fit an appropriate distribu-
tion to the input data.
L5.1 An Introduction to Stat::Fit

Stat::Fit is a utility packaged with the ProModel software, which is available
from the opening screen of ProModel (Figure L1.1) or from the Tools menu. It
can also be opened from the Windows Start → Programs menu. The Stat::Fit
opening screen is shown in Figure L5.1. Stat::Fit can be used for analyzing user-
input data and fitting an appropriate analytical distribution. (The student version
of ProModel is limited to three distributions: normal, uniform, and exponential.)
An analytical distribution can be either continuous or discrete. If there are enough
data points, say 100 or more, it may be appropriate to fit an analytical distribution
465
har01307_LB05_465-474.indd 465 12/17/10 9:54 AM

466 Part II Labs
FIGURE L5.1
Stat::Fit opening
screen.
to the data using conventional methods. Stat::Fit offers the capability to perform
input data analysis and fit probability distributions to empirical data. It allows
comparison among various distribution functions. It performs goodness-of-fit
tests using chi-square, Kolmogorov-Smirnov, and Anderson-Darling procedures.
It calculates appropriate parameters for distributions. It provides distribution ex-
pressions for use in the simulation model. When the amount of data is small, the
goodness-of-fit tests are of little use in selecting one distribution over another. It
is inappropriate to fit one distribution over another in such a situation. Also, when
conventional techniques have failed to fit a distribution, a frequency distribution
of the data is used directly as a user distribution (Chapter 5, section 5.9).
The opening menu of Stat::Fit is shown in Figure L5.2. Various options are
available in the opening menu:
1. File: This menu opens a new Stat::Fit project or an existing project or
data file. The File menu is also used to save a project.
2. Edit: The edit menu is used to access the Cut, Copy, Paste, Insert, Delete,
and Clear commands.
3. Input: The input menu is used to access various input options,
mathematical operations, transformations, and filters. It is also used to
repopulate the input data, generate distributions, and toggle between the
input data table and input graph.
4. Statistics: This menu is used to access descriptive statistics, binned data,
and run various independence tests for the input data.
5. Fit: This menu provides a Fit Setup dialog and a Distribution Graph
dialog. Other options are also available when a Stat::Fit project is
opened. The Fit Setup dialog lists all the distributions supported by
Stat::Fit and the relevant choices for goodness-of-fit tests. At least one
har01307_LB05_465-474.indd 466 12/17/10 9:54 AM

FIGURE L5.2
Stat::Fit opening
menu.
distribution must be chosen before the estimate, test, and graphing

commands become available. The Distribution Graph command uses the
distribution and parameters provided in the Distribution Graph dialog to
create a graph of any analytical distribution supported by Stat::Fit. This
graph is not connected to any input data or document.
6. Utilities: The Replications command allows the user to calculate the
number of independent data points or replications of an experiment
necessary to provide a given range or confidence interval for the estimate
of a parameter. The confidence interval is given for the confidence level
specified. The default is a 95 percent confidence interval. The resulting
number of replications is calculated using the t distribution.
7. View: This menu is used to access the toolbar (turn it on/off) and the
status bar (turn it on/off).
8. Window: The Window menu is used to either cascade or tile various
windows opened while working on the Stat::Fit projects.
9. Help: This menu allows us to access various help topics within Stat::Fit.
Figure L5.2 shows the data/document input screen. Figure L5.3 shows the
various data input options available in Stat::Fit. The type of distribution is also
specified here. The Input Options command can be accessed from the Input menu
as well as the Input Options button on the Speed bar.
L5.2 An Example Problem

Problem Statement
The times between arrivals of cars at San Dimas Gas Station were collected as
shown in Table L5.1. This information is also saved and available in an Excel work-
sheet named L5.2_Gas Station Time Between Arrivals. Use Stat::Fit to analyze the
har01307_LB05_465-474.indd 467 12/17/10 9:54 AM

468 Part II Labs
FIGURE L5.3
Stat::Fit data input
options.
TABLE L5.1 Times between Arrivals of Cars at San Dimas Gas Station
Number Times between Arrivals, Minutes
1 12.36
2 5.71
3 16.79
4 18.01
5 5.12
6 7.69
7 19.41
8 8.58
9 13.42
10 15.56
11 10
12 18
13 16.75
14 14.13
15 17.46
16 10.72
17 11.53
18 18.03
19 13.45
20 10.54
21 12.53
22 8.91
23 6.78
24 8.54
25 11.23
26 10.1
27 9.34
28 6.53
29 14.35
30 18.45
har01307_LB05_465-474.indd 468 12/20/10 6:03 PM

data and fit an appropriate continuous distribution to the data. Figure L5.4 shows
part of the actual data of times between arrivals of cars in minutes. The sample
data have 30 data points. Figure L5.5 shows the histogram of the input data, while
Figure L5.6 shows some of the descriptive statistics generated by Stat::Fit.
FIGURE L5.4
Times between arrivals
of cars at San Dimas
Gas Station.
FIGURE L5.5
Histogram of the times
between arrival data.
FIGURE L5.6
Descriptive statistics
for the input data.
har01307_LB05_465-474.indd 469 12/17/10 9:54 AM

470 Part II Labs
L5.3 Auto::Fit Input Data

Automatic fitting of continuous distributions can be performed by clicking on the
Auto::Fit icon or by selecting Fit from the Menu bar and then Auto::Fit from the
Submenu (Figure L5.7).
This command follows the same procedure as discussed in Chapter 5 for
manual fitting. Auto::Fit will automatically choose appropriate continuous distri-
butions to fit to the input data, calculate Maximum Likelihood Estimates for those
distributions, test the results for Goodness of Fit, and display the distributions
in order of their relative rank. The relative rank is determined by an empirical
method, which uses effective goodness-of-fit calculations. While a good rank usu-
ally indicates that the fitted distribution is a good representation of the input data,
an absolute indication of the goodness of fit is also given.
The Auto::Fit dialog allows the number of continuous distributions to be lim-
ited by choosing only those distributions with a lower bound or by forcing a lower
bound to a specific value as in Fit Setup. Also, the number of distributions will be
limited if the skewness of the input data is negative; many continuous distribu-
tions with lower bounds do not have good parameter estimates in this situation.
The acceptance of fit usually reflects the results of the goodness-of-fit tests
at the level of significance chosen by the user. However, the acceptance may be
modified if the fitted distribution would generate significantly more data points in
the tails of the distribution than are indicated by the input data.
The Auto::Fit function forces the setup of the document so that only continu-
ous distributions will be used. Figure L5.8 shows various continuous distributions
fitted to the input data of the San Dimas Gas Station (section L5.2) and their rank
FIGURE L5.7
The Auto::Fit
submenu.
FIGURE L5.8
Various continuous
distributions fitted to
the input data.
har01307_LB05_465-474.indd 470 12/17/10 9:54 AM

FIGURE L5.9
Goodness-of-fit tests performed on the input data.
FIGURE L5.10
Comparison of actual
data and fitted uniform
distribution.
in terms of the amount of fit. Both the Kolmogorov-Smirnov and the Anderson-
Darling goodness-of-fit tests will be performed on the input data as shown in Fig-
ure L5.9. The Maximum Likelihood Estimates will be used with an accuracy of
at least 0.00003. The actual data and the fitted uniform distribution are compared
and shown in Figure L5.10.
har01307_LB05_465-474.indd 471 12/17/10 9:54 AM

472 Part II Labs
Because the Auto::Fit function requires a specific setup, the Auto::Fit view
can be printed only as the active window or part of the active document, not
as part of a report. The Auto::Fit function will not fit discrete distributions. The
manual method, previously described, should be used instead.
L5.4 Exercises
1. Consider the operation of a fast-food restaurant where customers arrive
for ordering lunch. The following is a log of the time (minutes) between
arrivals of 40 successive customers. Use Stat::Fit to analyze the data and
fit an appropriate continuous distribution. What are the parameters of this
distribution?
11 11 12 8 15 14 15 13
9 13 14 9 14 9 13 7
12 12 7 13 12 16 7 10
8 8 17 15 10 7 16 11
11 10 16 10 11 12 14 15
2. The servers at the restaurant in Exercise 1 took the following time

(minutes) to serve food to these 40 customers. Use Stat::Fit to analyze
the data and fit an appropriate continuous distribution. What are the
parameters of this distribution?
11 11 12 8 15 14 15 13
9 13 14 10 14 9 13 12
12 12 11 13 12 16 11 10
10 8 17 12 10 7 13 11
11 10 13 10 11 12 14 15
3. The following is the number of incoming calls (each hour for 80 successive
hours) to a call center setup for serving customers of a certain Internet
service provider. Use Stat::Fit to analyze the data and fit an appropriate
discrete distribution. What are the parameters of this distribution?
12 12 11 13 12 16 11 10
9 13 14 10 14 9 13 12
12 12 11 13 12 16 11 10
10 8 17 12 10 7 13 11
11 11 12 8 15 14 15 13
9 13 14 10 14 9 13 12
12 12 11 13 12 16 11 10
10 8 17 12 10 7 13 11
11 10 13 10 11 12 14 15
10 8 17 12 10 7 13 11
har01307_LB05_465-474.indd 472 12/17/10 9:54 AM

4. Observations were taken on the times to serve online customers at a

stockbroker’s site (STOCK.com) on the Internet. The times (in seconds)
are shown here, sorted in ascending order. Use Stat::Fit and fit an
appropriate distribution. What are the parameters of this distribution?
1.39 21.47 39.49 58.78 82.10

3.59 22.55 39.99 60.61 83.52
7.11 28.04 41.42 63.38 85.90
8.34 28.97 42.53 65.99 88.04
11.14 29.05 47.08 66.00 88.40
11.97 35.26 51.53 73.55 88.47
13.53 37.65 55.11 73.81 92.63
16.87 38.21 55.75 74.14 93.11
17.63 38.32 55.85 79.79 93.74
19.44 39.17 56.96 81.66 98.82
5. Forty observations for a bagging operation were shown in Chapter 5,

Table 5.5. Use Stat::Fit to find out what distribution best fits the
data. Compare your results with the results obtained in Chapter 5.
(Suggestion: In Chapter 5, five equal probability cells were used for the
chi-square goodness-of-fit test. Use same number of cells in Stat::Fit to
match the results.)
6. One hundred observed inspection operation times were shown in
Chapter 5, Table 5.2. Use Stat::Fit to find out what distribution best fits
the data. Compare your results with the results obtained in Chapter 5.
(Suggestion: In Chapter 5, six equal probability cells were used for the
chi-square goodness-of-fit test. Use same number of cells in Stat::Fit to
match the results.)
7. While doing your homework you were interrupted frequently by
your friends. Thirty observations of times between interruptions were
recorded and shown in Exercise 20 (Chapter 5, section 5.14). Use
Stat::Fit to find out if the data are exponentially distributed.
8. A call center for Bank of USA wishes to determine if the length of
customers’ phone calls can be adequately modeled by an exponential
distribution. Last week, the agency recorded the length of all phone calls
and obtained the following results (in seconds): 4, 100, 6, 33, 37, 43, 5,
8, 88, 120, 9, 10, 12, 8, 16, 130, 20, 70, 78, 24, 27, 58, 50, 68. Do these
data indicate that the length of phone calls to the call center is governed
by an exponential distribution?
har01307_LB05_465-474.indd 473 12/17/10 9:54 AM

har01307_LB05_465-474.indd 474 12/17/10 9:54 AM
L A B
6 INTERMEDIATE MODEL
BUILDING
All truths are easy to understand once they are discovered; the point is to
discover them.
—Galileo Galilei
In this lab we expand on the basic ProModel concepts discussed in Lab 4. Sec-
tion L6.1 introduces the concept of attributes. Section L6.2 shows its application
in the calculation of cycle times. In section L6.3 we simulate the process of sorta-
tion, sampling inspection, and rework. Section L6.4 discusses various aspects of
machine breakdowns and maintenance; section L6.5 shows how to model shift-
working patterns; section L6.6 presents an application of ProModel in a job shop;
section L6.7 introduces the modeling of priorities; and section L6.8 has a couple
of examples of pull system applications in manufacturing. Costs are modeled and
tracked in section L6.9. Section L6.10 shows how to import background graphics
into a model, and section L6.11 shows how to define and display various views
of a model. Section L6.12 describes the creation of a model package for ease of
distribution to others.
L6.1 Attributes
Attributes can be defined for entities or for locations. Attributes are placeholders
similar to variables but are attached to specific entities or locations and usually
contain information about that entity or location. Attributes are changed and as-
signed when an entity executes the line of logic that contains an operator, much
like the way variables work. Some examples of attributes are part type, customer
number, and time of arrival of an entity, as well as length, weight, volume, or
some other characteristic of an entity.
475
har01307_LB06_475-516.indd 475 20/12/10 1:54 PM

476 Part II Labs
To define an attribute use the attribute editor, as follows:

1. Go to the Build Attributes menu and create a name (ID) for the attribute.
2. Select the type of attribute—integer or real.
3. Select the class of attribute—entity or location.
L6.1.1 Using Attributes to Track Customer Types

Problem Statement
Customers visit the neighborhood barbershop Fantastic Dan for a haircut. Among
the customers there are 20 percent children, 50 percent women, and 30 percent
men. The customer interarrival time is triangularly distributed with a minimum,
mode, and maximum of seven, eight, and nine minutes respectively. The haircut
time depends on the type of customer and is given in Table L6.1. This time also in-
cludes the initial greetings and the transaction of money at the end of the haircut.
Run the simulation model for one day (480 minutes). About how many customers
of each type does Dan serve per day? Also, track the number of customers (of
each type) in the salon throughout the day.
Two locations (Barber Dan and Waiting for Dan) and an entity (Customer)
are defined. Customer_Type is defined as an attribute (type—integer and
classification—entity) as shown in Figure L6.1. Declare three global variables—
man, woman, and child to keep track of the different types of customers
(Figure L6.2). The process/routing and customer arrivals are defined as shown in
Figures L6.3 and L6.4. A snapshot of the simulation model is shown in Fig-
ure L6.5. Figure L6.6 tracks the number of male customers in the salon through-
out the day.
TABLE L6.1 Haircut Times for All Customers
Haircut Time (minutes)
Customer Type Mean Half-Width
Children 8 2
Women 12 3
Men 10 2
FIGURE L6.1
Attribute declaration.
har01307_LB06_475-516.indd 476 20/12/10 1:54 PM

FIGURE L6.2
Variable definitions.
FIGURE L6.3
Process and routing tables for Fantastic Dan.
FIGURE L6.4
Arrival of customers at Fantastic Dan.
har01307_LB06_475-516.indd 477 20/12/10 1:54 PM

478 Part II Labs
FIGURE L6.5
A snapshot of the
simulation of model
for Fantastic Dan.
FIGURE L6.6
Tracking the number of male customers in the salon.
L6.1.2 Using Attributes and Local Variables

In this example we demonstrate the use of local variables in conjunction with at-
tributes to track and process various part types.
California Pipes and Tubes manufactures various types of steel pipes. The
three major product lines are stainless steel, alloy steel, and carbon steel pipes.
An inspector inspects all pipes after which they move to a packing station (Fig-
ure L6.7). A packer loads these pipes into boxes. The inspection time is normally
distributed N(6,2) minutes. The packing time is also normally distributed N(5,1)
minutes. The boxes are designed to hold only certain types of pipes. Therefore a
box designated to hold stainless steel pipes can only hold stainless steel pipes.
Boxes arrive to the Box Arrival queue. Then they move on to a multicapacity
location Packer. Each Box has a different entity attribute, b_type, describing the
har01307_LB06_475-516.indd 478 20/12/10 1:54 PM

FIGURE L6.7
Layout of California Pipes and Tubes.
type of pipe it can hold. The packing worker loads pipes into the Box. Workers must
load the stainless steel pipes into the box designated to hold stainless steel pipes,
alloy steel pipes into the box designated to hold alloy steel pipes and so on. There-
fore, the attribute value of the Pipe, p_type, must match the attribute value of the
Box, b_type. We can use local variables to accomplish this modeling task. In this
example, we have defined X as a local variable and set it equal to b_type. Figure 6.8
shows the process and routing tables for California Pipes and Tubes.
After the pipes have been inspected and packed into respective boxes they are
sent to separate rack locations for eventual shipment to customers. Simulate for
2,000 hours of operation and track the various types of pipes shipped.
L6.2 Cycle Time

The Clock and Log are functions built into ProModel to allow us to keep track
of the time entities spent in the system or in a segment of the system. The Clock
function returns the current simulation clock time in hours, minutes, seconds, or
internal time units. The value returned is of type real.
The Log function automatically subtracts the time stored in an attribute from
the current simulation clock time and stores the result with a text string header.
Time_In Clock()
Log “Cycle Time ”, Time_In
har01307_LB06_475-516.indd 479 20/12/10 1:54 PM

480 Part II Labs
FIGURE L6.8
Process and routing tables for California Pipes and Tubes.
For the example in L6.1, find

a. The cycle time for each type of customer of Barber Dan.
b. The average cycle time for all customers.
c. The average time spent in the salon by each type of customer.
d. The average waiting time for each type of customers.
We define Time_In as an attribute to track the time of arrival for all customer
types. Figure L6.9 shows the setting of the attribute Time_In and also the logging
of the average cycle time. The cycle times for children, women, men, and all cus-
tomers (of any type) are reported in Figure L6.10.
L6.3 Sorting, Inspecting a Sample, and Rework

Problem Statement
Discs-R-Us Manufacturing Inc. receives orders for two types of discs (disc A
and disc B). Disc A and disc B orders arrive on average every five and every ten
minutes (exponentially distributed), respectively. Both discs arrive at the input
queue. An attribute Part_Type is defined to differentiate between the two types
of discs.
Disc A goes to the lathe for turning operations that take Normal(5,1) minutes.
Disc B goes to the mill for processing that takes Uniform(6,2) minutes. Both type
of discs go on to an inspection queue, where every fifth part is inspected. Discs that
are not inspected are shipped directly. Inspection takes Normal(6,2) minutes. After
har01307_LB06_475-516.indd 480 20/12/10 1:54 PM

FIGURE L6.9
Setting the attribute Time_In and logging the cycle time.
FIGURE L6.10
The minimum, maximum, and average cycle times for various customers.
har01307_LB06_475-516.indd 481 20/12/10 1:54 PM

482 Part II Labs
FIGURE L6.11
Keeping track of machined_qty and modeling probabilistic routings at the Inspect location.
FIGURE L6.12
A snapshot of the simulation model for Discs-R-Us.
inspection 70 percent of the discs pass and leave the system; 30 percent of the discs
fail and are sent back to the input queue for rework. Determine the following:
a. Number of discs of each type shipped each week (40-hour week)?
b. Number of discs machined each week?
c. Number of discs inspected each week?
d. Average cycle time for each type of disc?
e. Number of discs reworked each week?
f. Average number of discs waiting in the input queue and the inspection
queue?
Five locations (Mill, Input_Q, Lathe, Inspect, and Inspect_Q) are defined for
this model. Four variables (qty, machined_qty, inspect_qty, rework_qty) are de-
fined. Figure L6.11 shows how we keep track of the machined quantity as well as
model the probabilistic routings after inspection. Figure L6.12 shows a snapshot
of the simulation model with variables added for keeping track of the number of
discs machined, inspected, and reworked.
har01307_LB06_475-516.indd 482 20/12/10 1:54 PM

L6.4 Preventive Maintenance and Machine Breakdowns

Downtime stops a location or resource from operating. Downtime can occur in
one of two ways: preventive maintenance or breakdown. A down resource (or
location) no longer functions and is not available for use. Downtimes may rep-
resent scheduled interruptions such as shifts, breaks, or scheduled maintenance.
They may also represent unscheduled and random interruptions such as equip-
ment failures.
The mean time between failures (MTBF) can be calculated as the recipro-
cal of the failure rate (distribution of failure over time). Often MTBF will fol-
low a negative exponential distribution. In particular, the probability of failure
before time T is given by 1 eT/MTBF. The reliability is viewed as performance
over time and is the probability that a given location or resource will perform
its intended function for a specified length of time T under normal conditions
of use:
Reliability P(no failure before time T ) eT/MTBF
The mean time to repair (MTTR) refers to the time the location or resource
remains down for repair. The MTTR depends on the ease and/or cost with which
a location or resource can be maintained or repaired.
For single-capacity locations, downtime of a location or resource can be
scheduled at regular intervals based on the clock time that has expired, the number
of entities processed at a location, the usage of the machine (resource) in time, a
change in entity type, or whenever some event occurs in the simulation. Unsched-
uled breakdowns of a machine can also be modeled in a similar fashion.
To model preventive maintenance or breakdowns, use the following
procedure:
1. Go to the Build/Locations menu.
2. Click on the DT button for a particular location record and select the
type of downtime you want to define from the menu.
a. Clock based.
b. Number of entries based.
c. Usage based.
d. Call based.
3. Fill in the information for each of the fields in the table that appears.
4. In the logic field, enter any logic associated with downtime (this is where
you enter the duration of the downtime).
Example:
DISPLAY “The Lathe is Down for Preventive Maintenance”
WAIT N(10,2) min
5. To disable the downtime feature, click Yes in the disable field.

Otherwise, leave the field as No.
har01307_LB06_475-516.indd 483 20/12/10 1:54 PM

484 Part II Labs
TABLE L6.2 Maintenance Schedule for Machines at Discs-R-Us

Manufacturing Inc.
Machine Time Between Repairs Time to Repair
Lathe 120 minutes N(10,2) minutes

Mill 200 minutes T(10,15,20) minutes
FIGURE L6.13
Logic for preventive maintenance at Discs-R-Us Manufacturing Inc.
L6.4.1 Downtime Using MTBF and MTTR Data

The lathe and the mill at Discs-R-Us Manufacturing Inc. (section L6.3) have
the maintenance schedule shown in Table L6.2. The logic for the preventive
maintenance operation for the lathe is shown in Figure L6.13. Note the use of
the DISPLAY statement to notify the user when the lathe goes down for preven-
tive maintenance. The processes and routings were shown in Figure L6.11.
Figure L6.14 shows a snapshot of the simulation model with the lathe down for
preventive maintenance.
L6.4.2 Downtime Using MTTF and MTTR Data

Reliability of equipment is frequently calibrated in the field using the mean time
to failure (MTTF) data instead of the MTBF data. Data collected in the field are
often based on time to failure (TTF) for clock-based failures.
har01307_LB06_475-516.indd 484 20/12/10 1:54 PM

FIGURE L6.14
A snapshot of the simulation model for Discs-R-Us Manufacturing Inc.
FIGURE L6.15
Layout of the machine
shop—modeling
breakdown with TTF
and TTR.
Problem Statement
The turning center in this machine shop (Figure L6.15) has a TTF distribution that
is exponential with a mean of 10 minutes. The repair time or time to repair (TTR)
is also distributed exponentially with a mean of 10 minutes.
This model shows how to get ProModel to implement downtimes that use
TTF rather than time between failures (TBF) data. In practice, you most likely
will want to use TTF because that is how data will likely be available to you, as-
suming you have unexpected failures (breakdowns). If you have regularly sched-
uled downtimes (preventive maintenance), it may make more sense to use TBF.
In this example, the theoretical percentage of uptime is MTTF/(MTTFMTTR),
where M indicates a mean value. The first time to failure and time to repair are set
in the variable initialization section (Figure L6.16). Others are set in the downtime
logic (Figure L6.17). Note that the downtime is not scheduled.
har01307_LB06_475-516.indd 485 20/12/10 1:54 PM

486 Part II Labs
FIGURE L6.16
Variable initializations.
FIGURE L6.17
Clock downtime logic.
FIGURE L6.18
Process and routing tables.
The processing and routing tables are shown in Figure L6.18. We will run
the simulation for 1,000 hours and collect batch mean statistics for analyzing
the output. The method of batch means, or interval batching, is a way to col-
lect independent samplesas an alternative to running multiple replications when
simulating steady-state systems. In batch means, the simulation is divided into
consecutive periods or intervals and statistics are collected on each period. Each
period is treated as though it were a replication. The advantage over running mul-
tiple replications is that the warm-up period runs only once (not a huge advantage
given modern computing speeds).
Batch mean statistics are collected by selecting the “batch mean” option in
the Simulation Options dialog (Figure L6.19). We will set the interval length to
1,200 minutes (20 hours) for batching the data. The number of intervals is deter-
mined by dividing the run length (1,000 hours) by the interval length. The batch
mean statistics for different states of the turning center (including the percent
down) are displayed by selecting the single-capacity state table as shown in Fig-
ure L6.20. The table that appears is shown in Figure L6.21 that automatically
displays the average of the interval means for each state. Additional statistics for
each state such as its min, max, std dev, and confidence interval values can be
displayed by selecting these options from the filter on the left (see Figure L6.22).
(Dr. Stephen Chick, INSEAD contributed this problem.)
har01307_LB06_475-516.indd 486 20/12/10 1:54 PM

FIGURE L6.19
Batch mean option
chosen in Simulation
dialog.
FIGURE L6.20
Output viewer showing selection of the single-capacity location states.
har01307_LB06_475-516.indd 487 20/12/10 1:54 PM

488 Part II Labs
FIGURE L6.21
Average of all the batches in the ProModel Output Viewer.
FIGURE L6.22
Additional statistics for batch mean, single-capacity location state values.
L6.5 Shift Working Schedule

ProModel offers an excellent interface to include the operator (location or re-
source) shift working schedules in the simulation model. The shift work for each
operator in the model can be defined separately. Work cycles that repeat either
daily or weekly can be defined.
Problem Statement
For the Discs-R-Us Manufacturing Inc. (section L6.3) an operator is used to
process the parts at both the lathe and the mill. The operator is also used to inspect
the parts. The operator is on a shift schedule from 8 A.M. until 5 P.M. with breaks
as shown in Table L6.3 and Figures L6.23 and L6.24.
har01307_LB06_475-516.indd 488 20/12/10 1:54 PM

TABLE L6.3 The Operator Break Schedule at

Discs-R-Us Manufacturing Inc.
Breaks From To
Coffee break 10 A.M. 10:15 A.M.

Lunch break 12 noon 12:45 P.M.
Coffee break 3 P.M. 3:15 P.M.
FIGURE L6.23
Weekly shift schedule at Discs-R-Us Manufacturing Inc.
FIGURE L6.24
Assigning the Shift File to the locations—Lathe, Mill, and Inspect.
har01307_LB06_475-516.indd 489 20/12/10 1:54 PM

490 Part II Labs
We are using the DISPLAY statement to notify the user when the operator is
on a break (Figure L6.25). The process and the routing tables were shown in Fig-
ure L6.11. Determine the following:
a. The quantity of discs of each type shipped each week (40-hour week).
b. The cycle time for each type of discs.
c. The maximum and minimum cycle times.
d. The number of discs reworked each week.
e. The average number of discs waiting in the inspection queue.
A snapshot of the model using the shift breaks is shown in Figure L6.26. Run
the model with and without shift breaks. What difference do you notice in these
statistics? How can you improve the system?
FIGURE L6.25
Use the DISPLAY
statement to notify the
user when the operator
is on a break.
FIGURE L6.26
A snapshot of the Discs-R-Us model using shift break schedule.
har01307_LB06_475-516.indd 490 20/12/10 1:54 PM

L6.6 Job Shop

A job shop is a collection of processing centers through which jobs are routed in
different sequences depending on the characteristics of the particular job. It is a
common method of grouping resources for producing small lots with widely vary-
ing processing requirements.
Each job in a job shop usually has a unique routing sequence. Job shops
usually have a process layout. In other words, similar processes are physically
grouped together.
Problem Statement
In Joe’s Jobshop there are three machines that process three types of jobs. All
jobs go to all machines, but with different process sequences. The data for job
routings and processing times are given in Table L6.4. The process times are
exponentially distributed with the given average values. Jobs arrive on average
every 30 minutes (exponentially distributed). Simulate for 10 days (80 hours).
How many jobs of each type are processed in 10 days?
Define the locations as shown in Figure L6.27. Define four entities—Job,
Job_A, Job_B, and Job_C. Figure L6.28 shows three variables defined to track
the three types of jobs processed. The processing and routing tables are shown in
Figure L6.29. A snapshot of the simulation model is shown in Figure L6.30.
TABLE L6.4 Summary of Process Plan for Joe’s Jobshop
Jobs Machines Process Times (minute) Job Mix (%)
A 2–3–1 45–60–75 25%

B 1–2–3 70–70–50 35
C 3–1–2 50–60–60 40
FIGURE L6.27
Locations at Joe’s
Jobshop.
FIGURE L6.28
Variables to track jobs
processed.
har01307_LB06_475-516.indd 491 12/20/10 7:00 PM

492 Part II Labs
FIGURE L6.29
Processing and routing tables at Joe’s Jobshop.
FIGURE L6.30
A snapshot of the
Joe’s Jobshop.
L6.7 Modeling Priorities

Priorities allow us to determine the order in which events occur in the simulation.
The most common uses of priorities are:
1. Selecting among processes.
2. Selecting among downstream processes—already discussed in Lab 4,
section L4.4.
har01307_LB06_475-516.indd 492 20/12/10 1:54 PM

3. Selecting resources.
4. Prioritizing downtimes—discussed in section L6.4.
L6.7.1 Selecting among Upstream Processes

This is the process of choosing among a set of upstream processes or queues for
removal and routing to a downstream process. When one downstream destination
exists and two or more upstream entities are competing to get there, priorities
can be used. In the example in Figure L6.31 two entities at different locations
(In_Q_Domestic and In_Q_Export) are both trying to get to the downstream pro-
cess (Machining Center).
Problem Statement
At Singh’s Export Machine Shop in the suburbs of Mumbai, two types of jobs
are processed: domestic and export. All jobs are processed through a machin-
ing center and a lathe. The processing times for all jobs on the machining center
and the lathe are Triangular(10,12,18) minutes and Triangular(12,15,20) minutes,
respectively.
Export jobs are given priority over domestic jobs for shop release—that is, in
moving from the input queue to the machining center. Five locations are defined:
Lathe, Machining_Center, In_Q_Domestic, In_Q_Export, and Outgoing_Q. Two
entities (domestic and export) are defined. The rate of arrival of both type of
jobs is Exponential(30) minutes. The processing and routing tables are shown in
Figure L6.32. Note that the priority of domestic jobs is set at a value of 1 while
that of export jobs is set at 10. Higher priority numbers signify higher priority in
terms of selecting the upstream process. The priorities can range from 0 to 999.
The default priority is 0. Simulate for 100 days (800 hours).
FIGURE L6.31
A snapshot of the
simulation model
for Singh’s Export
Machine Shop.
har01307_LB06_475-516.indd 493 20/12/10 1:54 PM

494 Part II Labs
FIGURE L6.32
Processing and routing tables for Singh’s Export Machine Shop.
FIGURE L6.33
Time plot of the input queue contents (domestic and export).
Note that the input queue for export jobs remains quite small because of
higher priority given to export jobs, while the input queue for domestic jobs
(lower priority) keeps growing to unsustainable levels (Figure L6.33).
L6.8 Modeling a Pull System

A pull system is a system in which locations produce parts only on downstream
demand. There are two types of pull systems:
1. Those based on limited buffer or queue size.
2. Those based on more distant “downstream” demand.
The first type of pull system is modeled in ProModel by defining locations
with limited capacity. In this type of system, upstream locations will be able to
har01307_LB06_475-516.indd 494 20/12/10 1:54 PM

send parts to downstream locations only when there is capacity available. In the
second type of pull system we model a kanban system using the Toyota Produc-
tion System (TPS) philosophy.
L6.8.1 Pull Based on Downstream Demand

The second type of pull system requires the use of a SEND statement from a down-
stream location to trigger part movement from an upstream location.
Problem Statement
At the Milwaukee Machine Shop (MMS), two types of aircraft parts are pro-
cessed in a machining cell. The cell consists of one lathe and one mill. Type A
jobs must be processed first on the lathe and then on the mill. Type B jobs are pro-
cessed only on the mill (Table L6.5). The aircraft manufacturer (customer) needs
a limited number of the two types of jobs (Table L6.5). All jobs are processed on
a first-in, first-out basis.
Brookfield Forgings is a supplier of forgings for MMS. Forgings are produced
in batches of five every day (exponential with a mean of 24 hours). However, the
forgings are supplied to MMS only on receipt of an order (for a batch of 5 forg-
ings) from the customer for a specific type of job. In other words, when customer
orders arrive at MMS, raw forgings are supplied by Brookfield Forgings (a pull
system of shop loading). The aircraft manufacturer (customer) needs only a lim-
ited number of the two types of jobs (Table L6.5). Track the work-in-process
inventories and the production quantities for both types of jobs using a simulation
model.
Six locations (Mill, Lathe, Brookfield_Forgings, Order_Arrival, Lathe_Q,
and Mill_Q) and four entities (Type_A, Type_B, Orders_A, and Orders_B) are
defined. The arrivals of various entities are shown in Figure L6.34. Five global
variables (wip_A, wip_B, wip_Total, prod_A, and prod_B) are defined to track
the work-in-process and production of the two types of jobs. The processes and
routings are shown in Figure L6.35. Note that as soon as a customer order is re-
ceived at the MMS, a signal is sent (SEND) to Brookfield Forgings to ship a forging
(of the appropriate type) to MMS. Thus the arrival of customer orders pulls the
raw material from the vendor. When the parts are fully machined, they are paired
(JOINed) with the appropriate customer order at the orders arrival location. Figure
L6.36 shows a layout of the MMS and a snapshot of the simulation model.
TABLE L6.5 Order Arrival and Processing Time Data

for the Milwaukee Machine Shop
Number of Time between Processing Time Processing Time

Job Type Orders Order Arrival on Lathe on Mill
Type_A 120 E(8) hrs. E(3) hrs. U(3,1) hrs.

Type_B 100 E(10) hrs. — U(4,1) hrs.
har01307_LB06_475-516.indd 495 20/12/10 1:54 PM

496 Part II Labs
FIGURE L6.34
Arrival of entities at the Milwaukee Machine Shop.
FIGURE L6.35
Processing and routing tables for the Milwaukee Machine Shop.
FIGURE L6.36
A snapshot of the
simulation model
for the Milwaukee
Machine Shop.
har01307_LB06_475-516.indd 496 20/12/10 1:54 PM

L6.8.2 Kanban System

The kanban system is one of the methods of control utilized within the Toyota
Production System (TPS). The basic philosophy in TPS is total elimination of
waste in machines, equipment, and personnel. To make the flow of things as close
as possible to this ideal condition, a system of just-in-time procurement of material
is used—that is, obtain material when needed and in the quantity needed.
“Kanban” literally means “visual record.” The word kanban refers to the
signboard of a store or shop, but at Toyota it simply means any small sign dis-
played in front of a worker. The kanban contains information that serves as a work
order. It gives information concerning what to produce, when to produce it, in
what quantity, by what means, and how to transport it.
Problem Statement
A consultant recommends implementing a production kanban system for the
Milwaukee Machine Shop (MMS) in section L6.8.1. Simulation is used to find out
how many kanbans should be used. Model the shop with a total of five kanbans.
The kanban procedure operates in the following manner:
1. As soon as an order is received by the MMS, they communicate it to the
Brookfield Forgings.
2. Brookfield Forgings holds the raw material in their own facility in the
forging queue in the sequence in which the orders were received.
3. The production of jobs at the MMS begins only when a production
kanban is available and attached to the production order.
4. As soon as the production of any job type is finished,
The locations at the MMS are defined as shown in Figure L6.37. A total of
five entities are defined—Type_A, Type_B, Orders_A, Orders_B, and Kanban.
Kanbans are defined as virtual entities for use in controlling the pull system
mechanism.
FIGURE L6.37
Locations at the Milwaukee Machine Shop.
har01307_LB06_475-516.indd 497 20/12/10 1:54 PM

498 Part II Labs
FIGURE L6.38
Arrival of entities at Milwaukee Machine Shop.
The arrival of two types of jobs at Brookfield Forgings is shown in the arriv-
als table (Figure L6.38). This table also shows the arrival of two types of customer
orders at the MMS. A total of five kanbans are generated at the beginning of the
simulation run.
Figure L6.39 shows a layout of the MMS and a snapshot of the simulation
model. The process and the routing tables are shown in Figure L6.40. The arrival of
a customer order (Orders_A or Orders_B) at the orders arrival location sends a signal
to Brookfield Forgings in the form of a production kanban. The kanban is temporarily
attached (LOADed) to a forging of the right type at the Order_Q. The forgings are sent
to the MMS for processing. After they are fully processed, the kanban is separated
(UNLOADed). The kanban goes back to the kanban square. The finished part is united
(JOINed) with the appropriate customer order at the orders arrival location.
FIGURE L6.39
A snapshot of the simulation model for the Milwaukee Machine Shop.
har01307_LB06_475-516.indd 498 20/12/10 1:54 PM

FIGURE L6.40
Processing and routing tables at the Milwaukee Machine Shop.
L6.9 Tracking Cost

ProModel includes a cost-tracking feature. The following costs can be monitored:
1. Location cost.
2. Resource cost.
3. Entity cost.
The cost dialog box can be accessed from the Build menu. From the menu bar
select Build Cost. This costing feature can be enabled/disabled from the Options
dialog box in the Simulation menu.
Locations
The Locations cost dialog box (Figure L6.41) has two fields—Operation Rate and
Per. Operation rate specifies the cost per unit time to process at the selected loca-
tion. Per is a pull-down menu to set the time unit for the operation rate as second,
minute, hour, or day.
Resources
The Resources cost dialog box (Figure L6.42) has three fields—Regular Rate, Per,
and Cost Per Use. Regular Rate specifies the cost per unit of time for a resource
used in the model. This rate can also be set or changed during run time using the
SETRATE operation statement. Per is a pull-down menu, defined before. Cost Per
har01307_LB06_475-516.indd 499 20/12/10 1:54 PM

500 Part II Labs
FIGURE L6.41
The cost dialog
box—Locations option.
FIGURE L6.42
The Cost dialog
box—Resources option.
FIGURE L6.43
The Cost dialog
box—Entities option.
Use is a field that allows you to define the actual dollar cost accrued every time
the resource is obtained and used.
Entities
The Entities cost dialog box (Figure L6.43) has only one field—Initial Cost. Ini-
tial cost is the cost of the entity when it arrives to the system through a scheduled
arrival.
har01307_LB06_475-516.indd 500 20/12/10 1:54 PM

Explicitly Adjusting Cost

Under certain circumstances, it may be useful to have more direct control over
costs. The costs of a location, resource, or entity can be explicitly incremented by
a positive or negative amount using the operation statements IncLocCost, IncRes-
Cost, or IncEntCost, respectively.
Special Features for Handling Cost

• As soon as you acquire a resource, it begins to accrue cost.
• Unless obtained in the move logic, ProModel charges the “Cost per use”
for a resource to the location that obtained it. Resources obtained in the
move logic do not charge a “per use” cost to any location.
• ProModel does not charge any resource time used during move logic to
any location.
• ProModel adds initial entity costs defined in the cost module only as
entity costs, not location costs.
• If a location uses a resource during a downtime, the location accrues that
resource’s cost.
• The USE statement counts as operation and resource cost.
• When you CREATE a new entity, it begins with new time statistics and an
initial cost.
• If you RENAME an entity, previous time statistics and costs continue with
the entity.
• The SPLIT AS statement divides the cost of the original entity between all
entities. Each new entity begins with new time statistics.
For more information on the cost-tracking feature of ProModel, refer to the
ProModelonline help system.
Problem Statement
Raja’s Manufacturing Cell consists of two mills and a lathe. All jobs are pro-
cessed in the same sequence through the following locations—arriving station,
lathe, mill1, mill 2, and exit station. The processing times and operation costs are
shown in Table L6.6. The arrival rate of jobs is exponentially distributed with a
mean of 20 minutes. The initial cost of the raw material is $100 per piece. Track
the number of jobs produced, the total cost of production, and the cost per piece
of production. Simulate for 200 hours.
TABLE L6.6 Processing Times and Operation Costs at Raja’s

Manufacturing Cell
Machine Processing Time Operation Cost
Lathe Normal(60,10) minutes $10/hour

Mill 1 Normal(30,5) minutes $18/hour
Mill 2 Normal(40,10) minutes $22/hour
har01307_LB06_475-516.indd 501 20/12/10 1:54 PM

502 Part II Labs
FIGURE L6.44
Processing and routing tables for Raja’s Manufacturing Cell.
FIGURE L6.45
A snapshot of the
Raja’s Manufacturing
Cell.
Five locations (Arrive Station, Lathe, Mill 1, Mill 2, and Exit Station) and
three variables (Total_Prod, Total Cost, and Cost_Per_Part) are defined. The pro-
cessing and routing tables are shown in Figure L6.44. A snapshot of the simula-
tion model is shown in Figure L6.45.
L6.10 Importing a Background

Background graphics work as a static wallpaper and enhance the look of a simu-
lation model. They make the model realistic and provide credibility during pre-
sentations. A background could show a floorplan of a factory or any item that is
not part of a location, entity, or resource. Many different graphic formats can be
imported such as BMP, EMF, WMF, GIF, and PCX. Drawings in CAD formats
like DWG must be saved in one of these file formats before they can be imported.
BMP and WMF formatted files can be copied to the clipboard from other applica-
tions and passed directly into the background.
AutoCAD drawings can also be copied to the clipboard and pasted into the
background on the layout of the simulation model. The procedure is as follows:
1. With the graphic on the screen, press <Ctrl> and <C> together.
Alternatively, choose Copy from the Edit menu. This will copy the
graphic into the Windows clipboard.
har01307_LB06_475-516.indd 502 20/12/10 1:54 PM

2. Open an existing or new model file in ProModel.

3. Press <Ctrl> and <V> together. Alternatively, choose Paste from the
Edit menu.
Another way to import backgrounds is to use the Edit menu in ProModel:
1. Choose Background Graphics from the Build menu.
2. Select Front of Grid or Behind Grid. Generally, most graphics should
be laid out in front of the grid. Large objects such as walls or imported
backgrounds are usually placed behind the grid. Front of grid means the
graphic will not be covered by the grid lines when the grid is on. Behind
grid means the graphic will be covered with grid lines when the grid
is on.
3. Choose Import Graphic from the Edit menu.
4. Select the desired file and file type. The image will be imported into the
layout of the model.
5. Left-click on the graphic to reposition and resize it, if necessary.
It is often a good idea to lock a background graphic in place to avoid uninten-
tionally moving it. To lock a graphic, right-click on the graphic and select Lock
from the drop-down menu. The graphic can be unlocked any time by selecting
Unlock from the same menu.
Open the model developed in section L6.9 (Raja’s Manufacturing Cell).
From the Build menu choose Background Graphics Behind Grid. From the Edit
menu choose Import Graphic. Search for and select the background graphic file
Tutorialback.WMF (Figure L6.46).
L6.11 Defining and Displaying Views

Specific areas of the model layout can be predefined and then quickly and easily
viewed in ProModel. Each view can be given a unique name and can have a suit-
able magnification. These views can be accessed during editing of the model by
selecting Views from the View menu or by using the keyboard shortcut key. The
Views dialog box is shown in Figure L6.47. Views can also be accessed in real
time during the running of the model with a VIEW “Mill_Machine” statement.
Problem Statement
For the Shipping Boxes Unlimited example in Lab 4, section L4.8.2, define the
following six views: Monitor Queue, Box Queue, Inspect Queue, Shipping Queue,
Pallet Queue, and Full View. Each view must be at 300 percent magnification.
Show Full View at start-up. Show the Pallet Queue when an empty pallet arrives at
the pallet queue location. Go back to showing the Full View (Figure L6.48) when
the box is at the shipping dock. Here are the steps in defining and naming views:
1. Select the View menu after the model layout is finished.
2. Select Views from the View menu.
har01307_LB06_475-516.indd 503 20/12/10 1:54 PM

504 Part II Labs
FIGURE L6.46
Importing a background for the Raja’s Manufacturing Cell.
Referencing a View in Model Logic

1. Select General Information (Figure L6.49) from the Build menu. Select
Initialization Logic and use the VIEW statement:
View “Full View”
2. Select Processing from the Build menu. Use the following statement in
the Operation field when the Pallet Empty arrives at the Pallet Queue
location (Figure L6.50).
View “Pallet Queue”
Also, use the following statement in the Operation field when the box arrives
at the shipping dock:
View “Full View”
har01307_LB06_475-516.indd 504 20/12/10 1:54 PM

FIGURE L6.47
The Views dialog box.
FIGURE L6.48
Full View of the Shipping Boxes Inc. model.
har01307_LB06_475-516.indd 505 20/12/10 1:54 PM

506 Part II Labs
FIGURE L6.49
General Information
for the Shipping Boxes
Inc. model.
FIGURE L6.50
Processing and routing tables for Shipping Boxes Inc. incorporating the change of views.
L6.12 Creating a Model Package

Promodel has an innovative feature of creating a package of all files associated
with the model file. This package file is titled <model name> .pkg—for example,
ATM.pkg. The package file can then be archived or distributed to others. This
file includes the model file (*.MOD), the graphic library (unless you check the
Exclude Graphics Library option), and any external files you defined (such as
har01307_LB06_475-516.indd 506 20/12/10 1:54 PM

read files, arrival files, and shift files). The model package automatically includes
bitmaps imported into the background graphics. The package file can be subse-
quently unpackaged by the receiver to run.
Use the following steps to create a model package:
1. Select Create Model Package from the File menu.
2. Enter the name you wish to use for the model package. ProModel uses
the name of the model file as the default name of the package. So ATM.
mod will be packaged as ATM.pkg. You can also Browse . . . to select the
model name and directory path.
3. Check/uncheck the Exclude Graphics Library box, if you want to
include/exclude the graphics library.
4. Check the Protect Model Data box if you want to protect the data in
your model and prevent other users from changing and even viewing the
model data files. When you load a model package, ProModel disables
the View Text, Print Text, and Expression Search features, plus the Build
menu and portions of the run-time Options and Information menus.
5. Click OK.
Problem Statement
For the Discs-R-Us Manufacturing Inc. in section L6.5, make a model pack-
age that includes the model file and the shift file for thelocations. Name the
model package L6_12ShiftWorkSchedule.pkg (Figure L6.51) and save it on a
USB flash drive.
To unpack and install the model package, double-click on the package file
(on the USB drive). In the Unpack Model Package dialog select the appropriate
destination drive and directory path to install the model file and its associated
FIGURE L6.51
Create Model Package
dialog.
har01307_LB06_475-516.indd 507 20/12/10 1:54 PM

508 Part II Labs
FIGURE L6.52
Unpack and install
Model package.
files (Figure L6.52). We chose the USB drive ( e:\ ) to install the model file, the
associated shift file, and the graphics library. Click OK. After completing the in-
stallation, ProModel prompts you for loading the model. Click Yes.
L6.13 Exercises
1. Five different types of equipment are available for processing a
special type of part for one day (six hours) of each week. Equipment 1
is available on Monday, equipment 2 on Tuesday, and so forth. The
processing time data is shown in Table L6.7.
Assume that parts arrive at a rate of one every 4 1 hours,
including weekends. How many parts are produced each week? How
large a storage area is needed for parts waiting for a machine? Is there a
bottleneck at any particular time? Why?
TABLE L6.7 Processing Time data
Time to Process
Equipment Available Day One Part (minutes)
1 Monday 52
2 Tuesday 42
3 Wednesday 3 1.5
4 Thursday 61
5 Friday 51
2. Customers visit the neighborhood barbershop Fantastic Dan for a

haircut. Among the customers there are 30 percent children, 50 percent
women, and 20 percent men. The customer interarrival time is
triangularly distributed with a minimum, mode, and maximum of 8, 11,
and 14 minutes, respectively. The haircut time (in minutes) depends on
the type of customer,as shown in Table L6.8.
har01307_LB06_475-516.indd 508 20/12/10 1:54 PM

TABLE L6.8 Haircut Time Data for Various Types of Customers
Haircut Time(min)
Customers Mean Half-width
Children 8 2
Women 12 3
Men 10 2
The initial greetings and signing in take Normal(2,0.2) minutes, and

the transaction of money at the end of the haircut takes Normal(3,0.3)
minutes. Run the simulation model for 2,000 hours.
a. About how many customers of each type does Dan process per day?
b. What is the average number of customers of each type waiting to get
a haircut? What is the maximum?
c. What is the average time spent by a customer of each type in the
salon? What is the maximum?
3. The Atithi Restaurant example in Lab 2, section L2.7 is a 24/7 facility.
Define an attribute “time_in” to track the customer’s time of arrival
to the restaurant and a log function to collect the statistics on the time
customers spend in the restaurant (waiting, being served, and eating).
Also, define a global variable “WIP” to keep track of the number of
customers in the system. Run the simulation model for a whole year.
Change the customer arrival rates per hour, from two to 12 in increments
of one and determine the effects of increasing traffic intensity () on,
a. average and maximum number of customers in the drive-through
facility of the restaurant.
b. average and maximum time a customer spends in the drive-through
facility.
c. server utilization.
What happens when the traffic intensity () exceeds 0.8 (80%)? Plot the
contents of the drive-in facility and note what happens when the traffic
intensity approaches values larger than 0.8.
4. Poly Forgings Inc. receives forgings from its suppliers in batches of
one every eleven minutes exponentially distributed. All forgings arrive
at the raw material store. Of these forgings, 70 percent are used to make
gear A, and the rest is used to make gear B. Gear A goes from the raw
material store to the mill, and then on to the grinder. Gear B goes directly
to the grinder. After grinding, all gears go to degrease for cleaning.
Finally, all gears are sent to the finished parts store. The processing time
data is shown in Table L6.9. Simulate for 2,000 hours.
Track the work-in-process inventory of both types of gears separately.
Also, track the production of finished gears (both types).
har01307_LB06_475-516.indd 509 20/12/10 1:54 PM

510 Part II Labs
TABLE L6.9 Processing Time Data for Gears
Gear Process 1 Process 2 Process 3
Gear A Mill [N(5,2) min.] Grinder [U(11,1) min.] Degrease [7 min.]

Gear B Grinder [U(9,1) min.] Degrease [7 min.]
5. Considerthe Salt Lake CityElectronics Exercise 10, section L4.13, with

the following enhancements. The soldering machine breaks down on
average after every 2,000 200 minutes of operation. The repair time is
normally distributed (100, 50) minutes. Create a simulation model of this
system. Simulate this manufacturing system for 2,000 hours. Analyze
the utilization of each station, associated queues, and the total number of
jobs manufactured during each eight-hour shift (average).
6. Consider the SoCal Casting Inc. example in Lab 4, section L4.5, and
answer the following questions:
a. What is the average and maximum time a casting spends in the system?
b. What is the average and maximum time a casting waits before being
loaded on a mill?
c. What is the average and maximum work-in-process of castings in the
system?
Embellishment: After a mill processes 25 castings, it is shut down for a
uniformly distributed time between 10 and 20 minutes for cleaning and
tool change.
d. What percentage of the time does each mill spend in cleaning and tool
change operations?
e. What is the average and maximum time a casting spends in the
system?
f. What is the average and maximum time a casting waits before being
loaded on a mill?
g. What is the average and maximum work-in-process of castings in the
system?
7. Consider the NoWaitBurger stand in Exercise 13, section L4.13, and
answer the following questions:
a. What is the average amount of time spent by a customer at the
hamburger stand?
b. Run 10 replications and compute a 90-percent confidence interval on
the average amount of time spent by a customer at the stand.
c. Develop a 90-percent confidence interval for the average number of
customers waiting at the burger stand.
8. Ram, Shyam, and Gopal wash cars at the Bollywood Car Wash. Cars
arrive every 10 6 minutes. They service customers at the rate of one
every 20 10 minutes. However, the customers prefer Ram to Shyam,
and Shyam to Gopal. If the attendant of choice is busy, the customers
har01307_LB06_475-516.indd 510 20/12/10 1:54 PM

choose the first available attendant. Simulate the car wash system for
1,000 service completions (car washes). Answer the following questions:
a. Estimate Ram’s, Shyam’s, and Gopal’s utilization.
b. On average, how long does a customer spend at the car wash?
c. What is the longest time any customer spent at the car wash?
d. What is the average number of customers at the car wash?
Embellishments:
The customers are forced to choose the first available attendant; no
individual preferences are allowed. Will this make a significant enough
difference in the performance of the system to justify this change?
Answer questions a through d to support your argument. Use the Paired-t
confidence interval approach (Lab 9, section L9.4) to compare the two
systems.
9. Amrita is a pharmacist and works at the Save-Here Drugstore. Walk-in
and drive-in customers arrive at a rate of 10 3 and 20 10 minutes
respectively. Drive-in customers are given higher priority than walk-in
customers. The number of items in a prescription varies uniformly from
1 to 5 [U(3,2)]. Amrita can fill one item in 6 1 minutes (uniform
distribution). She works from 8 A.M. until 5 P.M. Monday to Friday. Her
lunch break is from 12 noon until 1 P.M. She also takes two 15-minute
breaks: at 10 A.M. and at 3 P.M. Define a shift file for Amrita named
Amrita.sft. Model the pharmacy for a year and answer the following
questions:
a. Estimate the average time a customer (of each type) spends at the
drugstore.
b. What is the average number of customers (of each type) waiting for
service at the drug store?
c. What is the average number of customers (of each type) in the drug store?
d. What is the utilization of Amrita (percentage of time busy)?
e. Do you suggest that we add another pharmacist to assist Amrita? How
many pharmacists should we add?
f. Is it better to have a dedicated pharmacist for drive-in customers and
another for walk-in customers?
g. Create a package file called SaveHere.pkg that includes the
simulation model file and the shift file Amrita.sft.
10. A production line assembles a product on five workstations in sequence.
Parts arrive to the first workstation at random exponentially distributed
intervals with a mean of seven minutes. Service times at individual
stations are normally distributed (mean of 3 minutes and standard
deviation of 1 minute).
a. Develop a simulation model that collects data for time in the system
for each part. Determine the mean time in the system.
b. Additional analysis shows that products from three different priority
classes are assembled on the line and they appear in random order.
har01307_LB06_475-516.indd 511 20/12/10 1:54 PM

512 Part II Labs
Products from classes 1, 2, and 3 have highest, medium, and lowest

priorities, respectively. The part mix for priority classes 1, 2, and 3
is 30 percent, 50 percent, and 20 percent, respectively. Modify the
simulation model to reflect this situation. Determine the mean time in
the system for products from the three different priority classes.
c. The assembly process is modified so that each product to be
assembled is placed on a pallet. The production supervisor wants to
know how many pallets to place in operation at one time. It takes
Uniform(8,2) minutes to move the pallet from station 5 to station 1.
Pallet loading and unloading times are Uniform(3,1) minutes each.
Modify the simulation model to reflect this situation. Is it better to use
5 or 20 pallets?
11. Two types of parts arrive at a single machine where they are
processed one part at a time. One type of part has a higher priority
over the other. Both types of parts arrive exponentially with a mean
interarrival time of 30 minutes. Within priority classes, the parts
are serviced in FIFO order, but between priority classes the higher-
priority type will always be processed first. However, a lower-priority
part cannot be interrupted once it has begun processing. The machine
processing times for higher- and lower-priority types of parts are
Uniform(20,5) and Uniform(8,2) minutes, respectively. Simulate the
system for 2,000 hours.
a. What are the average time in queue and average time in system for
each type of part?
b. What are the average production rates of each type of part per hour?
12. The manufacture of a certain line of composite aerospace subassembly
involves a relatively lengthy assembly process, followed by a short
firing time in an oven, which holds only one subassembly at a time. An
assembler cannot begin assembling a new subassembly until he or she
has removed the old one from the oven. The following is the pattern of
processes followed by each assembler:
a. Assemble next subassembly.
b. Wait, first-come, first-served, to use the oven.
c. Use the oven.
d. Return to step a.
The operating times and relevant financial data are shown in
Tables L6.10 and L6.11, respectively.
TABLE L6.10 Operating Times Data
Operation Time Required (minutes)
Assemble U(30,5)
Fire U(8,2)
har01307_LB06_475-516.indd 512 20/12/10 1:54 PM

TABLE L6.11 Cost Data
Item Cost Information ($)
Assembler’s salary $35/hour

Oven cost $180 per 8-hour workday (independent of utilization)
Raw material $8 per subassembly
Sale price of finished subassembly $40 per subassembly
Build a simulation model of this manufacturing process. Use the model

to determine the optimal number of assemblers to be assigned to an
oven. The optimal number is understood in this context to be the one
maximizing profit. Base the determination on simulation equivalent
to 2,000 hours of simulated time. Assume there are no discontinuities
within a working day, or in moving between consecutive eight-hour
working days. (Adapted from T. Schriber, Simulation using GPSS, John
Wiley, 1974.)
13. At the downtown office of the California Commercial Real Estate, loan
applications arrive with interarrival times that are exponentially distributed
with a mean of 50 minutes. Loan application processing times are normally
distributed with an average time of 45 minutes and standard deviation of
10 minutes. All loan applications are audited after initial processing that
takes somewhere between 15 to 25 minutes, uniformly distributed. After
audit, 15 percent of the loans are found to require additional research and
processing work. The rework times are exponentially distributed with a
mean of 40 minutes. The rework is processed with a higher priority over
new loan applications. Simulate the loan processing office for 2,000 hours
of operation and respond to the following:
a. How long is a loan application in the system—average, maximum?
b. How long is a reworked loan application in the system—average,
maximum?
c. Develop a histogram of the mean time in the system for all the loan
applications.
d. Modify the simulation model so that a loan application is rejected if it
fails a second audit. What percentage of loan applications is rejected?
e. Modify the simulation model to count the number of loan applications
staying in the office longer than a specified length of time. What
percentage of loan applications remains in the office for more than
90 minutes?
f. Modify the simulation model to reflect that it takes uniform (10 ± 5)
minutes to move the loan applications from one station to another.
g. Modify the original simulation model so that loan applications that
are reworked for second or more times have a higher priority over
applications needing rework for the first time. Develop a histogram of
har01307_LB06_475-516.indd 513 20/12/10 1:54 PM

514 Part II Labs
the mean time in the system for all the loan applications. Does it seem
to make a difference to the mean time in the system for all the loan
applications?
14. Two types of jobs (regular and high-priority) arrive at two identical
assembly machines. Regular jobs arrive according to an exponential
distribution with a mean interarrival time of 10 minutes. High-priority jobs
arrive according to an exponential distribution with a mean interarrival
time of 20 minutes. One of the two machines is selected by turn.
Processing on either machine is uniformly distributed U(8,2) minutes.
Upon completion, jobs are sent to a third machine where they wait for
final assembly, which requires 5 minutes, normally distributed with a
standard deviation of 2 minutes. Jobs are selected from the waiting line
for assembly based on the estimated assembly time (jobs with shorter
assembly times having higher priority over jobs with longer assembly
times). Completed jobs are promptly shipped to the customer.
The final assembly machine is unreliable. It fails at intervals that may
be described by an exponential distribution with a mean of 60 minutes.
Repair time is uniformly distributed over the range 10 to 20 minutes.
The high-priority jobs have higher priority over regular jobs in terms of
processing on the three assembly machines. The processing times are
identical for both types of jobs. Simulate for 2,000 hours and determine:
a. Utilization of the three machines.
b. Average time in the shop for both types of jobs.
c. Average number of jobs (regular and high-priority) processed per
hour.
15. A production facility consists of four operations: milling, planing,
drilling, and inspection. There are two mills, a planer, two drills, and
one inspection station.Movement times associated with transfer of
parts between any pair of workstations and from workstations to the
scrap yard are triangularly distributed. The data for arrival, routing, and
processing are shown in Table L6.12.
Seventy percent of the jobs are type I, the others are type II. Type I
jobs arrive in batches of 5. Type II jobs arrive in batches of 10. A whole
batch of jobs is processed and moved together. Type I jobs are served
TABLE L6.12 Arrival, Routing, and Processing Time Data
Arrivals Exponential Mean = 4 minutes
Milling Triangular (9,12,14) minutes

Planing Uniform (7,1) minutes
Drilling Normal ( 5,1) minutes
Inspection Normal (2,.5) minutes
Movement Time Triangular (0.5,1.2,1.5) minutes
har01307_LB06_475-516.indd 514 20/12/10 1:54 PM

at milling then moved to drilling and finally to inspection. Type II jobs

go to planing then drilling, then inspection. There is no priority system
at drilling and inspection. Jobs fail inspection with probability of 0.05
(or 5%). Sixty percent of the failed jobs are returned to drilling, and the
remainder is sent to the scrap yard.
There are finite buffers in front of these machines. A buffer in front
of the mills, planer, drills, and inspection will hold 10 parts each. When
jobs cannot join a queue, because it is full, they will wait at the previous
process.
Simulate this system for 2,000 hours with 8 hours warm-up time and
determine:
a. Which is the bottleneck station?
b. Average time in system for Type I and Type II jobs.
c. Average production rate per hour for Type I and Type II jobs.
d. Average time in system for Type I and Type II jobs.
16. The machines (mills, planer, drills, and inspection) in the production
facility described in Exercise 15, breaks down on average after every
30 5 hours of operation. The repair time is normally distributed
(120,30) minutes. Create a simulation model of this system.
Simulate this system for 2000 hours with 8 hours warm-up time and
determine:
a. Which is the bottleneck station?
b. Average time in system for Type I and Type II jobs.
c. Average production rate per hour for Type I and Type II jobs.
d. Average time in system for Type I and Type II jobs.
har01307_LB06_475-516.indd 515 20/12/10 1:54 PM

har01307_LB06_475-516.indd 516 20/12/10 1:54 PM
L A B
7 MODEL VERIFICATION
AND VALIDATION
Dew knot trussed yore spell chequer two fined awl yore mistakes.
—Brendan Hills
In this lab we describe the verification and validation phases in the development
and analysis of simulation models. In section L7.1 we describe an inspection and
rework model. In section L7.2 we show how to verify the model by tracing the
events in it. Section L7.3 shows how to debug a model. The ProModel logic, basic
and advanced debugger options are also discussed.
L7.1 Verification of an Inspection and Rework Model

Problem Statement
Bombay Clothing Mill is a textile manufacturer, importer, and reseller. Containers
of garments arrive (three per hour, exponentially distributed) from an overseas sup-
plier to the mill warehouse in Bombay for identification and labeling, inspection,
and packaging (Figure L7.1). The identification and labeling take U(10,2), inspec-
tion takes N(5,1), and packaging takes N(4,0.25) minutes. At times the labeling
machine malfunctions. The labeling mistakes are detected during inspection. Sixty
percent of the garments are labeled properly and go on to the packaging area; how-
ever, 40 percent of the garments are sent back for rework (relabeling). Garments
sent back for rework have a higher priority over regular garments at the labeling
machine. Movement between all areas takes one minute. Simulate for 100 hours.
Priorities determine which waiting entity or downtime is permitted to access
a location or resource when the location or resource becomes available. Priorities
may be any value between 0 and 999 with higher values having higher priority.
For simple prioritizing, you should use priorities from 0 to 99.
517
har01307_LB07_517-526.indd 517 20/12/10 1:55 PM

518 Part II Labs
FIGURE L7.1
A snapshot of the
Bombay Clothing Mill
showing status lights
for the locations.
FIGURE L7.2
Locations at the Bombay Clothing Mill warehouse.
FIGURE L7.3
Processing and routing tables at the Bombay Clothing Mill.
Six locations (Ship_In, Garment_Q, Relabel_Q, Labeling, Inspection, and

Pack_Ship) are defined as shown in Figure L7.2. Two entities (Garments and
Relabel) are defined next. The processing and routing tables are shown in Fig-
ure L7.3. Note that the garments sent back for relabeling have a higher priority of
2 to access the labeling machine location as opposed to a lower priority of 1 for
incoming garments. Also, note the multiple routings created for garments from
the Inspection location. Sixty percent of the garments are labeled correctly and
go to the Pack_Ship location while 40 percent of the garments that are incorrectly
har01307_LB07_517-526.indd 518 20/12/10 1:55 PM

labeled go back to the Relabel_Q. This type of branching was first introduced in
Lab 4, section L4.6.
We also used status lights for the various locations in this model (Figure L7.1).
A status light is a circle that lights up with different colors depending on the status
of the location. You can place a status light anywhere relative to a location for
showing the status or current state of the location. At run time, you can display
a window showing what status each color represents. For a single-capacity loca-
tion, the states displayed are idle/empty (deep blue), in operation (green), blocked
(pink), down (red), waiting (yellow), and in setup (light blue). For multicapacity
locations, the displayed states are up–operational (blank) and down–off-shift, on
break, disabled (red).
L7.2 Verification by Tracing the Simulation Model

Now we will run the Bombay Clothing Mill model (section L7.1) with the Trace
feature turned on. Trace will list events as they happen during a simulation run. A
trace is a list of events that occur during a simulation. For example, a line in the
trace could read, “EntA arrives at Loc1, Downtime for Res1 begins.” A trace list-
ing also displays assignments, such as variable and array element assignments. A
trace listing of a simulation run may be viewed in several ways and is generated
in one of two modes, Step or Continuous.
• Off: Select this option to discontinue a trace.
• Step: Choose this option to step through the trace listing one event at a
time. Each time you click the left mouse button, the simulation will
advance one event.
• Continuous: Choose this option to write the trace continuously to the
output device selected from the Trace Output submenu. This is useful
when you do not know exactly where to begin or end the trace.
The user can launch the Trace feature TRACE statement within the model code
or from the Options menu. Run the simulation model. Choose Trace Options from
the Options menu (Figure L7.4). You can choose either the Trace Step or the Trace
Continuous mode for running Trace. Choose the Step mode for now. The trace
output can be sent to the Window or to a file. Let’s choose to send the output to
the Window (bottom of the Window) for now.
FIGURE L7.4
Trace modes in
ProModel.
har01307_LB07_517-526.indd 519 20/12/10 1:55 PM

520 Part II Labs
FIGURE L7.5
Tracing the simulation model of the Bombay Clothing Mill warehouse.
The trace prompts are displayed at the bottom of the screen as shown in
Figure L7.5. Follow the trace prompts and verify that the model is doing what it is
supposed to be doing. Make sure through visualization and the trace messages that
the rework items are indeed given higher priority than the first-time items. Also,
from the garment and relabel queue contents shown in Figure L7.6, notice that the
relabel queue is always quite small, whereas the garment queue grows quite big at
times (as a result of it lower priority for the labeling operation).
L7.3 Debugging the Simulation Model

The debugger is a convenient, efficient way to test or follow the processing of any
logic in your model. The debugger is used to step through the logic one statement
at a time and examine variables and attributes while a model is running. All local
variables, current entity attributes, and so forth are displayed, and the user has full
control over what logic gets tracked, even down to a specific entity at a specific
location (Figure L7.7). The user can launch the debugger using a DEBUG statement
har01307_LB07_517-526.indd 520 20/12/10 1:55 PM

FIGURE L7.6
Plots of garment and relabel queue contents averaged by hour.
FIGURE L7.7
The ProModel
Debugger menu.
within the model code or from the Options menu (Figure L7.4) during run time.
The system state can be monitored to see exactly when and why things occur.
Combined with the Trace window, which shows the events that are being sched-
uled and executed, the debugger enables a modeler to track down logic errors or
bugs in the model.
har01307_LB07_517-526.indd 521 20/12/10 1:55 PM

522 Part II Labs
FIGURE L7.8
Debugger options
menu.
Global changes and user specified conditions can be tracked using the
Debugger options menu (Figure L7.8). The user conditions can be checked—
before each statement, at each thread switch, or at each thread initiation.
L7.3.1 Debugging ProModel Logic

Before we discuss the details of the debugger, it is important to understand the
following terms:
• Statement: A statement in ProModel performs some operation or takes
some action – for example, JOIN, WAIT, USE, MOVE, and so forth.
• Logic: Logic is the complete set of statements defined for a particular
process record, downtime event, initialization, or termination of the
simulation, and so forth.
• Thread: A thread is a specific execution of any logic. A thread is initiated
whenever logic needs to be executed. This can be any entity running
through operation logic, the initialization logic, a resource running some
node logic, downtime logic, or any other logic. Note that the same logic
may be running in several threads at the same time. For example, three
entities of the same type being processed simultaneously at the same
multicapacity location would constitute three threads.
Even though several threads can execute the same logic at the same time in the
simulation, the simulation processor can process them only one at a time. So there
is really only one current thread, while all other threads are suspended (either
scheduled for some future simulation time or waiting to be executed after the cur-
rent thread at the same simulation instant).
Every time logic is executed, a thread that is assigned a unique number ex-
ecutes it. THREADNUM returns the number of the thread that called the function.
This function can be used with the IF-THEN and DEBUG statements to bring up the
debugger at a certain process.
har01307_LB07_517-526.indd 522 20/12/10 1:55 PM

For example, for the ElSegundo Composites, from Lab 5, section L5.6.1,
suppose the GROUP quantity has been typed as 50 (typo) instead of 5. A debug
statement IF THREADNUM () > 30 THEN DEBUG has been added to the operation
logic at the Oven location that causes the error, as shown in Figure L7.9. The
simulation will run until the proper process and then bring up the debugger (Fig-
ure L7.10). The debugger can then be used to step through the process to find out
the particular statement causing the error.
L7.3.2 Basic Debugger Options

The debugger can be used in two modes: Basic and Advanced. The Basic Debug-
ger appears initially, with the option of using the Advanced Debugger. The Basic
Debugger (Figure L7.10) has the following options:
FIGURE L7.9
An example of a DEBUG statement in the processing logic.
FIGURE L7.10
The Debugger window.
har01307_LB07_517-526.indd 523 20/12/10 1:55 PM

524 Part II Labs
• Error display box: Displays the error message or reason why the
debugger is displayed, such as the user condition becoming true.
• Logic display box: Displays the statements of the logic being executed.
• Information box: Displays either the context of the logic or local
information.
• Context: Displays the module, operation, and line number in which the
debugger stopped.
• Local information: Displays local variables and entity attributes with
nonzero values for the thread in the information box.
• End Simulation: Choose this option to terminate simulation. This will
prompt for whether or not you would like to collect statistics.
• Run: Continues the simulation, but still checks the debugger options
selected in the Debugger Options dialog box.
• Next Statement: Jumps to the next statement in the current logic. Note
that if the last statement executed suspends the thread (for example, if the
entity is waiting to capture a resource), another thread that also meets
the debugger conditions may be displayed as the next statement.
• Next Thread: Brings up the debugger at the next thread that is initiated or
resumed.
• Into Subroutine: Steps to the first statement in the next subroutine
executed by this thread. Again, if the last statement executed suspends
the thread, another thread that also meets debugger conditions may be
displayed first. If no subroutine is found in the current thread, a message
is displayed in the Error Display box.
• Options: Brings up the Debugger Options dialog box. You may also bring
up this dialog box from the Simulation menu.
• Advanced: Changes the debugger to the advanced mode.
L7.3.3 Advanced Debugger Options

The Advanced Debugger (Figure L7.11) contains all the options in the Basic
Debugger plus a few advanced features:
Thread
• Next: Jumps to the next thread that is initiated or resumed. This button has
the same functionality as the Next Thread button in the Basic Debugger.
• New: Jumps to the next thread that is initiated.
• Disable: Temporarily disables the debugger for the current thread.
• Exclusive: The Debugger displays the statements executed.
Logic
• Next: Jumps to the next initiated or resumed thread that is not executing
the same logic as the current thread.
har01307_LB07_517-526.indd 524 20/12/10 1:55 PM

FIGURE L7.11
The ProModel
Advanced Debugger
options.
• New: Jumps over any resumed threads to the next initiated thread that is
not executing the same logic as the current thread. This will automatically
jump to a new thread.
• Disable: Temporarily disables the debugger for all threads executing the
current logic.
• Exclusive: The debugger displays only the statements executed in any
thread that are an instance of the current logic.
Enable disabled threads and logic: Enables the threads and logics that were
disabled previously.
For more information and detailed examples of how to use the various debugger
options, consult the online ProModel Help System.
L7.4 Exercises
1. For the example in section L7.1, insert a DEBUG statement when a garment
is sent back for rework. Verify that the simulation model is actually sending
back garments for rework to the location named Label_Q.
2. For the example in section L4.1 (Pomona Electronics), trace the model
to verify that the type B circuit boards are indeed following the routing
given in Table L4.1.
har01307_LB07_517-526.indd 525 20/12/10 1:55 PM

526 Part II Labs
3. For the example in section L4.5 (SoCal Casting Inc.), run the simulation
model and launch the debugger from the Options menu. Turn on the
Local Information in the Basic Debugger. Verify the values of the
variables WIP and PROD_QTY. Use status lights for the Mill, Grinder,
and the Finish Parts Store to verify the operating, idle, and blocked status
for these locations.
4. For the example in section L4.4 (Bank of India), trace the model to
verify that successive customers are in fact being served by the three
tellers in turn.
5. For the example in section L4.8.1 (Shipping Boxes Unlimited), trace
the model to verify that monitors are indeed being packed (joined) into
boxes at the Packing_Table location.
the model to verify that the full boxes are in fact being loaded on empty
pallets at the Inspector location and are being unloaded at the Shipping
location.
7. For the example in section L4.11.1 (Bombay Restaurant), trace the
model to verify that customers are in fact leaving and going away when
there are six or more cars at the restaurant. Use status lights for all the
single capacity locations to verify the operating and idle status for these
locations.
the model to verify that the inspector is indeed only inspecting when at
least five or more full boxes are waiting in the Inspect_Q for inspection.
har01307_LB07_517-526.indd 526 20/12/10 1:55 PM

L A B
8 SIMULATION OUTPUT
ANALYSIS
Nothing has such power to broaden the mind as the ability to investigate
systematically and truly all that comes under thy observation in life.
—Marcus Aurelius
In this lab we review the differences between terminating and nonterminating

simulations. You will learn how to collect observations of a performance met-
ric from both terminating and nonterminating simulations using ProModel. Key
problems of identifying the warm-up period and selecting the run length of a non-
terminating simulation are addressed in the lab. We also illustrate how supporting
software can be used to help determine if the observations reported by ProModel
for a simulation are independent and normally distributed.
L8.1 Terminating versus Nonterminating Simulations

A terminating simulation has a fixed starting condition and a fixed ending condi-
tion (terminating point). For example, a simulation model of a store that opens
at 9:00 A.M. begins with a starting condition of no customers in the system and
terminates after the store closes at 7:00 P.M. and the last customer is processed.
Many service-oriented systems (banks, doctors’ offices, post offices, schools,
amusement parks, restaurants, and department stores) are analyzed with a termi-
nating simulation. The system returns to a fixed starting condition after reaching
a natural terminating event.
A nonterminating simulation has neither a fixed starting condition nor a
natural ending condition. The simulation cannot run indefinitely, so we end it after
its output reflects the long-term average behavior of the system. A hospital or a 24-
hour gas station could be analyzed with a nonterminating simulation depending
on the objective of the study. Many systems appear to be terminating but in reality
527
har01307_LB08_527-548.indd 527 20/12/10 1:55 PM

528 Part II Labs
are nonterminating. Systems that close/terminate but do not go back to a fixed

starting condition (empty for example) are usually analyzed with nonterminating
simulations. For example, a production system that uses the current day’s unfinished
work as the next day’s starting work is really a nonterminating system. If we com-
press time by removing the intervals during which the system is closed, it behaves
as a continuously operating system with no point at which the system starts anew.
In general, most manufacturing systems are nonterminating. The work-in-process
inventory from one day is typically carried forward to the next day.
When we report the simulation model’s response based on the output from
a terminating simulation or a nonterminating simulation, it is important to use
interval estimates rather than simple point estimates of the parameter being mea-
sured. The advantage of the interval estimate is that it provides a measure of the
variability in output produced by the simulation. Interval estimates or confidence
intervals provide a range of values around the point estimate that have a speci-
fied likelihood of containing the parameter’s true but unknown value. For a given
confidence level, a smaller confidence interval is considered better than a larger
one. Similarly, for a given confidence interval, a higher confidence level indicates
a better estimate. To construct the confidence interval, the observations produced
by replications or batch intervals need to be independent and identically distrib-
uted. Furthermore, the observations should be normally distributed when using
the Student’s t distribution to calculate the confidence interval. We can fudge a
little on the normality assumption but not on the independence and identically
distribution requirements. These topics are elaborated on in Chapter 8.
L8.2 Terminating Simulation

Problem Statement
Spuds-n-More by the Seashore is a favorite among vacationers at the seaside
resort. Spuds-n-More, a takeout fast-food restaurant, specializes in potato foods,
serving many different varieties of french fries, waffle fries, and potato logs along
with hamburgers and drinks. Customers walking along the boardwalk access the
restaurant via a single ordering window, where they place their orders, pay, and
receive their food. There is no seating in the restaurant. The interarrival time of
customers to Spuds-n-More is exponentially distributed with a mean of six min-
utes. The activity times are given in Table L8.1. The waiting area in front of
the walkup window can accommodate only five customers. If the waiting area is
TABLE L8.1 Process Times at Spuds-n-More by the Seashore
Activity Process Time (minutes)
Order food E(2)

Make payment E(1.5)
Pick up food E(5)
har01307_LB08_527-548.indd 528 20/12/10 1:55 PM

filled to capacity when a customer arrives, the customer bypasses Spuds-n-More.

Mr. Taylor closes the entryway to the waiting area in front of his restaurant after
eight hours of operation. However, any customers already in the waiting area by
the closing time are served before the kitchen shuts down for the day.
Mr. Taylor is thinking about expanding into the small newspaper stand next
to Spuds-n-More that has been abandoned. This would allow him to add a second
window to serve his customers. The first window would be used for taking orders
and collecting payments, and the second window would be used for filling orders.
Before investing in the new space, however, Mr. Taylor needs to convince the
major stockholder of Spuds-n-More, Pritchard Enterprises, that the investment
would allow the restaurant to serve more customers per day. The first step is to
build a baseline (as-is) model of the current restaurant configuration and then
to embellish the model with the new window and measure the effect it has on
the number of customers processed per day. Mr. Taylor is interested in seeing a
90 percent confidence interval for the number of customers processed per day in
the baseline simulation.
L8.2.1 Starting and Terminating Conditions (Run Length)

Build the baseline model of the as-is restaurant and conduct a terminating simula-
tion to estimate the expected number of customers served each day. The starting
condition for the simulation is that the system begins in the morning empty of cus-
tomers with the first customer arriving at time zero. Customers are admitted into
the restaurant’s queue over the eight hours of operation. The simulation termi-
nates after reaching eight hours and having processed the last customer admitted
into the queue. The layout of the simulation model is shown in Figure L8.1, and a
printout of the model is given in Figure L8.2. Note that the length of the Order_Q
location is 25 feet and the Customer entity travel speed is 150 feet per minute.
These values affect the entity’s travel time in the queue (time for a customer to
walk to the end of the queue) and are required to match the results presented in
this lab. The model is found in the Student folder installed with ProModel under
file name Lab 8_2_1 Spuds-n-More1.MOD.
Let’s take a look at the model listing in Figure L8.2 to see how the termina-
tion criterion was implemented. The Entry location serves as a gate for allowing
customers into the queue until the simulation clock time reaches eight hours. Enti-
ties arriving after the simulation clock reaches eight hours are turned away (routed
to Exit) by the routing rule User Condition (IF clock(hr) > 8) at the Entry
location. Entities arriving within eight hours are routed to the Order_Q location
(see routing rule User Condition IF clock(hr) < 8). Two global variables are
declared to count the number of entities arriving at the Order_Q location and the
number of entities processed at the Order_Clerk location. See the statement Inc
Arrivals at the Order_Q location and the statement Inc Processed at the Order_
Clerk location. The Inc statement increments the variable’s value by one each
time it is executed. As the entities that arrive after eight hours of simulation time
are turned away to the Exit, the number of entities processed by the Order_Clerk
har01307_LB08_527-548.indd 529 20/12/10 1:55 PM

530 Part II Labs
FIGURE L8.1
Layout of the Spuds-
n-More simulation
model.
is compared to the number of entities allowed into the Order_Q (see Move Logic).
If Processed Arrivals, then all entities that were allowed into the Order_Q have
been processed and the simulation is terminated by the Stop statement. Notice that
the combined capacity of the Entry and Order_Q locations is five in order to sat-
isfy the requirement that the waiting area in front of the restaurant accommodates
up to five customers.
L8.2.2 Replications
Run the simulation for five replications to record the number of customers served
each day for five successive days. To run the five replications, select Options from
under the Simulation main menu. Figure L8.3 illustrates the simulation options
set to run five replications of the simulation. Notice that no run hours are speci-
fied. After ProModel runs the five replications, it displays a message asking if you
wish to see the results. Answer yes.
In the Output Viewer, select Variable Summary from the Tables menu (Fig-
ure L8.4). From the Filter panel on the left, select <All> for Replications and
for Statistics select Average, Standard Deviation and 90 Percent Confidence
Interval to specify that you wish the output report to also include the sample
mean (average), sample standard deviation, and 90 percent confidence interval
of the five observations collected via the five replications (Figure L8.5).
har01307_LB08_527-548.indd 530 20/12/10 1:55 PM

FIGURE L8.2
The simulation model of Spuds-n-More.
The ProModel output report for variables is shown in Figure L8.6. The
number of customers processed each day can be found under the Current Value
column. The simulation results indicate that the number of customers processed
each day fluctuates randomly. The fluctuation may be attributed to variations in
processing times or customer arrivals. It is obvious that we cannot reliably use
the observation from a single day’s simulation to estimate the expected number
of customers served each day. The results from a single day could produce an es-
timate as small as 59 or as large as 67. Either value could be a very poor estimate
of the true expected number of customers served each day.
har01307_LB08_527-548.indd 531 20/12/10 1:55 PM

532 Part II Labs
FIGURE L8.3
ProModel’s Simulation
Options window set to
run five replications of
the simulation.
FIGURE L8.4
Selecting variable
summary statistics
from the ProModel
Output Viewer.
The average number of customers processed over the five days is computed
_
as x 62.20 (Figure L8.6). This provides a much better point estimate of the ex-
pected daily service rate (number of customers processed each day), but it fails to
capture any information on the variability of our estimate. The sample deviation
of s 3.27 customers per day provides that information. The most valuable piece
of information is the 90 percent confidence interval computed by ProModel using
har01307_LB08_527-548.indd 532 20/12/10 1:55 PM

FIGURE L8.5
Statistical Options in
Output Viewer set to
the average standard
deviation, and
90 percent confidence
interval.
FIGURE L8.6
ProModel output report for five replications of the Spuds-n-More simulation.
the sample mean and sample standard deviation. With approximately 90 percent
confidence, the true but unknown mean number of customers processed per day
falls between 59.08 and 65.32 customers. These results convince Mr. Taylor that
the model is a valid representation of the actual restaurant, but he wants to get a
better estimate of the number of customers served per day.
har01307_LB08_527-548.indd 533 20/12/10 1:55 PM

534 Part II Labs
L8.2.3 Required Number of Replications

The advantage of the confidence interval estimate is that it quantifies the error in
the point estimate. A wide confidence interval implies that the point estimate is
not very accurate and should be given little credence. The width of the confidence
interval is a function of both the variability of the system and the amount of data
collected for estimation. We can make the confidence interval smaller by collect-
ing more data (running more replications). However, for highly variable systems,
one may need a lot of data to make extremely precise statements (smaller confi-
dence intervals). Collecting the data may be very time consuming if the simula-
tion takes long to run. Hence, we sometimes have to make a judicious choice
about the amount of precision required.
The half-width of the 90 percent confidence interval based on the five replica-
tions is 3.12 customers (65.32 C.I. High – 62.20 Average). How many additional
replications are needed if Mr. Taylor desires a confidence interval half-width
of 2.0 customers? Using the technique described in section 8.2.3 of Chapter 8
we have
e hw 2.0 customers
␣ significance level 0.10 for a 0.90 confidence level

n ______
e
(Z/2)s 2
______
e
(t,/2)s 2 (1.645)3.27 2

__________ 7.23 replications
2.0
The value of 7.23 is appropriately rounded up to 8 replications. The n 8 pro-
vides a rough approximation of the number of replications that are needed to
achieve the requested error amount e 2.0 customers at the 0.10 significance
level. Running an additional three replications of the Spuds-n-More baseline sim-
ulation and combining the output with the output from the original five replica-
tions produces a sample mean of 61.50 customers and an approximate 90 percent
confidence interval of 58.99 to 64.01 customers. You should verify this. Although
the confidence interval half-width of 2.51 customers is larger than the target value
of 2.0 customers, it is smaller than the half-width of the confidence interval based
on five replications. We could keep running additional replications to get closer to
the target half-width of 2.0 customers if we like.
L8.2.4 Simulation Output Assumptions

To compute the confidence interval, we assumed that the observations of the num-
ber of customers processed each day were independent and normally distributed.
The observations are independent because ProModel used a different stream of
random numbers to simulate each replication. We do not know that the observations
are normally distributed, however. Therefore, we should specify that the confidence
level of the interval is approximately 0.90. Nevertheless, we followed an accept-
able procedure to get our estimate for the service rate, and we can stand by our
approximate confidence interval. To illustrate a point, however, let’s take a moment
to use the Stat::Fit software provided with ProModel to test if the simulation output
is normally distributed. To do so, we need a lot more than eight observations. Let’s
har01307_LB08_527-548.indd 534 20/12/10 1:55 PM

FIGURE L8.7
Exporting a ProModel output report into Microsoft Excel.
pick 100 just because it is a round number and we can quickly generate them by con-
figuring ProModel to run 100 replications of the simulation. Try this now and open
the Variable Summary output report from the Tables menu when the simulation ends.
Notice that you get a very small confidence interval with this many observations.
It would take a while to enter the 100 observations into Stat::Fit one at a time. A
faster method is to export the ProModel output report with the results from the 100
replications into Microsoft Excel (MS). From MS Excel, copy the observations into a
Stat::Fit data table. To do this, select Summary Data from the Export menu and export
the file to MS Excel (Figure L8.7). Open the MS Excel file and select the Variable
Summary Worksheet (Figure L8.8). From MS Excel, highlight the 100 Current Value
observations of the Processed variable—making sure not to include the average, stan-
dard deviation, 90% C.I. Low, and 90% C.I. High values (the last four values in the
column)—and paste the 100 observations into a Stat::Fit data table (Figure L8.9).
After the 100 observations are pasted into the Stat::Fit Data Table, display a
histogram of the data. Based on the histogram, the observations appear somewhat
normally distributed (Figure L8.9). Furthermore, Stat::Fit estimates that the nor-
mal distribution with mean ␮ 59.30 and standard deviation ␴ 4.08 provides
an acceptable fit to the data. Therefore, we probably could have dropped the word
“approximate” when we presented our confidence interval to Mr. Taylor. Be sure
to verify all of this for yourself using the software.
Note that if you were to repeat this procedure in practice because the problem
requires you to be as precise and as sure as possible in your conclusions, then you
would report your confidence interval based on the larger number of observations.
Never discard precious data.
Before leaving this section, go to the ProModel Entity Failed Arrivals output
report. The Total Failed column in the Failed Arrivals section of the output report
indicates the number of customers that arrived to eat at Spuds-n-More but left
har01307_LB08_527-548.indd 535 20/12/10 1:55 PM

536 Part II Labs
FIGURE L8.8
The ProModel output
report exported into
a Microsoft Excel
spreadsheet.
because the waiting line was full. Mr. Taylor thinks that his proposed expansion
plan will allow him to capture some of the customers he is currently losing. In
Lab Chapter 9, we will add embellishments to the as-is simulation model to reflect
Mr. Taylor’s expansion plan to see if he is right.
L8.3 Nonterminating Simulation

We run a nonterminating simulation of a system if we are interested in under-
standing the steady-state behavior of the system—that is, its long term average
behavior. Because a nonterminating system does not return to a fixed initial condi-
tion and theoretically never shuts down, it is not easy to select either the starting
condition for the simulation or the amount of time to run the simulation.
When a simulation has just begun running, its state (number of customers in
the system for example) will be greatly affected by the initial state of the simulation
and the elapsed time of the run. In this case, the simulation is said to be in a transient
condition. However, after sufficient time has elapsed, the state of the simulation
becomes essentially independent of the initial state and the elapsed time. When this
har01307_LB08_527-548.indd 536 20/12/10 1:55 PM

FIGURE L8.9
Stat::Fit analysis of the 100 observations of the number of customers processed per day
by Spuds-n-More.
happens, the simulation has reached a steady-state condition. Although the simula-
tion’s output will continue to be random, the statistical distribution of the simula-
tion’s output does not change after the simulation reaches steady state.
The period during which the simulation is in the transient phase is known as the
warm-up period. This is the amount of time to let the simulation run before gather-
ing statistics. Data collection for statistics begins at the end of the warm-up and
continues until the simulation has run long enough to allow all simulation events
(even rare ones) to occur many times (hundred to thousands of times if practical).
Problem Statement
A simulation model of the Green Machine Manufacturing Company (GMMC)
owned and operated by Mr. Robert Vaughn is shown in Figure L8.10. The interar-
rival time of jobs to the GMMC is constant at 1.175 minutes. Jobs require pro-
cessing by each of the four machines. The processing time for a job at each green
machine is given in Table L8.2.
A plentiful amount of storage is provided for waiting jobs in front of each
of the green machines. Our task is to conduct a nonterminating simulation to
estimate the time-average amount of work-in-process (WIP) inventory (the
har01307_LB08_527-548.indd 537 20/12/10 1:55 PM

538 Part II Labs
FIGURE L8.10
The layout of the Green Machine Manufacturing Company simulation model.
TABLE L8.2 Process Times at Green Machine Manufacturing Company
Activity Process Time (minutes) ProModel Format
Machine 1 Exponential with mean 1.0 Wait E(1.0) min

Machine 2 Weibull with minimum 0.10, ␣ 5, ␤ 1 Wait (0.10 W(5,1)) min
steady-state expected number of jobs in the system). A printout of the model is

given in Figure L8.11. The model is found in the Student folder installed with
ProModel under the file name Lab 8_3 GreenMachine.MOD.
L8.3.1 Warm-up Time and Run Length

In this section we estimate the amount of time needed for the GMMC model to
warm up, and the amount of time to run the model past its warm-up period. We
first declare a model variable called WIP to keep track of the work-in-process in-
ventory in the simulated manufacturing system. Notice we are collecting statistics
only for the WIP variable and the entity. You should configure your model the
same and run the simulation 250 hours to produce a time-series plot of the WIP
inventory levels (Figure L8.12(a)).
The simplest method for selecting the end of the warm-up phase is visual
determination. In this method, we plot the simulation model’s output over time,
har01307_LB08_527-548.indd 538 20/12/10 1:55 PM

FIGURE L8.11
The simulation model of the Green Machine Manufacturing Company.
har01307_LB08_527-548.indd 539 20/12/10 1:55 PM

540 Part II Labs
FIGURE L8.12
Work-in-process (WIP)
inventory value history
for one replication
of the GMMC
simulation. (a) WIP
value history without
warm-up phase
removed. Statistics
will be biased low
by the transient WIP
values. (b) WIP value
history with 100- (a)
hour warm-up phase
removed. Statistics will
not be biased low.
(b)
as illustrated in Figure L8.12(a), and then visually determine the point at which
the initial transient phase is over. In this example, it seems that the transient phase
created by the initial condition of zero WIP inventory is over at around 20 hours.
Usually you will want to select a value beyond this to guard against underestimat-
ing the end of the warm-up phase. So we will pick 100 hours as the end of the
warm-up. Actually, you should not make this decision based on the output from
one replication of the simulation. We are doing this here to quickly illustrate the
need for allowing the model to warm up. We will use the preferred method to
estimate the end of the warm-up phase in a moment.
In most nonterminating simulation models, the quality of the results substan-
tially improves if we throw away the output generated during the warm-up phase
of the simulation. In Figure L8.12(a), the observations during the transient phase
of the first 100 hours of the simulation are alleged to be artificially low. As such,
we do not want them to bias our estimate of the time-average WIP inventory.
Therefore, in our next run of the GMMC simulation, we specify a warm-up time
of 100 hours (Figure L8.13). Notice how the simulation output during the 100-hour
warm-up has been discarded in Figure L8.12(b). The discarded output will no
longer affect the observations reported in the simulation output.
har01307_LB08_527-548.indd 540 20/12/10 1:55 PM

FIGURE L8.13
Options window set
for a single replication
with a 100-hour
warm-up period
followed by a
150-hour run length.
The simplified process we just used to identify the end of the warm-up phase
should be avoided. The process can be substantially improved by running multiple
replications of the simulation and then combining the time-series output from
each replication into an average time-series plot. Averaging the time-series data
from multiple replications together will help filter out the large fluctuations (high
values from one replication are offset by small values from another replication)
and will give you a better estimate of the true but unknown time-series output.
In many cases, you may need to add an additional layer of filtering to smooth
the “averaged” time-series plot. This can be done by computing a Welch moving
average of the time-series data. Section 8.6.1 of Chapter 8 provides additional
details on this subject, and exercise 4 in Lab Chapter 10 illustrates how to use the
Welch method implemented in SimRunner.
Figure L8.14 was produced by SimRunner by recording the time-average
WIP levels of the GMMC simulation over successive one-hour time periods. The
results from each period were averaged across five replications to produce the
“raw data” plot, which is the more erratic line that appears red on the computer
screen. A 54-period moving average (the smooth line that appears green on the
computer) indicates that the end of the warm-up phase occurs between the 33rd
and 100th periods. Given that we need to avoid underestimating the warm-up,
let’s declare 100 hours as the end of the warm-up time. We feel much more com-
fortable basing our estimate of the warm-up time on five replications. Notice that
SimRunner indicates that at least 10 replications are needed to estimate the aver-
age WIP to within a 7 percent error and a confidence level of 90 percent using a
warm-up of 100 periods (hours). You will see how this was done with SimRunner
in Exercise 4 of section L10.4 in Lab Chapter 10.
har01307_LB08_527-548.indd 541 20/12/10 1:55 PM

542 Part II Labs
FIGURE L8.14
SimRunner screen for estimating the end of the warm-up time.
Why did we choose an initial run length of 250 hours to produce the time-
series plot in Figure L8.14? Well, you have to start with something, and we picked
250 hours. You can rerun the experiment with a longer run length if the time-
series plot, produced by averaging the output from several replications of the
simulation, does not indicate a steady-state condition. Long runs will help pre-
vent you from wondering what the time-series plot looks like beyond the point
at which you stopped it. Do you wonder what our plot does beyond 250 hours?
Let’s now direct our attention to answering the question of how long to run
the model past its warm-up time to estimate our steady-state statistic, mean WIP
inventory. We will somewhat arbitrarily pick 100 hours not because that is equal
to the warm-up time but because it will allow ample time for the simulation events
to happen thousands of times. In fact, the 100-hour duration will allow approxi-
mately 5,100 jobs to be processed per replication, which should give us decently
accurate results. How did we derive the estimate of 5,100 jobs processed per
har01307_LB08_527-548.indd 542 20/12/10 1:55 PM

replication? With a 1.175-minute interarrival time of jobs to the system, 51 jobs

arrive per hour (60 minutes/1.175 minutes) to the system. Running the simulation
for 100 hours should result in about 5,100 jobs (100 hours 51 jobs) exiting the
system. You will want to check the number of Total Exits in the Entity Activity
section of the ProModel output report that you just produced to verify this.
L8.3.2 Replications or Batch Intervals

We have a choice between running replications or batch intervals to estimate the
time-average WIP inventory. Replications are our favorite choice because we get
independent observations, guaranteed. Figure L8.15 illustrates how this is done
in ProModel as well as the specification of the 100 hours of run time beyond
the warm-up. Running the simulation for 10 replications produces the results in
Figure L8.16. We are approximately 90 percent confident that the true but un-
known mean WIP inventory is between 17.64 and 21.09 jobs.
If we decide to make one single long run and divide it into batch intervals to
estimate the expected WIP inventory, we would select the Batch Mean option in
the Output Reporting section of the Simulation Options menu (Figure L8.17). A
guideline in Chapter 8 for making an initial assignment to the length of time for
each batch interval was to set the batch interval length to the simulation run time
you would use for replications, in this case 100 hours. Based on the guideline,
the Simulation Options menu would be configured as shown in Figure L8.17 and
FIGURE L8.15
Specification of
warm-up hours and
run hours in the
Simulation Options
menu for running
replications.
har01307_LB08_527-548.indd 543 20/12/10 1:55 PM

544 Part II Labs
FIGURE L8.16
Ten replications of the Green Machine Manufacturing Company simulation using a 100-hour
warm-up and a 100-hour run length.
FIGURE L8.17
Warm-up hours, run
hours, and batch
interval length in the
Simulation Options
menu for running
batch intervals. Note
that the time units are
specified on the batch
interval length.
har01307_LB08_527-548.indd 544 20/12/10 1:55 PM

FIGURE L8.18
Ten batch intervals of the Green Machine Manufacturing Company simulation using a
100-hour warm-up and a 100-hour batch interval length.
would produce the output in Figure L8.18. To get the output report to display the
results of all 10 batch intervals, you specify “<All>” Periods in the Report selec-
tion settings panel. We are approximately 90 percent confident that the true but
unknown mean WIP inventory is between 18.88 and 23.19 jobs. If we desire a
smaller confidence interval, we can increase the sample size by extending the run
length of the simulation in increments of the batch interval length. For example,
we would specify run hours of 1500 to collect 15 observations of average WIP
inventory. Try this and see if the confidence interval becomes smaller. Note that
increasing the batch interval length is also helpful.
L8.3.3 Required Batch Interval Length

The worrisome part of the batch interval method is whether the batch interval
is long enough to acquire “approximately” independent observations. The word
“approximately” is emphasized because we know that the observations within a
simulation run are often autocorrelated (not independent). In fact, they are often
positively correlated, which would bias the sample standard deviation of the batch
means to the low side. And this would bias our estimate of the confidence inter-
val’s half-width to the low side. We do not know if this is the case with our WIP
inventory observations. However, if it was the case and we were not aware of
it, we would think that our estimate of the mean WIP inventory was better than
it really is. Therefore, we should do ourselves a favor and use Stat::Fit to check
har01307_LB08_527-548.indd 545 20/12/10 1:55 PM

546 Part II Labs
the lag-1 autocorrelation of the observations before we interpret the results. To

get a decent estimate of the lag-1 autocorrelation, however, we need more than
10 observations. Section 8.6.2 of Chapter 8 stated that at least 100 observations
are necessary to get a reliable estimate of the lag-1 autocorrelation. So let’s in-
crease the simulation run length to 10,000 hours and produce an output report
with 100 observations of the time-average WIP inventory.
It would take a while to enter the 100 observations into Stat::Fit one at a time.
Therefore, export the ProModel output report with the results from the 100 batch
intervals into Microsoft Excel so that they can then be copied into a Stat::Fit data
table. The procedure is the same as used in Section L8.2.4 to enter the observations
under the Current Value column of the ProModel Variable Summary report into
Stat::Fit—except that you are now entering the observations under the Average
Value column from the Variable Summary report into Stat::Fit.
Figure L8.19 illustrates the Stat::Fit results that you will want to verify. The
lag-1 autocorrelation value is the first value plotted in Stat::Fit’s Autocorrelation
of Input Data plot. Note that the plot begins at lag-1 and continues to lag-20. For
FIGURE L8.19
Stat::Fit Autocorrelation plot of the observations collected over 100 batch intervals.
Lag-1 autocorrelation is within the –0.20 to 0.20 range.
har01307_LB08_527-548.indd 546 20/12/10 1:55 PM

TABLE L8.3 Final Results of the 100 Batch Intervals

Combined into 10 Batch Intervals
WIP Average Value
21.04 (Batch 1)
19.40 (Batch 2)
21.12 (Batch 3)
20.29 (Batch 4)
20.33 (Batch 5)
20.41 (Batch 6)
22.38 (Batch 7)
22.02 (Batch 8)
19.93 (Batch 9)
22.23 (Batch 10)
20.92 (Average)
1.02 (Std. Dev.)
20.32 (90% C.I. Low)
21.51 (90% C.I. High)
this range of lag values, the highest autocorrelation is 0.192 and the lowest value
is 0.132 (see correlation (0.192, 0.132) at the bottom of the plot). Therefore,
we know that the lag-1 autocorrelation is within the 0.20 to 0.20 range rec-
ommended in section 8.6.2 of Chapter 8, which is required before proceeding to
the final step of “rebatching” the data into between 10 and 30 larger batches. We
have enough data to form 10 batches with a length of 1000 hours (10 batches at
1000 hours each equals 10,000 hours of simulation time, which we just did). The
results of “rebatching” the data in the Microsoft Excel spreadsheet are shown in
Table L8.3. Note that the first batch mean of 21.037 in Table L8.3 is the average
of the first 10 batch means that we originally collected. The second batch mean of
19.401 is the average of the next 10 batch means and so on. The confidence inter-
val is very narrow due to the long simulation time (1000 hours) of each batch. The
owner of GMMC, Mr. Robert Vaughn, should be very pleased with the precision
of the time-average WIP inventory estimate.
The batch interval method requires a lot of work to get it right. Therefore, if
the time to simulate through the warm-up period is relatively short, the replica-
tions method should be used. Reserve the batch interval method for simulations
that require a very long time to reach steady state.
L8.4 Exercises
1. An average of 100 customers per hour arrive to the Picayune Mutual
Bank. It takes a teller an average of two minutes to serve a customer.
Interarrival and service times are exponentially distributed. The bank
currently has four tellers working. Bank manager Rich Gold wants to
har01307_LB08_527-548.indd 547 20/12/10 1:55 PM

548 Part II Labs
compare the following two systems with regard to the average time
customers spend in the bank.
System #1
A separate queue is provided for each teller. Assume that customers
choose the shortest queue when entering the bank, and that customers
cannot jockey between queues (jump to another queue).
System #2
A single queue is provided for customers to wait for the first available
teller.
Assume that there is no move time within the queues. Run 15
replications of an eight-hour simulation to complete the following:
a. For each system, record the 90 percent confidence interval using a
0.10 level of significance for the average time customers spend in the
bank.
b. Estimate the number of replications needed to reduce the half-
width of each confidence interval by 25 percent. Run the additional
replications for each system and compute new confidence intervals
using a 0.10 level of significance.
c. Based on the results from part b, would you recommend one system
over the other? Explain.
2. Rerun the simulation of the Green Machine Manufacturing Company of
Section L8.3 to produce 100 batch means with a 50-hour interval length.
Use a 100-hour warm-up. Check the lag-1 autocorrelation of the 100 time-
average WIP observations to see if it falls between ⫺0.20 and ⫹0.20. If
it does not, increase the batch interval length by 50 percent and continue
the process until an acceptable lag-1 autocorrelation is achieved. Once the
lag-1 autocorrelation is acceptable, check to see if the 100 observations
are normally distributed. After that, rebatch the 100 observations into
20 batch means and compute a 95 percent confidence interval. Are you
comfortable with the resulting confidence interval? Explain.
3. Use the simulation model created for the DumpOnMe facility presented
in Exercise 11 of Lab 4 section L4.13 to complete the following:
a. For a simulation run length of 60 minutes, record the 90 percent
confidence interval for the scale’s average utilization based on
30 replications.
b. For a 60-minute warm-up and a 60-minute simulation run length
beyond the warm-up time, record the 90 percent confidence interval
for the scale’s average utilization based on 30 replications.
c. Based only on the two confidence intervals, do you think that a
warm-up period is necessary for the DumpOnMe model? Explain.
4. See Lab Chapter 10 for exercises involving the use of SimRunner’s
implementation of the Welch moving average technique.
har01307_LB08_527-548.indd 548 12/22/10 7:14 AM

L A B
9 COMPARING
ALTERNATIVE SYSTEMS
The great tragedy of Science—the slaying of a beautiful hypothesis by an

ugly fact.
—Thomas H. Huxley
In this lab we see how ProModel is used with some of the statistical methods
presented in Chapter 9 to compare alternative designs of a system with the goal
of identifying the superior system relative to some performance measure. We will
also learn how to program ProModel to use common random numbers (CRN)
to drive the simulation models that represent alternative designs for the systems
being compared. The use of CRN allows us to run the opposing simulations
under identical experimental conditions to facilitate an objective evaluation of
the systems.
L9.1 Overview of Statistical Methods

Often we want to compare alternative designs of a system. For example, we
may wish to compare two different material handling systems, production
scheduling rules, plant layouts, or staffing schedules to determine which yields
better results. We compare the two alternative system designs based on a certain
measure of performance—for example, production rate, number of customers
served, time in system, or waiting time for service. The comparison is based on
the estimates of the expected value of the performance measure that we derive
from simulating the alternative system designs. If the estimates provide evidence
that the performances of the two alternative simulation models are not equal, we
can justify picking the alternative that produced the better estimate as the best
alternative.
549
har01307_LB09_549-568.indd 549 12/17/10 10:16 AM

550 Part II Labs
Chapter 9 covers the paired-t confidence interval method and the Welch
confidence interval method for testing hypotheses about two alternatives. The
Bonferroni approach is used for comparing from three to about five alternatives.
Either the Welch method or paired-t method is used with the Bonferroni approach
to make the comparisons. The Bonferroni approach does not identify the best al-
ternative but rather identifies which pairs of alternatives perform differently. This
information is helpful for identifying the better alternatives, if in fact there is a
significant difference in the performance of the competing alternatives.
Chapter 9 recommends that the technique of analysis of variance (ANOVA)
be used when more than about five alternative system designs are being com-
pared. The first step in ANOVA is to evaluate the hypothesis that the mean perfor-
mances of the systems are equal against the alternative hypothesis that at least one
pair of the means is different. To figure out which means differ from which other
ones, a multiple comparison test is used. There are several multiple comparison
tests available, and Chapter 9 presents the Fisher’s protected least significant dif-
ference (LSD) test.
The decision to use paired-t confidence intervals, Welch confidence inter-
vals, the Bonferroni approach, or ANOVA with protected LSD depends on the
statistical requirements placed on the data (observations from the simulations) by
the procedure. In fact, simulation observations produced using common random
numbers (CRN) cannot be used with the Welch confidence interval method or the
technique of ANOVA. Please see Chapter 9 for details.
In this lab we will use the paired-t confidence interval method because it
places the fewest statistical requirements on the observations collected from the
simulation model. Therefore, we are less likely to get into trouble with faulty as-
sumptions about our data (observations) when making decision based on paired-
t confidence intervals. Furthermore, one of the main purposes of this lab is to
demonstrate how the CRN technique is implemented using ProModel. Out of the
comparison procedures of Chapter 9, only the ones based on paired-t confidence
intervals can be used in conjunction with the CRN technique.
L9.2 Three Alternative Systems

Problem Statement
The analysis of the Spuds-n-More by the Seashore restaurant started in Section
L8.2 of Lab Chapter 8 will be completed in this lab. The details of the system
and objective for the study are reviewed here. Spuds-n-More, a takeout fast-food
restaurant, is a favorite among vacationers at a seaside resort. Customers walking
along the boardwalk access the restaurant via a single ordering window, where they
place their orders, pay, and receive their food. There is no seating in the restaurant.
The interarrival time of customers to Spuds-n-More is exponentially distributed
with a mean of six minutes. The activity times are given in Table L8.1. The waiting
area in front of the walkup window can accommodate only five customers. If the
waiting area is filled to capacity when a customer arrives, the customer bypasses
har01307_LB09_549-568.indd 550 12/17/10 10:16 AM

Spuds-n-More. Mr. Taylor closes the entry to the waiting area in front of his res-
taurant after eight hours of operation. However, any customers already in the wait-
ing area by closing time are served before the kitchen shuts down for the day.
Mr. Taylor is thinking about expanding into the small newspaper stand next
to Spuds-n-More, which has been abandoned. This would allow him to add a
second window to serve his customers. The first window would be used for taking
orders and collecting payments, and the second window would be used for filling
orders. Before investing in the new space, however, Mr. Taylor needs to convince
the major stockholder of Spuds-n-More, Pritchard Enterprises, that the invest-
ment would allow the restaurant to serve more customers per day.
A baseline (as-is) model of the current restaurant configuration was devel-
oped and validated in Section L8.2 of Lab Chapter 8. See Figures L8.1 and L8.2
for the layout of the baseline model and the printout of the ProModel model.
Our task is to build a model of the proposed restaurant with a second window to
determine if the proposed design would serve more customers per day than does
the current restaurant. Let’s call the baseline model Spuds-n-More1 and the pro-
posed model Spuds-n-More2. The model is found in the Student folder installed
with ProModel.
The ProModel simulation layout of Spuds-n-More2 is shown in Figure L9.1,
and the ProModel printout is given in Figure L9.2. After customers wait in the
order queue (location Order_Q) for their turn to order, they move to the first win-
dow to place their orders and pay the order clerk (location Order_Clerk). Next
customers proceed to the pickup queue (location Pickup_Q) to wait for their turn
FIGURE L9.1
Layout of Spuds-n-
More2 simulation
model.
har01307_LB09_549-568.indd 551 12/17/10 10:16 AM

552 Part II Labs
FIGURE L9.2
The Spuds-n-More2 simulation model.
to be served by the pickup clerk (Pickup_Clerk) at the second window. There is

enough space for one customer in the pickup queue. The pickup clerk processes
one order at a time.
As the Spuds-n-More2 model was being built, Mr. Taylor determined that
for additional expense he could have a carpenter space the two customer service
windows far enough apart to accommodate up to three customers in the order
pickup queue. Therefore, he requested that a third alternative design with an order
har01307_LB09_549-568.indd 552 12/17/10 10:16 AM

pickup queue capacity of three be simulated. To do this, we only need to change

the capacity of the pickup queue from one to three in our Spuds-n-More2 model.
We shall call this model Spuds-n-More3. Note that for our Spuds-n-More2 and
Spuds-n-More3 models, we assigned a length of 25 feet to the Order_Q, a length
of 12 feet to the Pickup_Q, and a Customer entity travel speed of 150 feet per min-
ute. These values affect the entity’s travel time in the queues (time for a customer
to walk to the end of the queues) and are required to match the results presented
in this lab.
L9.3 Common Random Numbers

Let’s conduct the analysis for Mr. Taylor using Common Random Numbers
(CRN). The CRN technique provides a way of comparing alternative system de-
signs under more equal experimental conditions. This is helpful in ensuring that
the observed differences in the performance of system designs are due to the
differences in the designs and not to differences in experimental conditions. The
goal is to evaluate each system under the exact same set of circumstances to en-
sure that a fair comparison is made. Therefore, the simulation of each alternative
system design is driven by the same stream of random numbers to ensure that
the differences in key performance measures are due only to differences in the
alternative designs and not because some designs were simulated with a stream
of random numbers that produced more extreme conditions. Thus we need to use
the exact same random number from the random number stream for the exact
same purpose in each simulated system. A common practice that helps to keep
random numbers synchronized across systems is to assign a different random
number stream to each stochastic element in the model. For this reason, Pro-
Model provides 100 unique streams (Stream 1 to Stream 100) of random numbers
for generating random variates from the statistical distributions available in the
software.
The Spuds-n-More models use observations (random variates) from an expo-
nential distribution for the customer interarrival time, order time, order payment
time, and order pickup time. For CRN, in this example we can use Stream 1 to
generate interarrival times, Stream 2 to generate order times, Stream 3 to generate
payment times, and Stream 4 to generate pickup times. The ProModel function
for generating random variates from the exponential distribution is E(a, stream),
where the parameter a is the mean of the random variable and the parameter
stream is the ProModel stream number between 1 and 100. All ProModel distribu-
tion functions include the stream parameter as the last parameter of the function.
The stream parameter is optional (you do not have to specify a value). If you do
not provide a value for the stream parameter, ProModel uses stream 1 to gener-
ate observations (random variates) from the specified distribution. Figure L9.3
illustrates the desired assignment of random number streams in the processing
and arrivals section of a Spuds-n-More model. Configure all three of your Spuds-
n-More models accordingly.
har01307_LB09_549-568.indd 553 12/17/10 10:16 AM

554 Part II Labs
FIGURE L9.3
Unique random number streams assigned to the Spuds-n-More model.
L9.4 Bonferroni Approach with Paired-t Confidence Intervals

Our objective is to compare the performance of the three alternative Spuds-n-
More restaurant designs with respect to the number of customers processed per
day. Let µi represent the true number of customers processed per day by res-
taurant design i, where i 1 denotes Spuds-n-More1, i 2 denotes Spuds-n-
More2, and i 3 denotes Spuds-n-More3. Specifically, we wish to evaluate the
hypotheses
H0: µ1 µ2 µ3
H1: µ1 µ2 or µ1 µ3 or µ2 µ3
at the 0.06 significance level. To evaluate these hypotheses, we estimate
the number of customers processed by each restaurant design by simulating
each design for 25 days (25 replications). The simulations are to be driven by
Common Random Numbers. Therefore, select the Common Random Num-
bers (CRN) option in the Simulation Options window before running the
simulations (Figure L9.4). This option helps to synchronize the use of ran-
dom numbers across the three simulated systems. When the option is selected,
ProModel starts each replication at predetermined locations on the random
number streams. At the end of a replication, ProModel advances to the next
set of predetermined locations on the streams before starting the next repli-
cation. Even if one model uses more random numbers than another during a
replication, both models will begin the next replication from the same posi-
tions on the random number streams. The numbers of customers processed by
har01307_LB09_549-568.indd 554 12/17/10 10:16 AM

FIGURE L9.4
Options set to run
25 replications using
Common Random
Numbers.
each restaurant design for 25 replications using the CRN option are shown in
Table L9.1.
The Bonferroni approach with paired-t confidence intervals is used to evaluate
our hypotheses for the three alternative designs for the restaurant. The evaluation
of the three restaurant designs requires that three pairwise comparisons be made:
Spuds-n-More1 vs. Spuds-n-More2
Following the procedure described in Section 9.4.1 of Chapter 9 for the
Bonferroni approach, we begin constructing the required three paired-t confi-
dence intervals by letting 1 2 3 /3 0.06/3 0.02.
The computation of the three paired-t confidence intervals follows:
Spuds-n-More1 vs. Spuds-n-More2 (µ(1–2))

1 0.02
tn1, /2 t24,0.01 2.485 from Appendix B
1
(t24,0.01)s(12) (2.485)4.66
hw __________
冑n
__ __________
___
冑25 2.32 customers

_ _
x(12) hw µ(12) x(12) hw
16.20 2.32 µ(12) 16.20 2.32
18.52 µ(12) 13.88
har01307_LB09_549-568.indd 555 12/17/10 10:16 AM

556 Part II Labs
TABLE L9.1 Comparison of the Three Restaurant Designs Based on Paired Differences
(A) (B) (C) (D) (E) (F) (G)

Spuds-n-More1 Spuds-n-More2 Spuds-n-More3
Customers Customers Customers Difference Difference Difference
Rep. Processed Processed Processed (B − C) (B − D) (C − D)
(j) x1j x2j x3j x(1–2)j x(1–3) j x(2–3) j
1 60 72 75 12 15 3
2 57 79 81 22 24 2
3 58 84 88 26 30 4
4 53 69 72 16 19 3
5 54 67 69 13 15 2
6 56 72 74 16 18 2
7 57 70 71 13 14 1
8 55 65 65 10 10 0
9 61 84 84 23 23 0
10 60 76 76 16 16 0
11 56 66 70 10 14 4
12 66 87 90 21 24 3
13 64 77 79 13 15 2
14 58 66 66 8 8 0
15 65 85 89 20 24 4
16 56 78 83 22 27 5
17 60 72 72 12 12 0
18 55 75 77 20 22 2
19 57 78 78 21 21 0
20 50 69 70 19 20 1
21 59 74 77 15 18 3
22 58 73 76 15 18 3
23 58 71 76 13 18 5
24 56 70 72 14 16 2
25 53 68 69 15 16 1
_
Sample mean x(ii), for all i and i between 1 and 3, with i i 16.20 18.28 2.08
Sample standard dev s(ii), for all i and i between 1 and 3, with i i 4.66 5.23 1.61

2 0.02
tn1, /2 t24,0.01 2.485 from Appendix B
2
(t24,0.01__)s(13) __________
(2.485)5.23
hw _________ ___ 2.60 customers
冑n 冑25
_ _
x(13) hw µ(13) x(13) hw
18.28 2.60 µ(13) 18.28 2.60
20.88 µ(13) 15.68
har01307_LB09_549-568.indd 556 12/17/10 10:16 AM


2.88 µ(23) 1.28
Given that the confidence interval for µ(12) excludes zero, we conclude that
there is a significant difference in the mean number of customers processed by
the current “as-is” restaurant Spuds-n-More1 and the proposed Spuds-n-More2
design. The confidence interval further suggests that Spuds-n-More2 processes an
estimated 13.88 to 18.52 more customers per day on average than does Spuds-n-
More1. Likewise, there is a significant difference in the mean number of customers
processed by Spuds-n-More1 and Spuds-n-More3. The proposed Spuds-n-More3
processes an estimated 15.68 to 20.88 more customers per day on average than
does Spuds-n-More1. The third confidence interval suggests that the difference
between Spuds-n-More2 and Spuds-n-More3, although statistically significant,
is fairly small with Spuds-n-More3 processing an estimated 1.28 to 2.88 more
customers per day on average than does Spuds-n-More2.
With 0.06 the overall confidence for our conclusions is approximately
94 percent. Based on these results we are justified in concluding that both of the new
designs (Spuds-n-More2 and Spuds-n-More3) serve more customers per day than
does the current “as-is” restaurant. It appears that Spuds-n-More3 is not that much
better than Spuds-n-More2 from a customer service rate perspective. Perhaps we will
advise Mr. Taylor to request funding from the restaurant’s stockholders to implement
the Spuds-n-More2 design because the confidence interval suggests that it processes
an estimated 14 to 19 more customers per day than does the current restaurant. The
extra expense of Spuds-n-More3 above the implementation cost of Spuds-n-More2
for spacing the order and pickup windows farther apart will probably not be offset by
the extra income from processing only an estimated 1 to 3 more customers per day.
In the first exercise problem in section L9.6, we will find out how much of a
reduction, if any, we achieved in the half-widths of the confidence intervals by using
the CRN technique. Theoretically, the CRN technique produces a positive correla-
tion between the observations in columns B, C, and D of Table L9.1. The positive
correlation reduces the value of the sample standard deviation of the observations in
the difference columns (columns E, F, and G of Table L9.1). The smaller standard
deviation results in a smaller confidence interval half-width (a more precise state-
ment). This is important because, for example, if the third confidence interval was
10.93 µ(23) 1.79, which suggests that Spuds-n-More3 processes an esti-
mated 1.79 to 10.93 more customers per day on average than does Spuds-n-More2,
then we may not be so quick to rule out Spuds-n-More3 as a serious competitor to
Spuds-n-More2 due to the greater range of the confidence interval.
L9.5 Analysis of Designed Experiments

We first introduced you to the Design of Experiments in Chapter 9, section 9.4.3.
Experimental design consists of purposeful changes of the input factors to a
process or activity in order to observe the corresponding changes in the outputs
har01307_LB09_549-568.indd 557 12/17/10 10:16 AM

558 Part II Labs
(responses). A process is a combination of machines, materials, methods, people,

environment, and measurements which, when used together, perform a service,
produce a product, or both. Thus experimental design is a scientific approach that
allows the researcher to gain knowledge in order to better understand a process
and determine how the inputs affect the response(s).
The principal value of simulation analysis is to study change. Computer sim-
ulation models allow us to study the relationship between independent (input) and
dependent (output) factors in an experiment. In the real world, this capability is
to a major extent a function of the system under study. In a computer simulation,
an investigator must explicitly specify the sources of variation and the degree of
variation due to each source in order to make his/her simulation run.
When there are only a few input factors to test, often a full factorial design
(all possible combinations) is the best choice. In section L9.5.1 we describe a full
factorial design simulation experiment. When the number of input factors is large
a fraction factorial design is usually a more efficient alternative. In section L9.5.2
we present a fraction factorial design simulation experiment. A demonstration of
how to conduct a design of experiments in ProModel and then utilize Minitab
to plot the results is shown at http://www.promodel.com/solutionscafe/webinars/
DOE%20Webinar/DOE%20Webinar.html.
L9.5.1 Full Factorial Design Simulation Experiment

Apna Bazaar, a supermarket store in India, has two regular checkout counters.
The time between arrivals of customers is exponentially distributed with a mean
of six minutes. Upon arrival customers grab a shopping cart (currently there are
5 shopping carts). A customer waits until a cart is available. The shopping time
and checkout time of customers are as shown in Table L9.2.
Customers chose the first available checkout line. Once a customer joins a
line, they cannot switch to another line. The supermarket is open 24 hours/day,
365 days a year.
As a result of intense advertising on TV, management anticipates an increase
in business. A consultant is hired and asked to look at the following options:
a. Do nothing.
b. Add more cashiers (checkout counters).
c. Install a new procedure (bar code reader) that reduces the checkout time.
d. Increase the number of shopping carts.
TABLE L9.2 Shopping and Checkout Times for Customers at Apna Bazaar
Type of Customer Shopping Time Checkout Time
Regular Normal(12,3) min. Normal(5,1) min.
har01307_LB09_549-568.indd 558 12/17/10 10:16 AM

e. Increase the frequency of customer arrivals.

f. Make a combination of some of the above changes.
The consultant identified four independent (or input) variables—number of shop-
ping carts, number of checkout counters, checkout time, and frequency of cus-
tomer arrivals. He set up a full factorial experiment with two levels of each input
factor to investigate the effects of each of these factors and their interactions on
the time a customer spends in the store (time in system).
The high and low levels of these input factors were selected as shown in
Table L9.3. A snapshot of the simulation model of the Apna Bazaar supermar-
ket is shown in Figure L9.5. An attribute Time_In is defined in the simulation
model and a Log function is used to keep track of the time customers spend in
the grocery store (Lab 10, section L10.2). Tables L9.4 and L9.5 shows the full
factorial design matrix and the resulting time in system data for five replications.
Figure L9.6 shows a Pareto chart for the absolute values of the half-effects of
all the input factors and their interactions. It is interesting to note that factors D
(customer arrival frequency), A (number of shopping carts), and the AD interac-
tion are most important in having an effect on the customer time in the system
TABLE L9.3 Input Factors and their Levels
Factors Factor Description High Level Low Level
A Number of Shopping carts 10 5

B Number of Checkout Counters 6 2
C Checkout Time N(5,1) min. N(2,1) min.
D Customer Arrival Freq. 30/hr. 10/hr.
FIGURE L9.5
A snapshot of the Apna
Bazaar supermarket
Simulation Model.
har01307_LB09_549-568.indd 559 12/17/10 10:16 AM

560 Part II Labs
TABLE L9.4 Full Factorial Experiment Design Matrix
Runs A B C D AB AC AD BC BD CD ABC ABD BCD ACD ABCD
1 −1 −1 −1 −1 1 1 1 1 1 1 −1 −1 −1 −1 1
2 −1 −1 −1 1 1 1 −1 1 −1 −1 −1 1 1 1 −1
3 −1 −1 1 −1 1 −1 1 −1 1 −1 1 −1 1 1 −1
4 −1 −1 1 1 1 −1 −1 −1 −1 1 1 1 −1 −1 1
5 −1 1 −1 −1 −1 1 1 −1 −1 1 1 1 1 −1 −1
6 −1 1 −1 1 −1 1 −1 −1 1 −1 1 −1 −1 1 1
7 −1 1 1 −1 −1 −1 1 1 −1 −1 −1 1 −1 1 1
8 −1 1 1 1 −1 −1 −1 1 1 1 −1 −1 1 −1 −1
9 1 −1 −1 −1 −1 −1 −1 1 1 1 1 1 −1 1 −1
10 1 −1 −1 1 −1 −1 1 1 −1 −1 1 −1 1 −1 1
11 1 −1 1 −1 −1 1 −1 −1 1 −1 −1 1 1 −1 1
12 1 −1 1 1 −1 1 1 −1 −1 1 −1 −1 −1 1 −1
13 1 1 −1 −1 1 −1 −1 −1 −1 1 −1 −1 1 1 1
14 1 1 −1 1 1 −1 1 −1 1 −1 −1 1 −1 −1 −1
15 1 1 1 −1 1 1 −1 1 −1 −1 1 −1 −1 −1 −1
16 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
TABLE L9.5 Results of the Full Factorial Experiment
Runs Replications—Time in System (min) Average Std. Dev.
1 15.55 15.81 15.95 15.84 15.74 15.78 0.15

2 1095.18 1063.45 1047.33 1010.99 1055.75 1054.54 30.34
3 19.43 20.02 20.25 19.96 19.64 19.86 0.32
4 1393.02 1425.18 1329.9 1399.34 1368.53 1383.19 35.97
5 15.53 15.76 15.97 15.64 15.73 15.73 0.16
6 1070.56 1084.67 1078.01 1005.89 992.54 1046.33 43.56
7 19.2 19.82 19.52 19.33 19.05 19.38 0.30
8 1379.92 1381.4 1324.76 1402.07 1374.98 1372.63 28.70
9 15.19 15.23 15.36 15.28 15.21 15.25 0.07
10 17.27 16.85 16.57 16.92 16.61 16.84 0.28
11 18.97 18.9 18.94 18.79 18.7 18.86 0.11
12 548.9 599.58 662.99 577.47 591.62 596.11 42.08
13 15.14 15.22 15.31 15.27 15.21 15.23 0.06
14 16.68 16.65 16.19 15.83 16.43 16.36 0.35
15 18.18 18.2 18.32 18.27 18.18 18.23 0.06
16 24.33 27.47 23.56 23.6 24.76 24.74 1.61
(grocery store). An analysis of variance (ANOVA) was performed to determine

which factor effects and/or interactions are significant. The ANOVA results are
shown in Table L9.6.
A summary of the full factorial experiment is shown in Table L9.4
har01307_LB09_549-568.indd 560 12/17/10 10:16 AM

FIGURE L9.6
Pareto Chart of
Half-Effects of
Input Factors—Full
Factorial Design
TABLE L9.6 ANOVA Results for the Full Factorial Experiment
Input Significant
Factors Abs(delta) MSB Feffect @95%
D 671.55 9019588 21512.64 Yes

A 525.73 5527841 13184.46 Yes
AD 524.93 5511030 13144.37 Yes
C 157.12 493734 1177.606 Yes
CD 153.53 471429 1124.407 Yes
B 73.98 109461 261.0752 Yes
BD 73.68 108575 258.9621 Yes
BC 71.78 103047 245.7785 Yes
BCD 71.53 102331 244.0695 Yes
ABC 71.09 101076 241.076 Yes
ABCD 71.04 100934 240.737 Yes
AB 69.15 95634.5 228.0979 Yes
ABD 69.12 95551.5 227.9001 Yes
AC 8.55 1462.05 3.487139 NO
ACD 8.27 1367.86 3.262481 NO
MSE 419.269
Fcritical(0.95,1,64) 3.99
L9.5.2 Fraction Factorial Experiment

Instead of running a full factorial experiment, the consultant designed a fraction
factorial experiment based on the Taguchi design. More specifically, he designed
a half-factorial L8 experiment incorporating the factors shown in Table L9.3. The
actual design of the experiment is shown in Table L9.7. The results of the fraction
har01307_LB09_549-568.indd 561 12/17/10 10:16 AM

562 Part II Labs
TABLE L9.7 Fraction Factorial (L8) Experiment Design Matrix
Runs A B ABCD C ACBD BCAD DABC
1 1 1 1 1 1 1 1
2 1 1 1 1 1 1 1
3 1 1 1 1 1 1 1
4 1 1 1 1 1 1 1
5 1 1 1 1 1 1 1
6 1 1 1 1 1 1 1
7 1 1 1 1 1 1 1
8 1 1 1 1 1 1 1
TABLE L9.8 Results of the Fraction Factorial Experiment
Runs Replications-Time in System(min) Average Std.Dev.
1 15.67 15.85 15.25 15.42 15.4 15.52 0.24

2 1387.5 1384.27 1298.31 1391.9 1389.44 1370.28 40.33
3 1028.32 967.1 912.35 1022.7 1013.86 988.87 49.13
4 19.62 19.76 18.71 18.94 18.87 19.18 0.48
5 16.88 16.21 16.53 16.37 16.51 16.50 0.25
6 18.67 18.79 18.3 18.53 18.46 18.56 0.17
7 15.11 15.12 14.87 15.04 15.1 15.05 0.10
8 24.67 23.44 22.51 24.28 22.62 23.50 0.97
factorial experiment are shown in Table L9.8. Figure L9.7 shows a Pareto chart
for the absolute values of the half-effects of all the input factors and their inter-
actions. It is interesting to note that factors D (customer arrival frequency), A
(number of shopping carts), and the AD interaction are most important in having
an effect on the customer time in the system (grocery store). An analysis of vari-
ance (ANOVA) was performed to determine which factor effects and/or interac-
tions are significant. The ANOVA results for the fraction factorial experiment
are shown in Table L9.9. The analysis of variance confirms what we had already
concluded from observing the Pareto chart. Factors D, A, and the AD interaction
are significant at 0.95 percent. Factors C and B and interactions AB, CD, AC,
and BD are all insignificant. We have arrived at very similar conclusions using a
fraction factorial design with only half the effort (8 runs
5 replications each
40 experiments) what we had concluded after running the full-factorial experi-
ment (16 runs
5 replications each 80 experiments).
har01307_LB09_549-568.indd 562 12/17/10 10:16 AM

FIGURE L9.7
Pareto Chart of
Half-Effects of Input
Factors—Fraction
Factorial Design.
TABLE L9.9 ANOVA Results for the Fraction Factorial Experiment
Input Factors Abs(delta) MSB Feffect Significant
DABC 582.71 3395509 105.0185 Yes

BCAD 581.11 3376888 104.4426 Yes
A 580.06 3364696 104.0655 Yes
C 98.9 97812.1 3.025196 No
ABCD 95.31 90839.96 2.809557 No
ACBD 93.64 87684.5 2.711962 No
B 93.57 87553.45 2.707909 No
MSE 32332.49
Fcritical(0.95,1,32) 4.149
L9.6 Exercises
1. Without using the Common Random Numbers (CRN) technique,
compare the Spuds-n-More1, Spuds-n-More2, and Spuds-n-More3
models with respect to the average number of customers served per
day. To avoid the use of CRN, assign ProModel Stream 1 to each
stochastic element in the model and do not select the CRN option from
the Simulation Options menu. Simulate each of the three restaurants for
25 replications.
a. Use the Bonferroni approach with an overall significance level of
0.06, and construct the confidence intervals using the paired-t method.
How do the half-widths of the confidence intervals compare to the
ones computed in section L9.4? Are the differences in half-widths
what you expected? Explain.
har01307_LB09_549-568.indd 563 12/17/10 10:16 AM

564 Part II Labs
b. Use the Analysis of Variance (ANOVA) technique with the Fisher’s

protected Least Significant Difference (LSD) test to compare the
three systems. Use a 0.05 significance level. Explain any differences
between your conclusions here and those made in section L9.4 about
the three systems.
2. Increase the capacity of the Pickup_Q location in the Spuds-n-More3
model from three to six. Call this model Spuds-n-More4. Using the
Common Random Numbers (CRN) technique with ProModel Streams 1,
2, 3, and 4 assigned as in Figure L9.3, see if Spuds-n-More4 processes
more customers per day on average than does Spuds-n-More3. Use a
paired-t confidence interval with a 0.02 significance level. Run each
simulation model for 25 replications.
3. For the Picayune Mutual Bank (Exercise 1 of section L8.4), run 15 eight-
hour replications of each alternative system design with the objective of
identifying which design minimizes the average time customers spend in
the bank. Use the Welch confidence interval method with a 0.10 level of
significance to make the comparison. Contrast this comparison with the
ballpark guess you made in section L8.4.
4. The DumpOnMe facility presented in Exercise 11 of Lab section L4.13
operates with six dump trucks. The owner of the facility wants to
increase the amount of coal hauled to the railroad. The loader operators
recommend adding two more trucks for a total of eight trucks. Not to
be outdone, the scale operator recommends running the operation with
10 trucks. Trucks are expensive, so the owner hires you to simulate
the system as is and then with the additional trucks before making her
decision. Since each truck hauls approximately the same amount of coal,
you will determine which alternative delivers the most coal to the railroad
by evaluating the number of trucks processed by the scale. To compare
the performance of the 6-, 8-, and 10-truck systems, run five replications
of each alternative system using a 100-minute warm-up time followed
by a 480-minute simulation run length beyond the warm-up time. Use
the Welch confidence interval method with a 0.06 level of significance to
make the comparison. Who is correct, the loader operators or the scale
operator? Are you a little surprised by the answer? Look over the output
reports from the simulations to find an explanation for the answer.
5. For the Manufacturing Costing Optimization model (Mfg_Cost.mod in
demos subdirectory) design a fraction factorial experiment to determine
the effects of the input factors (shown in Table L9.10) and interactions
on the average cycle time. Design a fraction factorial experiment of
16 runs (1/4 factorial) and 5 replications for each run. Draw a Pareto
chart for the absolute values of the half-effects (|delta/2|). What can
you conclude from the Pareto Chart? Complete an analysis of variance.
Which factors and factor interactions are significant (95%)? Determine a
prediction equation for the estimated time in system, Ŷ.
har01307_LB09_549-568.indd 564 12/17/10 10:16 AM

TABLE L9.10 Input Factors for the Manufacturing Costing Optimization

Model Experiment
Input Factors Low Level High Level
Number of operators 1 3
Capacity of Receive Location 2 4
Capacity of Loc1 1 5
Capacity of Degrease 2 4
Bearing Queue 20 100
Number of Pallets 2 6
TABLE L9.11 Input Factors for the Tube Distribution Supply Chain Model
Experiment
Capacity of Scrapln 5 15
Capacity of FurnaceQ 5 10
Capacity of Furnace 5 10
Capacity of Insp1 5 15
Number of Units—Transport1 1 3
6. In the Tube Distribution Supply Chain model (Logistics.mod in demos

subdirectory), design a fraction factorial experiment with the input
factors shown in Table L9.11. Design a fraction factorial experiment
of 16 runs, 5 replications each. Each simulation experiment should be
run for 100 hours. Determine the effects of the input factors and factor
interactions on the average time in system for the entity Pipe. Draw a
Pareto chart for the absolute values of the half-effects (|delta/2|). What
can you conclude from the Pareto Chart? Complete an analysis of
variance. Which factors and factor interactions are significant (95%)?
Determine a prediction equation for the estimated time in system, Ŷ.
7. In the Order Processing Simulation model (Orders.mod in the demos
subdirectory) for the High Staffing Level scenario we would like to vary
some of the input factors and determine their effects on the average
process time. The factors identified to be included in the experiment and
the levels for these factors are as shown in Table L9.12.
har01307_LB09_549-568.indd 565 12/17/10 10:16 AM

566 Part II Labs
TABLE L9.12 Input Factors for the Order Processing Simulation Model
Experiment
Number of Forklifts in Shipping 3 5

Number of Forklifts in Picking 2 6
Number of Dock Workers 3 5
Number of Order Pickers 4 9
Number of Floor Spotters 3 5
TABLE L9.13 Input Factors for the SMPhase6 Model Experiment
Staff Cross Trained 1 No 2 Yes

Capacity of the Waiting Area 5 10
Number of Secretary 1 2
Number of Loan Officers 1 3
Number of Auditors 1 3
Number of Service Persons 1 3
Design a fraction factorial experiment (half factorial) of 16 runs,

5 replications each. Each simulation replication should be run for
100 hours. Determine the effects of the input factors and factor
interactions on the average time in the system. Draw a Pareto chart for
the absolute values of the half-effects (|delta/2|). What can you conclude
from the Pareto Chart? Complete an analysis of variance. Which factors
and factor interactions are significant (95%)? Determine a prediction
equation for the estimated time in system, Ŷ.
8. For the SMPhase6 model in the Training subdirectory, design a fraction
factorial experiment that includes the factors shown in Table L9.13.
Design a fraction factorial experiment (half factorial) of 16 runs,
5 replications each. Each simulation replication should be run for
100 hours.
Determine the effects of the input factors and factor interactions
on the average time in the system. Draw a Pareto chart for the absolute
values of the half-effects (|delta/2|). What can you conclude from the
Pareto Chart? Complete an analysis of variance. Which factors and factor
interactions are significant (95%)? Determine a prediction equation for
the estimated time in system, Ŷ.
har01307_LB09_549-568.indd 566 12/17/10 10:16 AM

TABLE L9.14 Input Factors for Drawings R Us Fraction Factorial Design
Capacity of Rework 1 4
Design—Number of Units 1 3
Review—Number of Units 1 3
Produce Drawings—Number of Units 1 3
Production Bob 1 3
9. For the Drawings R Us simulation model in the Training subdirectory,

design a fraction factorial experiment that includes the factors (all at two
levels) shown in Table L9.14. Determine the effects of the input factors
and factor interactions on the average time in the system. Draw a Pareto
chart for the absolute values of the half-effects (ldelta/2l). What can
you conclude from the Pareto Chart? Complete an analysis of variance.
Which factors and factor interactions are significant (95%)? Determine a
prediction equation for the estimated time in system, Ŷ.
har01307_LB09_549-568.indd 567 12/17/10 10:16 AM

har01307_LB09_549-568.indd 568 12/17/10 10:16 AM
L A B
10 SIMULATION OPTIMIZATION
WITH SIMRUNNER
Climb mountains to see lowlands.

—Chinese Proverb
The purpose of this lab is to demonstrate how to solve simulation-based optimi-

zation problems using SimRunner. The lab introduces the five major steps for
formulating and solving optimization problems with SimRunner. After stepping
through an example application of SimRunner, we provide additional application
scenarios to help you gain experience using the software.
L10.1 Introduction to SimRunner

When you conduct an analysis using SimRunner, you build and run projects. With
each project, SimRunner applies its evolutionary algorithms to your simulation
model to seek optimal values for multiple decision variables. In SimRunner, de-
cision variables are called input factors (Figure L10.1). For each project, you
will need to give SimRunner a model to optimize, identify which input factors
to change, and define how to measure system performance using an objective
function. The following describes the terminology and procedure used to conduct
experiments using SimRunner.
Step 1. Create, verify, and validate a simulation model using ProModel,

MedModel, or ServiceModel. Next, create a macro and include it in the run-
time interface for each input factor that is believed to influence the output of
the simulation model. The input factors are the variables for which you are
seeking optimal values, such as the number of nurses assigned to a shift or
the number of machines to be placed in a work cell. Note that SimRunner
569
har01307_LB10_569-594.indd 569 12/17/10 10:17 AM

570 Part II Labs
FIGURE L10.1
SimRunner’s
optimization Optimization
Simulation
algorithms and
ProModel simulation
model.
Algorithms
Optimization
Input Factors
Response (Objective Function)
can test only those factors identified as macros in ProModel, MedModel, or

ServiceModel.
Step 2. Create a new SimRunner project and select the input factors you wish
to test. For each input factor, define its numeric data type (integer or real) and its
lower bound (lowest possible value) and upper bound (highest possible value).
SimRunner will generate solutions by varying the values of the input factors ac-
cording to their data type, lower bounds, and upper bounds. Care should be taken
when defining the lower and upper bounds of the input factors to ensure that a
combination of values will not be created that leads to a solution that was not
envisioned when the model was built.
Step 3. After selecting the input factors, define an objective function to measure
the utility of the solutions tested by SimRunner. The objective function is built
using terms taken from the output report generated at the end of the simulation
run. For example, the objective function could be based on entity statistics, loca-
tion statistics, resource statistics, variable statistics, and so forth. In designing the
objective function, the user specifies whether a term is to be minimized or maxi-
mized as well as the overall weighting of that term in the objective function. Some
terms may be more important than other terms to the decision maker. SimRunner
also allows you to seek a target value for an objective function term.
Step 4. Select the optimization profile and begin the search by starting the
optimization algorithms. The optimization profile sets the size of the evolution-
ary algorithm’s population. The population size defines the number of solutions
har01307_LB10_569-594.indd 570 12/17/10 10:17 AM

FIGURE L10.2
Generally, the larger the size of the population the better the result.
Max f(x)
Cautious
Moderate
Aggressive
Time
evaluated by the algorithm during each generation of its search. SimRunner

provides three population sizes: small, medium, and large. The small popula-
tion size corresponds to the aggressive optimization profile, the medium pop-
ulation size corresponds to the moderate optimization profile, and the large
population size corresponds to the cautious profile. In general, as the population
size is increased, the likelihood that SimRunner will find the optimal solution
increases, as does the time required to conduct the search (Figure L10.2).
Step 5. Study the top solutions found by SimRunner and pick the best. SimRunner
will show the user the data from all experiments conducted and will rank each so-
lution based on its utility, as measured by the objective function. Remember that
the value of an objective function is a random variable because it is produced from
the output of a stochastic simulation model. Therefore, be sure that each experi-
ment is replicated an appropriate number of times during the optimization.
Another point to keep in mind is that the list of solutions presented by Sim-
Runner represents a rich source of information about the behavior, or response
surface, of the simulation model. SimRunner can sort and graph the solutions
many different ways to help you interpret the “meaning” of the data.
har01307_LB10_569-594.indd 571 12/17/10 10:17 AM

572 Part II Labs
L10.2 SimRunner Projects

Problem Statement
Prosperity Company has selected what it thinks is the ideal product to manu-
facture and has also designed the “ideal production system” (see Figure L10.3).
Plates of raw material arrive to the ideal production system and are transformed
into gears by a milling operation. The time between arrivals of plates is exponen-
tially distributed, as is the processing time at the milling machine. Plates are pro-
cessed in a first-in, first-out (FIFO) fashion. The time to move material between
the pallets and the milling machine is negligible. The input and output queues
have infinite capacities and the milling machine never fails or requires mainte-
nance. It is the ideal production system. However, we have been asked to look for
optimal operational parameters under three different scenarios.
We begin by building a simulation model of the ideal production system.
Set the default time units to minutes in the General Information dialog box.
The model consists of three locations: InputPalletQueue, MillingMachine, and
OutputPalletQueue. Set the capacity of the InputPalletQueue to infinity (Inf) and
the capacity of the milling machine and OutputPalletQueue to one. For simplic-
ity, use a single entity type to represent both plates and gears. Assign Gear as the
name of the entity. The parameters for the exponentially distributed time between
arrivals and processing time will be given later. For now, set the model’s run hours
to 250 and warm-up hours to 50. The complete model is shown in Figure L10.4.
The model is found in the Student folder installed with ProModel under file name
Lab 10_2 ProsperityCo.Mod.
Before continuing, we would like to point out that this fictitious production
system was chosen for its simplicity. The system is not complex, nor are the
example application scenarios that follow. This deliberate choice will allow us
FIGURE L10.3
The ideal production Ideal Production System
system for Prosperity
Company. Milling Machine
Input Queue Output Queue
• Infinite queue capacity

• Raw materials are plentiful
• Milling Machine never fails
har01307_LB10_569-594.indd 572 12/17/10 10:17 AM

FIGURE L10.4
ProModel model of Prosperity Company.
har01307_LB10_569-594.indd 573 12/17/10 10:17 AM

574 Part II Labs
to focus on learning about the SimRunner software as opposed to getting bogged

down in modeling details. Additionally, the problems that are presented are easily
solved using queuing theory. Therefore, if so inclined, you may want to compare
the estimates obtained by SimRunner and ProModel with the actual values ob-
tained using queuing theory. In the real world, we seldom have such opportuni-
ties. We can take refuge in knowing that ProModel and SimRunner are robust
tools that can be effectively applied to both trivial and complex real-world prob-
lems, as was demonstrated in Chapter 10.
L10.2.1 Single Term Objective Functions

In the first scenario, the average time between arrivals of a plate of raw material to
the input pallet queue is E(3.0) minutes. For this scenario, our objective is to find
a value for the mean processing time at the milling machine that minimizes the
average number of plates waiting at the input pallet queue. (We have a great deal
of latitude when it comes to adjusting processing times for the ideal production
system.) To do this, we will create a macro in ProModel that represents the mean
processing time and give it an identification of ProcessTime. Thus, the processing
time for entities at the milling machine is modeled as E(ProcessTime) minutes.
See Operation field for the MillingMachine location in Figure L10.4.
Before continuing, let’s cheat by taking a moment to look at how varying the
mean processing time affects the mean number of plates waiting in the input pal-
let queue. The plot would resemble the one appearing in Figure L10.5. Therefore,
FIGURE L10.5
the mean processing Response Surface
time and the mean
number of entities
waiting in the queue 100
given a mean time 90 Location (Queue) Contents
between arrivals of 80
three.
No. in Queue
70
60
50
40
30
20
10
0
-10
0 1.5 3
Process Time
har01307_LB10_569-594.indd 574 12/17/10 10:17 AM

FIGURE L10.6
ProModel macro editor.
we can minimize the average number of plates waiting in the queue by setting the
mean processing time of the milling machine to zero minutes, which of course is a
theoretical value. For complex systems, you would not normally know the answer
in advance, but it will be fun to see how SimRunner moves through this known
response surface as it seeks the optimal solution.
The first step in the five-step process for setting up a SimRunner project
is to define the macros and their Run-Time Interface (RTI) in the simulation
model. In addition to defining ProcessTime as a macro (Figure L10.6), we shall
also define the time between arrivals (TBA) of plates to the system as a macro to
be used later in the second scenario that management has asked us to look into.
The identification for this macro is entered as TBA. Be sure to set each macro’s
“Text . . .” value as shown in Figure L10.6. The Text value is the default value
of the macro. In this case, the default value for ProcessTime is 2 and the default
value of TBA is 3. If you have difficulty creating the macros or their RTI, please
see Lab Chapter 11, section L11.1.
Next we activate SimRunner from ProModel’s Simulation menu. SimRunner
opens in the Setup Project mode (Figure L10.7). The first step in the Setup Project
module is to select a model to optimize or to select an existing optimization project
har01307_LB10_569-594.indd 575 12/17/10 10:17 AM

576 Part II Labs
FIGURE L10.7
The opening SimRunner screen.
(the results of a prior SimRunner session). For this scenario, we are optimizing
a model for the first time. Launching SimRunner from the ProModel Simulation
menu will automatically load the model you are working on into SimRunner. See
the model file name loaded in the box under “Create new project—Select model”
in Figure L10.7. Note that the authors named their model ProsperityCo.Mod.
With the model loaded, the input factors and objective function are defined
to complete the Setup Project module. Before doing so, however, let’s take a mo-
ment to review SimRunner’s features and user interface. After completing the
Setup Project module, you would next run either the Analyze Model module or
the Optimize Model module. The Analyze Model module helps you determine
the number of replications to run to estimate the expected value of performance
measures and/or to determine the end of a model’s warm-up period using the
techniques described in Chapter 8. The Optimize Model module automatically
seeks the values for the input factors that optimize the objective function using
the techniques described in Chapter 10. You can navigate through SimRunner by
har01307_LB10_569-594.indd 576 12/17/10 10:17 AM

FIGURE L10.8
Single term objective function setup for scenario one.
selecting items from the menus across the top of the window and along the left
side of the window or by clicking the <Previous or Next> buttons near the bottom
right corner of the window.
Clicking the Next> button takes you to the section for defining the objective
function. The objective function, illustrated in Figure L10.8, indicates the desire to
minimize the average contents (in this case, plates) that wait in the location called
InputPalletQue. The InputPalletQue is a location category. Therefore, to enter this
objective, we select Location from the Response Category list under Performance
Measures by clicking on Location. This will cause SimRunner to display the list
of location statistics in the Response Statistic area. Click on the response statistic
InputPalletQue—AverageContents and then press the button below with the down
arrows. This adds the statistic to the list of response statistics selected for the objec-
tive function. The default objective for each response statistic is maximize. In this
example, however, we wish to minimize the average contents of the input pallet
queue. Therefore, click on Location:Max:1*InputPalletQue—AverageContents,
har01307_LB10_569-594.indd 577 12/17/10 10:17 AM

578 Part II Labs
which appears under the area labeled Response Statistics Selected for the Objec-
tive Function; change the objective for the response statistic to Min; and click the
Update button. Note that we accepted the default value of one for the weight of
the factor. Please refer to the SimRunner Users Guide if you have difficulty per-
forming this step.
Clicking the Next> button takes you to the section for defining the input
factors. The list of possible input factors (macros) to optimize is displayed at the
top of this section under Macros Available for Input (Figure L10.9). The input
factor to be optimized in this scenario is the mean processing time of the milling
machine, ProcessTime. Select this macro by clicking on it and then clicking the
button below with the down arrows. This moves the ProcessTime macro to the
list of Macros Selected as Input Factors (Figure L10.9). Next, indicate that you
want to consider integer values between one and five for the ProcessTime macro.
Ignore the default value of 2.00. If you wish to change the data type or lower and
upper bounds, click on the input factor, make the desired changes, and click the
Update button. Please note that an input factor is designated as an integer when
FIGURE L10.9
Single input factor setup for scenario one.
har01307_LB10_569-594.indd 578 12/17/10 10:17 AM

the lower and upper bounds appear without a decimal point in the Macros proper-
ties section. When complete, SimRunner should look like Figure L10.9.
From here, you click the Next> button until you enter the Optimize Model
module, or click on the Optimize Model module button near the top right corner
of the window to go directly to it. The first step here is to specify Optimization
options (Figure L10.10). Select the Aggressive Optimization Profile. Accept the
default value of 0.01 for Convergence Percentage.
The convergence percentage controls how long SimRunner’s optimization
algorithms will run experiments before stopping. With each experiment, SimRunner
records the objective function’s value for a solution in the population. The evalu-
ation of all solutions in the population marks the completion of a generation. At
the end of a generation, SimRunner computes the population’s average objec-
tive function value and compares it with the population’s best (highest) objective
function value. When the best and the average are at or near the same value at
FIGURE L10.10
Optimization and simulation options.
har01307_LB10_569-594.indd 579 12/17/10 10:17 AM

580 Part II Labs
the end of a generation, all the solutions in the population are beginning to look
alike (their input factors are converging to the same setting). It is difficult for the
algorithms to locate a better solution to the problem once the population of so-
lutions has converged. Therefore, the optimization algorithm’s search is usually
terminated at this point.
The convergence percentage controls how close the best and the average must
be to each other before the optimization stops. A convergence percentage near
zero means that the average and the best must be nearly equal before the optimiza-
tion stops. A high percentage value will stop the search early, while a very small
percentage value will run the optimization longer.
After you specify the optimization options, set the simulation options. Typi-
cally you will want to disable the animation as shown in Figure L10.10 to make the
simulation run faster. Usually you will want to run more than one replication to es-
timate the expected value of the objective function for a solution in the population.
When more than one replication is specified, SimRunner will display the objective
function’s confidence interval for each solution it evaluates. Note that the confi-
dence level for the confidence interval is specified here. Confidence intervals can
help you to make better decisions at the end of an optimization as discussed in sec-
tion 10.6.2 of Chapter 10. In this case, however, use one replication to speed things
along so that you can continue learning other features of the SimRunner software.
As an exercise, you should revisit the problem and determine an acceptable number
of replications to run per experiment. As indicated in Figure L10.10, set the simula-
tion warm-up time to 50 hours and the simulation run time to 250 hours. You are
now ready to have SimRunner seek the optimal solution to the problem.
With these operations completed, click the Next> button (Figure L10.10) and
then click the Run button on the Optimize Model module (Figure L10.11) to start
the optimization. For this scenario, SimRunner runs all possible experiments,
locating the optimum processing time of one minute on its third experiment. The
Experimental Results table shown in Figure L10.11 records the history of Sim-
Runner’s search. The first solution SimRunner evaluated called for a mean process-
ing time at the milling machine of three minutes. The second solution evaluated
assigned a processing time of two minutes. These sequence numbers are recorded
in the Experiment column, and the values for the processing time (ProcessTime)
input factor are recorded in the ProcessTime column. The value of the term used
to define the objective function (minimize the mean number of plates waiting in
the input pallet queue) is recorded in the InputPalletQue:AverageContents col-
umn. This value is taken from the output report generated at the end of a simula-
tion. Therefore, for the third experiment, we can see that setting the ProcessTime
macro equal to one results in an average of 0.162 plates waiting in the input
pallet queue. If you were to conduct this experiment manually with ProModel,
you would set the ProcessTime macro to one, run the simulation, display output
results at the end of the run, and read the average contents for the InputPalletQue
location from the report. You may want to verify this as an exercise.
Because the objective function was to minimize the mean num-
ber of plates waiting in the input pallet queue, the same values from the
har01307_LB10_569-594.indd 580 12/17/10 10:17 AM

FIGURE L10.11
Experimental results table for scenario one.
InputPalletQue:AverageContents column also appear in the Objective Func-

tion column. However, notice that the values in the Objective Function column
are preceded by a negative sign (Figure L10.11). This has to do with the way
SimRunner treats a minimization objective. SimRunner’s optimization algorithms
view all problems as maximization problems. Therefore, if we want to minimize
a term called Contents in an objective function, SimRunner multiplies the term
by a negative one {(−1)Contents}. Thus SimRunner seeks the minimal value by
seeking the maximum negative value. Figure L10.12 illustrates this for the ideal
production system’s response surface.
Figure L10.13 illustrates SimRunner’s Performance Measures Plot for this
optimization project. The darker colored line (which appears red on the computer
screen) at the top of the Performance Measures Plot represents the best value of
the objective function found by SimRunner as it seeks the optimum. The lighter
colored line (which appears green on the computer screen) represents the value of
the objective function for all of the solutions that SimRunner tried.
har01307_LB10_569-594.indd 581 12/17/10 10:17 AM

582 Part II Labs
FIGURE L10.12
SimRunner’s process
for converting 100
minimization problems 90 Location (Queue) Contents
80
to maximization 70
problems. 60
50
1 x Contents
40
30
20
10
0
-10
0 1.5 3
0
-50 -1 x Contents
-100
-110
0 1.5 3
Process Time
FIGURE L10.13
SimRunner’s
Performance Plot
indicates the progress
of the optimization for
scenario one.
The last menu item of the Optimize Model module is the Response Plot (Fig-
ure L10.11), which is a plot of the model’s output response surface based on the
solutions evaluated during the search. We will skip this feature for now and cover
it at the end of the lab chapter.
har01307_LB10_569-594.indd 582 12/17/10 10:17 AM

L10.2.2 Multiterm Objective Functions

Objective functions may be composed of any number of terms taken from the
output of a simulation model. This application scenario demonstrates the con-
struction of an objective function with two terms.
The managers of the ideal production system presented in section L10.2.1
have a not-so-ideal objective for the system. They have conflicting objectives of
(1) maximizing the number of gears produced and (2) minimizing the amount of
space allocated for storing work-in-process at the input pallet queue area. Because
space has to be available for the maximum number of plates that could wait in the
queue, the second objective can be restated as minimizing the maximum number of
plates waiting at the input pallet queue. In attempting to satisfy the managers’ objec-
tives, both the mean processing time at the milling machine (ProcessTime) and the
mean time between arrivals (TBA) of plates to the input pallet queue can be varied.
Each of these input factors can be assigned integer values between one and five min-
utes. To allow SimRunner to change the mean time between arrivals (TBA) in the
simulation model, enter E(TBA) in the Frequency column of the model’s Arrivals
table (Figure L10.4). Remember that you previously defined TBA as a macro.
The SimRunner objective function for this scenario is shown in Figure L10.14.
The first segment of the objective function can be implemented using the output
response statistic that records the total number of gear entities that exit the system
(Gear:TotalExits), which is an Entity response category. The second segment is
implemented using the output response statistic that records the maximum number
of plate entities that occupied the input pallet queue location during the simulation
(InputPalletQue:MaximumContents), which is a Location response category. Man-
agement has indicated that a fairly high priority should be assigned to minimizing the
space required for the input pallet queue area. Therefore, a weight of 100 is assigned
to the second term in the objective function and a weight of one is assigned to the first
term. Thus the objective function consists of the following two terms:
Maximize [(1)(Gear:TotalExits)]
and Minimize [(100)(InputPalletQue:Maximum Contents)]
SimRunner minimizes terms by first multiplying each minimization term
appearing in the objective function by a negative one, as explained in Sec-
tion L10.2.1. Similarly, SimRunner multiplies each maximization term by a positive
one. Next the terms are arranged into a linear combination as follows:
(1)[(1)(Gear:TotalExits)] (1)[(100)(InputPalletQue:MaximumContents)]
which reduces to
[(1)(Gear:TotalExits)] [(100)(InputPalletQue:MaximumContents)]
Given that SimRunner’s optimization algorithms view all problems as maxi-
mization problems, the objective function F becomes
F Maximize {[(1)(Gear:TotalExits)]
[(100)(InputPalletQue:MaximumContents)]}
har01307_LB10_569-594.indd 583 12/17/10 10:17 AM

584 Part II Labs
FIGURE L10.14
Multiterm objective function for scenario two.
The highest reward is given to solutions that produce the largest number of gears
without allowing many plates to accumulate at the input pallet queue. In fact, a
solution is penalized by 100 points for each unit increase in the maximum number
of plates waiting in the input pallet queue. This is one way to handle competing
objectives with SimRunner.
Develop a SimRunner project using this objective function to seek the op-
timal values for the input factors (macros) TBA and ProcessTime, which are in-
tegers between one and five (Figure L10.15). Use the aggressive optimization
profile with the convergence percentage set to 0.01. To save time, specify one
replication per experiment. (Remember, you will want to run multiple replications
on real applications.) Also, set the simulation run hours to 250 and the warm-up
hours to 50 for now.
At the conclusion of the optimization, SimRunner will have run four genera-
tions as it conducted 23 experiments (Figure L10.16). What values do you recom-
mend to management for TBA and ProcessTime?
har01307_LB10_569-594.indd 584 12/17/10 10:17 AM

FIGURE L10.15
Multiple input factors setup for scenario two.
Explore how sensitive the solutions listed in the Experimental Results

table for this project are to changes in the weight assigned to the maximum
contents statistic. Change the weight of this second term in the objective func-
tion from 100 to 50. To do this, go back to the Define Objectives section of the
Setup Project module and update the weight assigned to the InputPalletQue—
MaximumContents response statistic from 100 to 50. Upon doing this,
SimRunner warns you that the action will clear the optimization data that you
just created. You can save the optimization project with the File Save option
if you wish to keep the results from the original optimization. For now, do not
worry about saving the data and click the Yes button below the warning mes-
sage. Now rerun the optimization and study the result (Figure L10.17). Notice
that a different solution is reported as optimum for the new objective function.
Running a set of preliminary experiments with SimRunner is a good way to
help fine-tune the weights assigned to terms in an objective function. Addition-
ally, you may decide to delete terms or add additional ones to better express
har01307_LB10_569-594.indd 585 12/17/10 10:17 AM

586 Part II Labs
FIGURE L10.16
Experimental results table for scenario two.
your desires. Once the objective function takes its final form, rerun the optimi-
zation with the proper number of replications.
L10.2.3 Target Range Objective Functions

The target range objective function option directs SimRunner to seek a “target”
value for the objective function term instead of a maximum or minimum value.
For example, you may wish to find an arrangement for the ideal production sys-
tem that produces from 100 to 125 gears per day. Like the maximization and mini-
mization objective options, the target range objective option can be used alone or
in combination with the maximization and minimization options.
For this application scenario, the managers of the ideal production system
have specified that the mean time to process gears through the system should
range between four and seven minutes. This time includes the time a plate waits in
the input pallet queue plus the machining time at the mill. Recall that we built the
model with a single entity type, named Gear, to represent both plates and gears.
har01307_LB10_569-594.indd 586 12/17/10 10:17 AM

FIGURE L10.17
Experimental results table for scenario two with modified objective function.
Therefore, the statistic of interest is the average time that the gear entity is in the
system. Our task is to determine values for the input factors ProcessTime and
TBA that satisfy management’s objective.
The target range objective function is represented in SimRunner as shown
in Figure L10.18. Develop a SimRunner project using this objective function to
seek the optimal values for the input factors (macros) TBA and ProcessTime.
Specify that the input factors are integers between one and five, and use the
aggressive optimization profile with the convergence percentage set to 0.01. To
save time, set the number of replications per experiment to one. (Remember,
you will want to run multiple replications on real applications.) Also, set the
simulation run hours to 250 and the warm-up hours to 50 and run the optimi-
zation. Notice that only the solutions producing a mean time in the system of
between four and seven minutes for the gear received a nonzero value for the
objective function (Figure L10.19). What values do you recommend to manage-
ment for TBA and ProcessTime?
har01307_LB10_569-594.indd 587 12/17/10 10:17 AM

588 Part II Labs
FIGURE L10.18
Target range objective function setup for scenario three.
Now plot the solutions SimRunner presented in the Experimental Results

table by selecting the Response Plot button on the Optimize Model module (Fig-
ure L10.19). Select the independent variables as shown in Figure L10.20 and
click the Update Chart button. The graph should appear similar to the one in Fig-
ure L10.20. The plot gives you an idea of the response surface for this objective
function based on the solutions that were evaluated by SimRunner. Click the Edit
Chart button to access the 3D graph controls to format the plot and to reposition
it for different views of the response surface.
L10.3 Conclusions
Sometimes it is useful to conduct a preliminary optimization project using only
one replication to help you set up the project. However, you should rarely, if ever,
make decisions based on an optimization project that used only one replication
har01307_LB10_569-594.indd 588 12/17/10 10:17 AM

FIGURE L10.19
Experimental results table with Performance Measures Plot for scenario three.
per experiment. Therefore, you will generally conduct your final project using
multiple replications. In fact, SimRunner displays a confidence interval about the
objective function when experiments are replicated more than once. Confidence
intervals indicate how accurate the estimate of the expected value of the objective
function is and can help you make better decisions, as noted in section 10.6.2 of
Chapter 10.
Even though it is easy to use SimRunner, do not fall into the trap of letting
SimRunner, or any other optimizer, become the decision maker. Study the top
solutions found by SimRunner as you might study the performance records of
different cars for a possible purchase. Kick their tires, look under their hoods, and
drive them around the block before buying. Always remember that the optimizer
is not the decision maker. SimRunner can only suggest a possible course of action.
It is your responsibility to make the final decision.
har01307_LB10_569-594.indd 589 12/17/10 10:17 AM

590 Part II Labs
FIGURE L10.20
Surface response plot for scenario three.
L10.4 Exercises
Simulation Optimization Exercises
1. Rerun the optimization project presented in section L10.2.1, setting the
number of replications to five. How do the results differ from the original
results?
2. Conduct an optimization project on the buffer allocation problem presented
in section 10.6 of Chapter 10. The model’s file name is Lab 10_4 BufferOpt
Ch10.Mod and is found in the Student folder installed with ProModel. To
get your results to appear as shown in Figure 10.5 of Chapter 10, enter
Buffer3Cap as the first input factor, Buffer2Cap as the second input factor,
and Buffer1Cap as the third input factor. For each input factor, the lower
bound is one and the upper bound is nine. The objective is to maximize
profit. Profit is computed in the model’s termination logic by
Profit (10*Throughput)
(1000*(Buffer1Cap Buffer2Cap Buffer3Cap))
har01307_LB10_569-594.indd 590 12/17/10 10:17 AM

Figure L10.21 is a printout of the model. See section 10.6.2 of Chapter 10

for additional details. Use the Aggressive optimization profile and set the
number of replications per experiment to 10. Specify a warm-up time of
240 hours, a run time of 720 hours, and a confidence level of 95 percent.
Note that the student version of SimRunner will halt at 25 experiments,
which will be before the search is completed. However, it will provide
the data necessary for answering these questions:
a. How do the results differ from those presented in Chapter 10 when
only five replications were run per experiment?
b. Are the half-widths of the confidence intervals narrower?
c. Do you think that the better estimates obtained by using 10 replications
will make it more likely that SimRunner will find the true optimal
solution?
3. In Exercise 4 of Lab section L9.6, you increased the amount of coal
delivered to the railroad by the DumpOnMe facility by adding more
dump trucks to the system. Your solution received high praise from
everyone but the lead engineer at the facility. He is concerned about
the maintenance needs for the scale because it is now consistently
operated in excess of 90 percent. A breakdown of the scale will incur
substantial repair costs and loss of profit due to reduced coal deliveries.
He wants to know the number of trucks needed at the facility to achieve
a target scale utilization of between 70 percent and 75 percent. This
will allow time for proper preventive maintenance on the scale. Add a
macro to the simulation model to control the number of dump trucks
circulating in the system. Use the macro in the Arrivals Table to specify
the number of dump trucks that are placed into the system at the start of
each simulation. In SimRunner, select the macro as an input factor and
assign it a lower bound of one and an upper bound of 15. Conduct an
optimization project to seek the number of trucks that will achieve the
target scale utilization using the Aggressive optimization profile, five
replications per experiment, a warm-up time of 100 minutes, and a run
time of 480 minutes.
SimRunner Warm-up Detection Exercises

SimRunner’s Analyze Model module implements the Welch moving average
technique presented in section 8.6.1 of Chapter 8 to help you determine the end
of a nonterminating simulation’s warm-up phase before beginning an optimiza-
tion project. It also helps you evaluate the number of replications needed to ob-
tain a point estimate to within a specified percentage error and confidence level.
Although the module was created to help you set up an optimization project, it is
also useful for nonoptimization projects. To use the module without optimization,
you declare a dummy macro in your simulation model and select it as an input
factor in SimRunner. Then you select the output statistic that you wish to use
in order to evaluate the end of the simulation’s warm-up. The output statistic is
har01307_LB10_569-594.indd 591 12/17/10 10:17 AM

FIGURE L10.21
Buffer allocation model from Chapter 10 (Section 10.6.2).
har01307_LB10_569-594.indd 592 12/17/10 10:17 AM

entered as an objective function term. Exercises 4 and 5 here involve this feature
of SimRunner.
4. This exercise will help you duplicate the SimRunner result presented
in Figure L8.14 of Lab Chapter 8, which was used to estimate the end
of the warm-up phase for the Green Machine Manufacturing Company
(GMMC) simulation model. Load the GMMC model into ProModel,
declare a dummy macro, and define its Run-Time Interface (RTI). Start
SimRunner. The purpose of the GMMC model was to estimate the
steady-state value of the time-average amount of work-in-process (WIP)
inventory in the system. Therefore, select the WIP—Average Value
response statistic from the Variable response category as a SimRunner
maximization objective. Select the dummy macro as an input factor.
Click the Next> button to move into the Analyze Model module, and fill
in the parameters as shown in Figure L10.22.
Percentage error in the objective function estimate is the desired
amount of relative error in our average WIP inventory estimate expressed
as a percentage (see section 8.2.3 of Chapter 8 for additional details on
relative error). In this case, we are seeking to approximate the number
of replications needed to estimate the average WIP with a percentage
error of 7 percent and a confidence level of 90 percent as we also
FIGURE L10.22
SimRunner parameters
for the GMMC
warm-up example.
har01307_LB10_569-594.indd 593 12/17/10 10:17 AM

594 Part II Labs
estimate the end of the warm-up phase. Click the Next> button and then
the Run button on the Conduct Analysis window to start the analysis.
After SimRunner runs the simulation for five replications, your screen
should appear similar to Figure L8.14. Here you adjust the number of
periods for the moving average window to help you identify the end
of simulation’s warm-up, which seems to occur between periods 33
and 100. SimRunner computes that at least 10 replications are needed
to estimate the average WIP inventory with a 7 percent error and a
confidence level of 90 percent assuming a 100-period (hour) warm-up.
5. Use SimRunner’s Analyze Model module to determine the end of the
warm-up phase of the DumpOnMe simulation model with six dump
trucks as presented in Exercise 11 of Lab section L4.13. Base your
assessment of the warm-up phase on five replications of the simulation
with a run time of 300 minutes, and an output recording time interval of
one minute.
har01307_LB10_569-594.indd 594 12/17/10 10:17 AM

L A B
11 MODELING MANUFACTURING
SYSTEMS
All things good to know are difficult to learn.

—Greek Proverb
In this lab we discuss models of various manufacturing systems and introduce

you to some additional features of ProModel that are very useful. In section L11.1
we introduce the concepts of macros and run-time interfaces using a small forg-
ing shop. In section L11.2 we show how to generate multiple scenarios for a
packaging facility. In section L11.3 we introduce a machining facility and show
how to set up and import data from external files. In section L11.4 model a small
jobshop and discuss arrays of data. Subroutines are explained with the example
of a gear fabrication and inspection facility in section L11.5. Section L11.6 we
model preventive maintenance in a machine shop and introduce the concepts
and use of random number streams. In section L11.7 we show how to merge the
models of two manufacturing facilities and introduce the concept of merging
submodules.
L11.1 Macros and Run-time Interface

If text, a set of statements, or a block of code is used many times in the model, it is
convenient to substitute a macro and use the macro as many times as needed. As
will be seen in section L11.5, subroutines are preferable to macros when defining
a block of code that is used multiple places, but macros have the advantage that
they can be used to represent individual numeric expressions and objects as well
as entire statements. Macros are created and edited in the following manner:
1. Select More Elements from the Build menu.
2. Select Macros.
595
har01307_LB11_595-616.indd 595 20/12/10 1:57 PM

596 Part II Labs
Macros are also used to define a run-time interface. A run-time interface (RTI)
allows the user to change model parameters without directly editing the model itself.
Every time the simulation is run, the RTI allows the user to change model parameters
defined in the RTI. If you recall the Manufacturing Cost model from Lab 1, it had a
run-time interface defined for it that allowed you to select the number of operators to
use in the model. The RTI provides a user-friendly menu to change only the macros
that the modeler wants the user to change. An RTI is a custom interface defined by a
modeler that allows others to modify the model or conduct multiple-scenario experi-
ments without altering the actual model data. All changes are saved along with the
model so they are preserved from run to run. RTI parameters are based on macros,
so they may be used to change any model parameter that can be defined using a
macro (that is, any field that allows an expression or any logic definition).
An RTI is created and used in the following manner:
1. Select Macros from the Build menu and type in a macro ID.
2. Click the RTI button and choose Define from the submenu. This opens
the RTI Definition dialog box.
3. Enter the Parameter Name that you want the macro to represent.
4. Enter an optional Prompt to appear for editing this model parameter.
5. Select the parameter type, either Unrestricted Text or Numeric Range.
6. For numeric parameters:
a. Enter the lower value in the From box.
b. Enter the upper value in the To box.
7. Click OK.
8. Enter the default text or numeric value in the Macro Text field.
9. Use the macro ID in the model to refer to the run-time parameter (such
as operation time or resource usage time) in the model.
10. Before running the model, use the Model Parameters dialog box
(Figure L11.1) or the Scenarios dialog box to edit the RTI parameter.
FIGURE L11.1
Model Parameters
dialog box.
har01307_LB11_595-616.indd 596 20/12/10 1:57 PM

Problem Statement
Forgings R Us receives various kinds of forging orders. Raw_Forgings arrive in
batches of one every five minutes. Some forgings go from the raw material store
to the mill, and then on to the grinder. Other forgings go directly to the grinder.
After grinding, all forgings go to degrease for cleaning. Finally, all finished forg-
ings are sent to the finished parts store. The milling and grinding times vary de-
pending on the forging design. However, the degrease time is seven minutes per
forging. The layout of Forgings R Us is shown in Figure L11.2.
Define a run-time interface to allow the user to change the milling and grind-
ing times every time the simulation is run. Also, allow the user to change the total
quantity of widgets being processed. Track the work-in-process inventory (WIP)
of widgets. In addition, define another variable to track the production (PROD_QTY)
of finished widgets.
The macros are defined as shown in Figure L11.3. The run-time interface
for the milling time is shown in Figure L11.4. Figure L11.5 shows the use of the
parameters “Mill_Time” and “Grind_Time” in the processing and routing tables.
To view or change any of the model parameters, select Model Parameters from
the Simulation menu. The model parameters dialog box is shown in Figure L11.6.
To change the Grind_Time_Halfwidth, first select it from the model parameters
list, and then press the Change button. The Grind_Time_Halfwidth dialog box is
shown in Figure L11.7.
FIGURE L11.2
Layout of the Forgings
R Us.
FIGURE L11.3
Macros created
for Forgings R Us
simulation model.
har01307_LB11_595-616.indd 597 20/12/10 1:57 PM

598 Part II Labs
FIGURE L11.4
The run-time interface
defined for the milling
operation.
FIGURE L11.5
The processing and routing tables showing the Mill_Time and Grind_Time Parameters.
FIGURE L11.6
Model Parameters
dialog box.
FIGURE L11.7
The Grind_Time_
Halfwidth parameter
dialog box.
har01307_LB11_595-616.indd 598 20/12/10 1:57 PM

L11.2 Generating Scenarios

When a modeler alters one or more input parameters or elements of the model, he
or she creates model scenarios. For example, running a model first with five units
of pallets and then with 10 units to test the effect of increased resource availability
is the same thing as creating two scenarios.
A model builder has the option of defining multiple scenarios for a model
using the run-time interface. Using scenarios allows the model user to alter vari-
ous model parameters to run a series of “what-if” analyses without changing
the model directly. This technique makes it easy for analysts who have no in-
terest in dealing with the nuts and bolts of a model to make changes safely
and conveniently to the model to learn what would happen under a variety of
circumstances.
The Scenarios dialog box is accessed from the Simulation menu. Scenarios
are saved with the model for future use. They are created using the Scenarios
dialog box (Figure L11.8).
For the Shipping Boxes Unlimited example in Lab 4, section L4.8.2, gener-
ate two scenarios (Figure L11.8):
Scenario 1: Five pallets in circulation in the system.
Scenario 2: Ten pallets in circulation in the system.
Define a model parameter for the number of pallets in the system (no_of_
pallets) in the run-time interface. The RTI dialog box is obtained from the
Build More Elements Macros menu. Edit the scenarios (Figure L11.9) and
FIGURE L11.8
Scenarios defined
for Shipping Boxes
Unlimited.
FIGURE L11.9
Editing the scenario
parameter.
har01307_LB11_595-616.indd 599 20/12/10 1:57 PM

600 Part II Labs
FIGURE L11.10
Arrivals table for Shipping Boxes Unlimited.
FIGURE L11.11
Entity Summary Reports created for both scenarios.
change the parameter no_of_pallets to the values of 5 and 10, respectively, for
the two scenarios. In the Arrivals menu, change the occurrences for Pallet to the
parameter name no_of_pallets (Figure L11.10). This parameter will take on the
values of 5 and 10 in scenarios 1 and 2, respectively.
To run the scenarios, click on the Run Scenarios button in the Scenarios
dialog box. Both scenarios will be run in a sequential manner, and reports for
both scenarios created. Figure L11.11 shows the entity summary report for
both scenarios.
L11.3 External Files

External files are used to read data directly into the simulation model or to write
output data from the simulation model. External files are also used to store and
access process sheets, arrival schedules, shift schedules, bill of material, delivery
schedules, and so forth. All external files used with a model must be listed in
the External Files Editor and are accessed from the Build menu. The file format
(.XLS, .TXT, etc.) varies depending on the type of file, which can be any of the
following:
har01307_LB11_595-616.indd 600 20/12/10 1:57 PM

a. General Read File: These files contain numeric values read into a
simulation model using a READ statement. A space, comma, or end
of line can separate the data values. Any nonnumeric data is skipped
automatically. The syntax of the READ statement is as follows:
READ <file ID>, <variable name>
If the same file is to be read more than once in the model, it may be
necessary to reset the file between each reading. This can be achieved by
adding an arbitrary end-of-file marker 99999 and the following two lines
of code:
Read MydataFile1, Value1
If Value1=99999 Then Reset MydataFile1
The data stored in a general read file must be in ASCII format.
Spreadsheet programs (like Excel) can convert spreadsheets to ASCII
files (MydataFile1.TXT).
b. General Write: These files are used to write text strings and numeric
values using the WRITE, and WRITELINE statements. Text strings are
enclosed in quotes when written to the file, with commas automatically
appended to strings. This enables the files to be read into spreadsheet
programs like Excel for viewing. Write files can also be written using the
XWRITE statement, which gives the modeler full control over the output
and formatting. If you write to the same file more than once, either
during multiple replications or within a single replication, the new data
are appended to the previous data.
The WRITE statement writes to a general write file. The next item is
written to the file immediately after the previous item.
WRITE <file ID>, <string or numeric expression>
The WRITELINE statement writes information to a general write file and
starts a new line. It always appends to the file unless you have RESET the
file. Any file that is written to with the WRITE or WRITELINE statement
becomes an ASCII text file, and ProModel attaches an end-of-file marker
automatically when it closes the file.
The syntax of the WRITE, WRITELINE, and the XWRITE statements is as
follows:
WRITE MyReport, “Customer Service Completed At:”
WRITELINE MyReport, CLOCK(min)
The XWRITE statement allows the user to write in any format he or she
chooses.
XWRITE <file ID>, <string or numeric expression>
XWRITE MyReport2, “Customer Service Completed At:”
$FORMAT(Var1,5,2)
c. Entity Location: An entity-location file contains numeric expressions
listed by entity and location names. Entity names should appear across
top row, beginning in column 2, while location names should be entered
har01307_LB11_595-616.indd 601 20/12/10 1:57 PM

602 Part II Labs
TABLE L11.1 Entries in an External

Arrivals File
Column Data
A Entity Name
B Location Name
C Quantity per Arrival
D Time of First Arrival
E Number of Arrivals
F Frequency of Arrivals
G Attribute Assignments
down the first column beginning on row 2. A numeric expression for

each entity-location combination is entered in the cell where the names
intersect. These files must be in either an .XLS (Excel) or .WK1 (old
Excel) format.
d. Arrival: An arrivals file is a spreadsheet file (.XLS or .WK1 only)
containing arrival information normally specified in the Arrival Editor.
One or more arrival files may be defined and referenced in the External
Files Editor. Arrival files are automatically read in following reading of
the Arrival Editor data. The column entries must be as in Table L11.1.
e. Shift: A shift file record is automatically created in the External Files
Editor when a shift is assigned to a location or resource. If shifts
have been assigned, the name(s) of the shift file(s) will be created
automatically in the External Files Editor. Do not attempt to create a shift
file record in the External Files Editor yourself.
f. .DLL: A .DLL file is needed when using external subroutines through the
XSUB() function.
External files can also be used to populate or extract data from a model array (see
Arrays further below).
Problem Statement
At the Pomona Castings, Inc. castings arrive for processing at the rate of 12 per
hour (average interarrival time assumed to be five minutes). Seventy percent of
the castings are processed as casting type A, while the rest are processed as cast-
ing type B.
For Pomona Castings, Inc., create an entity-location file named P11_4.xls
(Figure L11.12) to store the process routing and process time information. In the
FIGURE L11.12
An external entity-
location file in .xls
format.
har01307_LB11_595-616.indd 602 20/12/10 1:57 PM

simulation model, read from this external file to obtain all the process informa-
tion. Build a simulation model and run it for 100 hours. Keep track of the work-
in-process inventory and the production quantity.
Choose Build More Elements External Files. Define the ID as SvcTms.
The Type of file is Entity Location. The file name (and the correct path) is also
provided (Figure L11.13). In the Process definition, use the file ID (Figure L11.14)
instead of the actual process time—for example, Wait SvcTms(). Change the
file path to point to the appropriate directory and drive where the external file is
located. A snapshot of the simulation model is shown in Figure L11.15.
FIGURE L11.13
File ID and file name created for the external file.
FIGURE L11.14
Processing and routing tables referring to the file ID of the external file.
FIGURE L11.15
A snapshot of the
Pomona Castings, Inc.
har01307_LB11_595-616.indd 603 20/12/10 1:57 PM

604 Part II Labs
L11.4 Arrays
An array is a collection of values that are related in some way such as a list of test
scores, a collection of measurements from some experiment, or a sales tax table.
An array is a structured way of representing such data.
An array can have one or more dimensions. A two-dimensional array is
useful when the data can be arranged in rows and columns. Similarly, a three-
dimensional array is appropriate when the data can be arranged in rows, columns,
and ranks. When several characteristics are associated with the data, still higher
dimensions may be appropriate, with each dimension corresponding to one of
these characteristics.
Each cell in an array works much like a variable. A reference to a cell in an array
can be used anywhere a variable can be used. Cells in arrays are usually initialized to
zero, although Initializing cells to some other value can be done in the initialization
logic. A WHILE...DO loop can be used for initializing array cell values.
Suppose that electroplating bath temperatures are recorded four times a day
at each of three locations in the tank. These temperature readings can be arranged
in an array having four rows and three columns (Table L11.2). These 12 data items
can be conveniently stored in a two-dimensional array named Temp[4,3] with
four rows and three columns.
An external Excel file (BathTemperature.xls) contains these bath temperature
data (Figure L11.16). The information from this file can be imported into an array
in ProModel (Figure L11.17) using the Array Editor. When you import data from
TABLE L11.2 Electroplating Bath Temperatures
Time Location 1 Location 2 Location 3
1 75.5 78.7 72.0

2 78.8 78.9 74.5
3 80.4 79.4 76.3
4 78.5 79.1 75.8
FIGURE L11.16
An external file
containing bath
temperatures.
FIGURE L11.17
The Array Editor in ProModel.
har01307_LB11_595-616.indd 604 20/12/10 1:57 PM

FIGURE L11.18
The Import File dialog
in the Array Editor.
TABLE L11.3 Data for Pending Orders TABLE L11.4 Process Routings and
at Joe’s Jobshop Average Process Times
Jobs No. of Pending Orders Jobs Machines Process Times
A 25 A 2-3-1 45-60-75
B 27 B 1-2-3 70-70-50
C 17 C 3-1-2 50-60-60
an external Excel spreadsheet into an array, ProModel loads the data from left
to right, top to bottom. Although, there is no limit to the quantity of values you
may use, ProModel supports data import for only two-dimensional arrays. Fig-
ure L11.18 shows the Import File dialog in the Array Editor.
Value of an array can also be exported to an Excel spreadsheet (*.xls) during
the termination logic. Microsoft Excel must have been installed in your computer
in order to export data from an array at run time. Array exporting can be turned
off in the Simulations Options dialog.
Problem Statement
Table L11.3 shows the status of orders at the beginning of the month at Joe’s
Jobshop. In his shop Joe has three machines through which three types of jobs
are routed. All jobs go to all machines, but with different routings. The data for
job routings and processing times (exponential) are given in Table L11.4. The
processing times are given in minutes. Use a one-dimensional array to hold the
information on the order status. Simulate and find out how long will it take for Joe
to finish all his pending orders.
har01307_LB11_595-616.indd 605 20/12/10 1:57 PM

606 Part II Labs
The layout, locations and the arrival of jobs at Joe’s Jobshop are shown in
Figures L11.19, L11.20, and L11.21. The pending order array is shown in Fig-
ure L11.22. The initialization logic is shown in Figure L11.23. The processes and
routings are shown in Figure L11.24. It took about 73 hours to complete all the
pending work orders.
FIGURE L11.19
Layout of Joe’s
Jobshop.
FIGURE L11.20
Locations in Joe’s
Jobshop.
FIGURE L11.21
Arrival of jobs at Joe’s Jobshop.
FIGURE L11.22
Pending order array in Joe’s Jobshop.
har01307_LB11_595-616.indd 606 20/12/10 1:57 PM

FIGURE L11.23
Initialization logic
in the general
information menu.
FIGURE L11.24
Processing and routing tables at Joe’s Jobshop.
L11.5 Subroutines
A subroutine is a defined block of code that can be called from anywhere in the
model. It can be called upon to perform a block of logic and optionally return a
value. Subroutines may have parameters or local variables (local to the subrou-
tine) that take on the values of arguments passed to the subroutine. There are three
variations of the use of subroutines in ProModel.
har01307_LB11_595-616.indd 607 20/12/10 1:57 PM

608 Part II Labs
1. A subroutine is called by its name from some other logic in the model.
When the subroutine finishes executing, the calling logic resumes
execution.
2. A subroutine is processed independently of the calling logic so that
the calling logic continues without waiting for the subroutine to
finish. An ACTIVATE statement followed by the name of the subroutine
accomplishes this.
3. Subroutines written in an external programming language can be called
using the XSUB() function.
Subroutines are defined in the Subroutines Editor in the More Elements section of
the Build menu. For more information on ProModel subroutines, please refer to
the ProModel Help System.
Problem Statement
At California Gears, gear blanks are routed from a fabrication center to one of
three manual inspection stations. The operation logic for the gears is identical
at each station except for the processing times, which are a function of the indi-
vidual inspectors. Each gear is assigned two attributes during fabrication. The
first attribute, OuterDia, is the dimension of the outer diameter of the gear. The
second attribute, InnerDia, is the dimension of the inner diameter of the gear.
During the fabrication process the outer and inner diameters are machined with
an average of 4.015 0.01 and 2.015 0.01 inches (uniformly distributed).
These dimensions are tested at the inspection stations and the values entered
into a text file (Quality.doc) for quality tracking. After inspection, gears are
routed to a shipping location if they pass inspection, or to a scrap location if
they fail inspection.
Gear blanks arrive at the rate of 12 per hour (interarrival time exponentially
distributed with a mean of five minutes). The fabrication and inspection times
(normal) are shown in Table L11.5. The specification limits for the outer and
inner diameters are given in Table L11.6. The layout, locations, entities, and
arrival of raw material are shown in Figures L11.25, L11.26, L11.27, and L11.28
TABLE L11.5 Fabrication and Inspection

Times at California Gears
TABLE L11.6 The Specification
Mean Std. Dev.
Limits for Gears
Fabrication Time 3.2 .1 USL LSL

Inspector 1 4 .3
Inspector 2 5 .2 Outer Diameter 4.01" 4.02"
Inspector 3 4 .1 Inner Diameter 2.01" 2.02"
har01307_LB11_595-616.indd 608 20/12/10 1:57 PM

FIGURE L11.25
Layout of California
Gears.
FIGURE L11.26
Locations at the
California Gears.
FIGURE L11.27
Entities at the
California Gears.
FIGURE L11.28
Arrival of raw material at the California Gears.
har01307_LB11_595-616.indd 609 20/12/10 1:57 PM

610 Part II Labs
respectively. The subroutine defining routing logic is shown in Figure L11.29.

Figure L11.30 shows the processes and routing logic. The external file in which
quality data will be written is defined in Figure L11.31. Figure L11.32 shows a
portion of the actual gear rejection report.
FIGURE L11.29
Subroutines defining routing logic.
FIGURE L11.30
Processing and routing tables for California Gears.
FIGURE L11.31
External file for
California Gears.
har01307_LB11_595-616.indd 610 20/12/10 1:57 PM

FIGURE L11.32
Gear rejection report
(partial) at California
Gears.
L11.6 Random Number Streams

The sequence of random numbers generated by a given seed is referred to as a random
number stream. A random number stream is a sequence of an independently cycling
series of random numbers. The number of distinct values generated before the stream
repeats is called the stream’s cycle length. Ideally, the cycle length should be long
enough so that values do not repeat within the simulation. For a more thorough dis-
cussion on generating random numbers, please refer to Chapter 3, section 3.4.1.
In ProModel, up to 100 streams may be used in the model. A random number
stream always generates numbers between 0 and 1, which are then used to sample
from various hypothetical distributions by a method called inverse transforma-
tion, which is discussed in Chapter 3, section 3.4.2.
By default, all streams use a seed value of 1. Also, the random number
streams are not reset between each replication if multiple replications are run. The
Streams editor is used to assign a different seed to any stream and also to reset the
seed value to the initial seed value between replications.
Problem Statement
Salt Lake Machine Shop has two machines: Mach A and Mach B. A mainte-
nance mechanic, mechanic Dan, is hired for preventive maintenance on these
machines. The mean time between preventive maintenance is 120 10 minutes.
Both machines are shut down at exactly the same time. The actual maintenance
takes 10 5 minutes. Jobs arrive at the rate of 7.5 per hour (exponential mean
time between arrival). Jobs go to either Mach A or Mach B selected on a random
basis. Simulate for 2,000 hours.
The layout, resources, path network of the mechanic, processing and rout-
ing tables are shown in Figures L11.33, L11.34, L11.35, and L11.36. The ar-
rival of customers is defined by an exponential interarrival time with a mean of
three minutes. The downtimes and the definition of random number streams are
shown in Figures L11.37 and L11.38, respectively. Note that the same seed value
(Figure L11.38) is used in both the random number streams to ensure that both
machines are shut down at exactly the same time.
L11.7 Merging a Submodel

Sometimes a large model is built in smaller segments. A model segment can be a man-
ufacturing cell or a department. Different analysts can build each segment. After all
segments are complete, they can be merged together to form a single unified model.
har01307_LB11_595-616.indd 611 20/12/10 1:57 PM

612 Part II Labs
FIGURE L11.33
Layout of the Salt Lake
Machine Shop.
FIGURE L11.34
Resources at Salt Lake Machine Shop.
FIGURE L11.35
Path network of the mechanic at Salt Lake Machine Shop.
FIGURE L11.36
Processing and routing tables at Salt Lake Machine Shop.
har01307_LB11_595-616.indd 612 20/12/10 1:57 PM

FIGURE L11.37
Clock downtimes for machines A and B at Salt Lake Machine Shop.
FIGURE L11.38
Definition of random
number streams.
The Merge Model option in the File menu allows two or more independent
(complete or incomplete) models to be merged together into a single model. Entity
and attribute names common to both models are considered common elements in
the merged model. Duplicate locations, resources, or path networks must first be
renamed or deleted from the original merging model. If the graphic libraries are
different, the user has an option to append the merging model’s graphic library to
the base model’s graphic library. All other duplicate elements cause a prompt to
appear with a chance to delete the duplicate element.
Problem Statement
SoCal Casting Inc. (Lab 4, section L4.5) decides to merge with El Segundo
Composites (Lab 4, section L4.7.1). The new company is named El Segundo
Castings N’ Composites. Merge the model for section L4.5 with the model for
section L4.7.1. All the finished products—castings as well as composites—are
now sent to the shipping queue and shipping clerk. The model in section L4.5
is shown in Figure L4.20. The complete simulation model, after the model for
section L4.7.1 is merged, is shown in Figure L11.39. After merging, make the
necessary modifications in the process and routing tables.
We will make suitable modifications in the Processing module to reflect these
changes (Figure L11.40). Also, the two original variables in section L4.5 (WIP
and Prod_Qty) are deleted and four variables are added—WIPCasting, WIPCom-
posite, Prod_Qty_Casting, and Prod_Qty_Composite.
har01307_LB11_595-616.indd 613 20/12/10 1:57 PM

614 Part II Labs
FIGURE L11.39
Merging the models from section L4.5 and section L4.7.1.
FIGURE L11.40
Changes made in the process and routing tables after merging.
L11.8 Exercises
1. Differentiate between the following:
a. Table functions versus arrays.
b. Subroutines versus macros.
c. Arrivals versus arrival cycles.
d. Scenarios versus replications.
e. Built-in distribution versus user distribution.
har01307_LB11_595-616.indd 614 20/12/10 1:57 PM

TABLE L11.7 Processing Time Distributions for Various Job Types
Job Type Ratio Minimum Mean Maximum
Job A 0.4 30 sec 60 sec 90 sec

Job B 0.5 20 sec 30 sec 40 sec
Job C 0.1 90 sec 120 sec 200 sec
2. What are some of the advantages of using an external file in ProModel?

3. HiTek Molding, a small mold shop, produces three types of parts:
Jobs A, B, and C. The ratio of each part and the processing times
(minimum, mean, and maximum of a triangular distribution) are shown
in Table L11.7.
Create an array (mold-time) and an external file (L11_8_3.xls) for the
molding processing time data.
Orders for jobs arrive at an average rate of 4 every minute (interarrival
time exponentially distributed). Each machine can mold any type of
job, one job at a time. Develop scenarios with one, two, three, or four
molding machines to compare with each other. Simulate each scenario
for 100 hours and compare them with each other. What would be
appropriate criteria for such a comparison?
4. For the Southern California Airline (Exercise 4, section L4.13), use an
external file to read interarrival and service time distribution data directly
into the simulation model.
5. For the Bengal Toy Company (Exercise 9, section L4.13), use
an external file to store all the data. Read this file directly into the
simulation model.
6. In Exercise 1, of section L6.13, use an array to store the process time
information for the five pieces of equipment. Read this information
directly into the simulation model.
7. For the Detroit ToolNDie plant (Exercise 14, section L4.13), generate
the following three scenarios:
a. Scenario I: One tool crib clerk.
b. Scenario II: Two tool crib clerks.
c. Scenario III: Three tool crib clerks.
Run 10 replications of each scenario. Analyze and compare the results.
How many clerks would you recommend hiring?
8. In United Electronics (Exercise 7, section L4.13), use an array to store
the process time information. Read this information from an external
Excel spreadsheet into the simulation model.
har01307_LB11_595-616.indd 615 20/12/10 1:57 PM

616 Part II Labs
TABLE L11.8 Macros and RTIs for Various Processing

Time Parameters
Average Grind time Half-width of Grind time

Average Shipping Clerk time (casting) Half-width of Shipping Clerk time (casting)
Average Shipping Clerk time (composite) Half-width of Shipping Clerk time
(composite)
Average Mill time Std. Dev. Of Mill Time
Average Ply Cutting time Half-width of Ply Cutting time
Average Layup time Half-width of Layup time
9. For Salt Lake City Electronics (Exercise 10, section L4.13), use
external files (arrivals and entity_location files) to store all the data. Read
this file directly into the simulation model.
10. For El Segundo Castings N’ Composites example in section L11.7,
create macros and suitable run-time interfaces for the processing time
parameters shown in Table L11.8.
har01307_LB11_595-616.indd 616 20/12/10 1:57 PM

L A B
12 MATERIAL HANDLING
CONCEPTS
With malice toward none, with charity for all, with firmness in the right, as God
gives us to see the right, let us strive on to finish the work we are in . . .
—Abraham Lincoln
In this lab we introduce you to various material handling concepts in ProModel.

First in section L12.1 we introduce you to the concepts of conveyors. In the next
section (section L12.2) we introduce manual and other discrete material handling
systems, such as the forklift truck. Section L12.3 introduces you to the concepts
of a crane system. For a more detailed discussion of modeling various material
handling systems, please refer to Chapter 12.
L12.1 Conveyors
Conveyors are continuous material handling devices for transferring or moving
objects along a fixed path having fixed, predefined loading and unloading points.
Some examples of conveyors are belt, chain, overhead, power-and-free, roller,
skate, and bucket conveyors.
In ProModel, conveyors are locations represented by a conveyor graphic. A
conveyor is defined graphically by a conveyor path on the layout. Once the path
has been laid out, the length, speed, and visual representation of the conveyor can
be edited by double-clicking on the path and opening the Conveyor/Queue dialog
box (Figure L2.17). The various conveyor options are specified in the Conveyor
Options dialog box, which is accessed by clicking on the Conveyor Options but-
ton in the Conveyor/Queue dialog box (Figure L12.1).
617
har01307_LB12_617-638.indd 617 20/12/10 1:57 PM

618 Part II Labs
FIGURE L12.1
The Conveyor Options
dialog box.
FIGURE L12.2
A snapshot of the
San Dimas Car Wash
system model.
FIGURE L12.3
Processing and routing tables at San Dimas Car Wash.
L12.1.1 Single Conveyor

At the San Dimas Car Wash, dirty cars comes in randomly with an exponential
interarrival time distribution [e(15) minutes]. After the initial greetings, custom-
ers select the type of car wash desired, pay for it, and then hand over the keys to
the car wash attendant. All this takes a time that is randomly distributed [N(8,2)
minutes]. After that the car is placed on a conveyor that goes through the auto-
mated wash system. The conveyor is 400 feet long and moves at a speed of 20
feet per minute. After wash the car is handed over to the customer at the car wash
exit station. During the handing over of the car the customer tips the attendant. All
this takes a random amount of time that is normally distributed [N(5,1) minutes].
Simulate this car wash system and run it for 100 hours of operation. Figure L12.2
shows a snapshot of the car wash system in operation. Figure L12.3 shows the
processing and routing tables for this simulation model. Note the use of the # and
the // symbols for inserting comments in the simulation code.
L12.1.2 Multiple Conveyors

Problem Statement
At Ship’n Boxes Inc., boxes of Type A arrive at ConvA and Type B arrive at ConvB
at the rate of one every five minutes and one every two minutes (exponentially
har01307_LB12_617-638.indd 618 20/12/10 1:57 PM

TABLE L12.1 Conveyor Lengths and Speeds
Name Length (feet) Speed (feet per min.)
ConvA 100 30
ConvB 100 50
ShippingConv 200 20
FIGURE L12.4
Processing and routing tables for the Ship’n Boxes Inc. model.
FIGURE L12.5
A snapshot of the
the Ship’n Boxes Inc.
distributed). Both types of boxes are sent to the shipping dock by the shipping con-
veyor. The speed and length of the conveyors are shown in Table L12.1. Develop
a simulation model and run it for 10 hours. Figure out how many boxes of each
type are shipped in 10 hours.
Four locations (ConvA, ConvB, ShippingConv, and Shipping_Dock) and
two entities (BoxA and BoxB) are defined. The processing and routing tables are
shown in Figure L12.4. Figure L12.5 shows a snapshot of the simulation model
for the Ship’n Boxes Inc.
har01307_LB12_617-638.indd 619 20/12/10 1:57 PM

620 Part II Labs
L12.1.3 Recirculating Conveyor

At PCBA Electronics, a full-service printed circuit board assembly facility in Los
Angeles, a recirculating conveyor transports material between various locations
in the facility. After bare printed circuit boards are received they are processed on
one of two identical high-speed placement machines quickly and with high accu-
racy. The processing time (on either machine) is exponentially distributed [e(10)
minutes]. Bare boards arrive for assembly to the PCB Receive area with a ran-
dom exponential interarrival time distribution [e(12) minutes]. After bare boards
arrive the customer ID and serial numbers are printed on the board which takes
N(5,2) minutes. After component placement at one of the two machines the com-
pleted boards are sent to the PCB Ship area where they are properly packed, ship-
ping documents printed and shipped. All of this work at the PCB Ship area takes
N(10,3) minutes. The PCB Ship area as well as machines 1 and 2 have a capacity
to hold one board each. If a board arrives when a machine (or the shipping area) is
occupied it is recirculated using the recirculating conveyor system. Figure L12.6
shows the processing and routing tables. Figure L12.7 shows a snapshot of the
PCBA assembly facility. Note the use of four conveyor segments to form a recir-
culating conveyor system. Each segment of the conveyor is 100 feet long and the
speed of the conveyor segments are 20 feet per minute each. Simulate for 2,000
hours of operation. Track the WIP and the production quantity.
FIGURE L12.6
Processing and routing tables for PCBA Electronics with recirculating conveyor.
har01307_LB12_617-638.indd 620 20/12/10 1:57 PM

FIGURE L12.7
A snapshot of the
PCBA Electronics
simulation model.
L12.2 Resources, Path Networks, and Interfaces

A resource is a person, piece of equipment (viz. lift truck, pallet truck, etc.)
or some other discrete device that is used for one or more of the following
functions: transporting entities, assisting in performing a process on entities at
locations, performing maintenance on locations, or performing maintenance on
other resources. Resources consist of one or more units with common charac-
teristics, such as a pool of service technicians or a fleet of lift trucks. Resources
may be dynamic, meaning that they move along a path network, or static, in
which no movement occurs. In Lab 4, section L4.3 we had introduced you to
the concept of static resources. Resource downtimes were introduced in Lab 6,
section L6.4.
Dynamic resources are resources that move along an assigned path network
and may transport entities between locations as a forklift would. They may also
need to process entities at several locations, such as an operator performing tasks
at more than one location. For these reasons, it is usually preferable to model the
resource’s movement using a path network. Defined properly, the resource will
travel along the path network during the simulation run.
A path network defines the way a resource (dynamic resource) travels be-
tween locations. Entities moving by themselves between locations may also move
on path networks if referenced in the move logic of the routing. Otherwise, they
follow the routing path. Multiple entities and resources may share a common
path network. There are three types of path networks: passing, nonpassing, and
crane. A passing network is used for open path movement where entities and
resources are free to overtake one another. Nonpassing networks consist of single-
file tracks or guide paths such as those used for AGVs where vehicles are not able
to pass. Crane path networks are described in more detail in the section L12.3.
The specifications of path networks allow you to define the nodes at which the
resource parks, the motion of the resource, and the path on which the resource
travels. Path networks consist of nodes that are connected by path segments. A
beginning node and an ending node define a path segment. Path segments may be
har01307_LB12_617-638.indd 621 20/12/10 1:57 PM

622 Part II Labs
unidirectional or bidirectional. Multiple path segments, which may be straight or

jointed, are connected at path nodes. For all path networks, path nodes define the
points where the resources using the network interface with processing locations.
To create path networks:
1. Select the Path button and then left-click in the layout where you want
the path to start.
2. Subsequently, left-click to put joints in the path and right-click to end
the path.
Interfaces are where the resource interacts with the location when it is on the
path network. To create an interface between a node and a location:
1. Left-click and release on the node (a dashed line appears).
2. The left-click and release on the location.
Multiple interfaces from a single node to locations can be created, but only
one interface may be created from the same path network to a particular loca-
tion. More than one resource can use the same path network. For more details
on resources, path networks, and interfaces please refer to the ProModelhelp
system.
L12.2.1 Resource as Material Handler

Problem Statement
At Ghosh’s Gear Shop, a small automotive OEM facility, there are two NC
lathes, a degreaser, and inspection equipment. Blanks are received at the receiv-
ing area. The blanks are machined either on NC Lathe 1 or NC Lathe 2 (machines
are assigned by turn for successive blanks). After machining, the cogs go to the
degreaser for washing. Then the cogs are inspected. An operator moves the parts
between machines. The operator can travel at a speed of 100 feet per minute with
or without load. The operator also loads and unloads the parts, which takes 2 min-
utes each. All machines run on automated cycles.
Blanks arrive at the rate of 3 per hour (exponentially distributed). The ma-
chining, washing, and inspection processes take Uniform(15⫾5), Uniform(5⫾1),
and Normal(8,4) minutes, respectively. The distances between each area are given
in Table L12.2.
TABLE L12.2 Distances between Each Area at Ghosh’s Gear Shop
Receive NC Lathe 1 500 feet

NC Lathe 1 NC Lathe 2 200 feet
NC Lathe 2 Degrease 200 feet
Degrease Inspect 200 feet
Inspect Receive 500 feet
har01307_LB12_617-638.indd 622 20/12/10 1:57 PM

Build a simulation model, run it for 2,000 hours, and determine how many
operators the shop should hire so that the work-in-process inventory remains at
stable and low levels.
Five locations (Receive, NC_Lathe_1, NC_Lathe_2, Inspect, and Degrease),
two entities (Blank and Cog), and a resource (Operator) are defined. The processes
and routings are shown in Figure L12.8. The blanks arrive with an interarrival
time that is exponentially distributed with a mean of 20 minutes. The simula-
tion run hours is set at 2000 in the Simulation Option menu. The model is titled
Ghosh’s Gear Shop in the General Information menu. The operator is defined as
a resource (Figure L12.9). Figure L12.10 shows the path network (Net1) specifi-
cations. All the paths in the path network (Net1) are defined with five nodes and
FIGURE L12.8
Processing and Routing Tables at Ghosh’s Gear Shop—Operator as Material Handler.
FIGURE L12.9
Operator as a resource.
FIGURE L12.10
Path Network for
Operator.
har01307_LB12_617-638.indd 623 20/12/10 1:57 PM

624 Part II Labs
five interfaces (Figure L12.11). The layout of Ghosh’s Gear Shop is shown in
Figure L12.12.
We ran this model with one, two, or three operators working together. The
results are summarized in Table L12.3. The production quantities, average and
maximum WIP, and the average time in system (minutes) are obtained from the
output of the simulation analysis. Note that with one operator both the WIP levels
and the average time in system grow to unsustainable levels. With two or more
operators WIP levels and the average time in system remains at stable and reason-
ably low values. While there are slight improvements in the production rate, WIP
levels, and time in system by going from two to three operators, the improvements
are not large enough to justify hiring the third operator.
FIGURE L12.11
Paths defined in the Path Network.
FIGURE L12.12
Layout of Ghosh’s
Gear Shop.
har01307_LB12_617-638.indd 624 20/12/10 1:57 PM

TABLE L12.3 Production Performance at Ghosh’s Gear Shop—Operator

as Material Handler
Production Average
No. of Quantity in Time in
Operators 2,000 hrs Prod. Rate/hr. Average WIP Maximum WIP System
1 3,156 1.58 1481 2919 29,252 min.

2 5,635 2.82 169.7 369 3,382 min.
3 6,062 3.03 10.5 36 208 min.
FIGURE L12.13
Processing and routing tables at Ghosh’s Gear Shop—operator as material handler plus machine operator.
L12.2.2 Resource as Machine Operator

and Material Handler
For the Ghosh’s Gear Shop in section L12.2.1, we will now consider the impact
of using the operator not only as a material handler moving products between
various machines but also operating all the machines. All the other assumptions
remain the same. Build a simulation model, run it for 2,000 hours, and determine
how many operators the shop should hire so that the work-in-process inventory
remains at stable and low levels. Figure L12.13 shows the processing and routing
tables for Ghosh’s Gear Shop with the operator responsible for moving material
between machines plus operating all the machines.
We ran this model with one, two, three, or four operators working together.
The results are summarized in Table L12.4. Note that with one, two, or even three
operators working together both the WIP levels and the average time in system
grow to unsustainable levels. With four operators WIP levels and the average time
in system stays at stable and reasonably low values. We recommend hiring four
operators.
har01307_LB12_617-638.indd 625 20/12/10 1:57 PM

626 Part II Labs
TABLE L12.4 Production Performance at Ghosh’s Gear Shop—Operator as Material Handler

Plus Machine Operator
Production
No. of Quantity in Average Time
Operators 2,000 hrs Prod. Rate/hr. Average WIP Maximum WIP in System
1 1,665 0.83 2186 4416 43,386 min.

2 3,445 1.72 1294.5 2556 25,956 min.
3 5,105 2.55 413 904 8,209 min.
4 5,976 2.99 50.6 128 997.6 min.
L12.3 Crane Systems

ProModel provides the constructs for modeling crane systems. The path networks
module and the resources module provide modeling features for the crane systems.
Modeling one or multiple cranes operating in the same bay is easily possible.
Users can define the following regarding crane systems:
Crane envelope: The crane envelope is a parallelogram-shaped area
represented by a crane-type path network. Two rails and two lines
connecting the endpoints of the rails bound this envelope. The lines
connecting the endpoints of the rails are the two extreme positions of the
centerlines of the crane bridge.
Bridge Separation Distance: This is the minimum distance between the
centerlines of any two neighboring cranes. Make sure to define the crane
envelope wide enough to allow sufficient space beyond any serviceable
nodes.
Priorities: Two types of priorities are associated with the crane construct in
ProModel: resource requestpriority and crane movepriority.
1. Resource request priorities are used to resolve any conflicts between

multiple entities requesting the same crane at the same time. Lower-
priority tasks are preempted to serve higher-priority tasks.
2. Crane move priorities are used when multiple cranes are operating in the
same work envelope to decide which crane has priority over another to
move. Crane move priorities can be subdivided into three categories:
a. Task move priority is used as the base move priority for travel to pick
up and travel to drop a load.
b. Claim priority is same as the task move priority if the crane is moving
to its ultimate (task) destination. Otherwise, if it is moving under the
influence of another crane, claim priority is equal to the claim priority
of the claim inducer bridge.
c. Effective claim priority applies only when there are three or more
cranes operating in the same envelope. If multiple cranes are moving
in the same direction with overlapping claims, the effective claim
har01307_LB12_617-638.indd 626 20/12/10 1:57 PM

priority in the overlapping zone is the maximum of all the claim

priorities of those cranes.
For more details on crane operations, and in particular to understand how
zone claims are handled, please refer to the ProModel user’s manual.
Problem Statement
Raja owns a manufacturing facility (Raja’s Manufacturing Cell) consisting of
two mills, and a lathe. All jobs are processed in the same sequence, consisting of
arrival station, lathe, mill1, and mill2, and exit station. The processing time on
each machine is normally distributed with a mean of 30 minutes and a standard
deviation of 10 minutes. The arrival rate of jobs is exponentially distributed with a
mean of 60 minutes. In this facility, an overhead crane system is used to handle all
the material movements between the stations. Job pickup and release times by the
crane are uniformly distributed [U(3,1) min]. The path network is shown in Fig-
ure L12.14. The coordinates of all the locations in the cell are given in Table L12.5.
The crane system resource, and processes and routings are shown in Figures L12.15
and L12.16. The crane can travel at the rate of 150 feet/minute with or without a
load. Simulate for 2,000 hours.
FIGURE L12.14
Path networks for the crane system at Raja’s manufacturing cell.
TABLE L12.5 Coordinates of All the Locations in Raja’s

Manufacturing Cell (in feet)
Location X(rail) Y(bridge)
Arrive_station 10 100
Lathe 50 80
Mill1 90 80
Mill2 80 20
Exit_station 0 20
har01307_LB12_617-638.indd 627 20/12/10 1:57 PM

628 Part II Labs
FIGURE L12.15
The crane system resource defined.
FIGURE L12.16
Processing and routing tables at Raja’s Manufacturing Cell.
Select Path Network from the Build menu. From the Type menu, select Crane.
The following four nodes are automatically created when we select the crane type
of path network: Origin, Rail1 End, Rail2 End, and Bridge End. Define the five
nodes N1 through N5 to represent the three machines, the arrive_station, and the
exit_station. Click on the Interface button in the Path Network menu and define
all the interfaces for these five nodes.
Select Resource from the Build menu. Name the crane resource. Enter 1 in
the Units column (one crane unit). Click on the Specs button. The Specifications
menu opens up. Select Net1 for the path network. Enter the empty and full speed
of the crane as 150 ft./min. Also, enter Uniform(180,60) seconds as the pickup
and deposit time.
L12.4 Exercises
1. Modify the example in section L6.11 by having a material handler
(Joe) return the Empty_Pallet from the shipping dock to the Inspector.
Eliminate the pallet queue. Also, let’s have Joe load the boxes onto
pallets and the loaded pallets onto the shipping conveyor [Uniform(3,1)
min]. Assume one pallet being recirculated in the system. Simulate for
2,000 hours and determine the following:
a. The number of monitors shipped per hour.
b. The utilization of the material handler (Joe).
2. Pritha takes over the manufacturing cell (section L12.3) from Raja. She
replaces the crane system and hires a material handler to move material
har01307_LB12_617-638.indd 628 20/12/10 1:57 PM

TABLE L12.6 Distances between Stations at Pritha’s

Manufacturing Cell
Arrival station Lathe 40 feet

Lathe Mill 1 80 feet
Mill 1 Mill 2 60 feet
Mill 2 Exit 50 feet
Exit Arrival station 80 feet
between all the stations. The distances between stations at Pritha’s

Manufacturing Cell are shown in Table L12.6. The processing time on
each machine is normally distributed with a mean of 30 minutes and a
standard deviation of 10 minutes. The arrival rate of jobs is exponentially
distributed with a mean of 40 minutes. Raja can travel at the rate of
150 feet per minute when carrying no load. However, he can walk at
the rate of only 80 feet per minute when carrying a load. Simulate for
2,000 hours.
3. Consider the DumpOnMe facility in Exercise 11 of section L4.13 with
the following enhancements. Consider the dump trucks as material
handling resources. Assume that 10 loads of coal arrive to the loaders
every hour (randomly; the interarrival time is exponentially distributed).
Create a simulation model, with animation, of this system. Simulate for
100 days, 8 hours each day. Collect statistics to estimate the loader and
scale utilization (percentage of time busy). About how many trucks are
loaded each day on average.
4. For the Discs-R-Us Manufacturing Inc. facility in section L6.3, a
maintenance mechanic (a resource) will be hired to do the repair work
on the lathe and the mill. Modify the simulation model and run it or
2,000 hours. What is the utilization of the maintenance mechanic?
Hint: Define maintenance_mech as resource. In the Logic field of the
Clock downtimes for Lathe enter the following code:
GET maintenance_mech
DISPLAY “The Lathe is Down for Maintenance”
Wait N(10,2) min
FREE maintenance_mech
Enter similar code for the Mill.

The maintenance mechanic is an independent contractor and works
four hours each day from 10 A.M. until 2 P.M., Monday through Friday,
with no lunch break. The rest of the shop works from 8 A.M. until 4 P.M.
What will be the impact on the shop (average cycle time and number of
discs made per day) if we hire him full-time and have him work from
8 A.M. until 4 P.M. instead of working part-time?
har01307_LB12_617-638.indd 629 20/12/10 1:57 PM

630 Part II Labs
FIGURE L12.17
Forging
Layout of LA
Forge Inc. Loader
Machine 1 LA Forge Inc. Machine 3
Machine 2
TABLE L12.7 Distances between Various Machines
Loader Machine 1 100 feet

Machine 1 Machine 2 100 feet
Machine 2 Machine 3 100 feet
Machine 3 Loader 100 feet
5. At LA Forge Inc. raw forgings arrive at a recirculating conveyor system

(Figure L12.17). They await a loader in front of the conveyor belt.
The conveyor delivers the forgings (one foot long) to three identical
machines (machine 1, machine 2, and machine 3) located around the
conveyor. A forging is off-loaded for processing to a machine only if
the machine is not in use. Otherwise, the forging moves on to the next
machine. After processing at a machine the machined forging leaves
the system. If the forging cannot gain access to any machine, it is
recirculated. The conveyor moves at 30 feet per minute. The distances
are shown in Table L12.7.
Loading and unloading take 3 minutes each. Forgings arrive to the
loader with an exponential interarrival time and a mean of 10 minutes.
The machining times are normally distributed (mean 20 minutes and
standard deviation 10 minutes). Simulate the conveyor system for 2,000
hours of operation. Analyze the following:
a. Loader and machine utilization.
b. Average time a forging spends in the system.
c. Fraction of forgings that cannot gain access to any machine in the first
pass and need to recirculate one or more times.
d. Average and maximum number of forgings in the conveyor system.
e. Production rate of forgings per hour at LA Forge Inc.
har01307_LB12_617-638.indd 630 20/12/10 1:57 PM

6. U.S. Construction Company has one bulldozer, four trucks, and two
loaders. The bulldozer stockpiles material for the loaders. Two piles of
material must be stocked prior to the initiation of any load operation.
The time for the bulldozer to stockpile material is Erlang distributed and
consists of the sum of two exponential variables, each with a mean of 4
(this corresponds to an Erlang variable with a mean of 8 and a variance
of 32). In addition to this material, a loader and an unloaded truck must
be available before the loading operation can begin. Loading time is
exponentially distributed with a mean time of 14 minutes, for loader 1
and 12 minutes for loader 2.
After a truck is loaded, it is hauled and then dumped; it must be
returned before the truck is available for further loading. Hauling time
is normally distributed. The average hauling time is 22 minutes when
loaded and 18 minutes when unloaded. In both cases, the standard
deviation is three minutes. Dumping time is uniformly distributed
between two and eight minutes. Following a loading operation, the
loader must rest for five minutes before it is available to begin loading
again. Simulate this system at the U.S. Construction Company for a
period of 2,000 hours of operation and analyze the results.
7. At Walnut Automotive, machined castings arrive randomly (exponential,
mean of six minutes) from the supplier to be assembled at one of
five identical engine assembly stations. A fork truck delivers the
castings from the shipping dock to the engine assembly department. A
recirculatingconveyor connects the assembly stations.
The forklift truck moves at a velocity of five feet per second.
The distance from the shipping dock to the assembly department is
1,000 feet. The conveyor is 5,000 feet long and moves at a velocity of
five feet per second.
At each assembly station, no more than three castings can be waiting
for assembly. If a casting arrives at an assembly station and there is
no room for the casting (there are already three castings waiting), it is
recirculated. The assembly time is normally distributed with a mean
of five minutes and standard deviation of two minutes. The assembly
stations are equally distributed around the belt. The load/unload station
is located halfway between stations 5 and 1. The forklift truck delivers
the castings to the load/unload station. It also picks up the completed
assemblies from the load/unload station and delivers them back to the
shipping dock.
Create a simulation model, with animation, of this system. Run the
simulation model until 1,000 engines have been assembled and analyze
the following:
a. Average throughput time for the castings in the manufacturing
system.
b. Utilization of the forklift truck and the conveyor.
har01307_LB12_617-638.indd 631 20/12/10 1:57 PM

632 Part II Labs
FIGURE L12.18 Incoming conveyor

The conveyor sorting
system at the Grocery
Warehouse. Lane 1 Lane 2 Lane 3
Group of Sorters
Shipping
Dock
TABLE L12.8 Forklift Operation Data Collected at the Grocery Warehouse
Load time 1 min.

Unload time 3 min.
Travel speed (with and without load) 10 ft./min.
Distance from shipping dock to sorting conveyors 250 ft.
c. Average and maximum number of castings in the manufacturing

system.
d. Average and maximum number of castings on the conveyor.
8. At the Grocery Warehouse, a sorting system consists of one incoming
conveyor and three sorting conveyors, as shown in Figure L12.18.
Cases enter the system from the left at a rate of 100 per minute at
random times. The incoming conveyor is 150 feet long. The sorting
conveyors are 100 feet long. They are numbered 1 to 3 from left to
right, and are 10 feet apart. The incoming conveyor runs at 10 feet
per minute and all the sorting conveyors at 15 feet per minute. All
conveyors are accumulating type. Incoming cases are distributed to
the three lanes in the following proportions: Lane 1—30 percent;
Lane 2—50 percent; Lane 3—20 percent.
At the end of each sorting conveyor, a sorter (from a group of
available sorters) scans each case with a bar code scanner, applies a
label, and then places the case on a pallet. One sorter can handle 10
cases per minute on the average (normally distributed with a standard
deviation of 0.5 minute). When a pallet is full (40 cases), a forklift
arrives from the shipping dock to take it away, unload the cases, and
bring back the empty pallet to an empty pallet queue near the end of the
sorting conveyor. A total of five pallets are in circulation. The data shown
in Table L12.8 are available for the forklift operation.
har01307_LB12_617-638.indd 632 20/12/10 1:57 PM

Simulate for 2,000 hours. Respond to the following questions:

a. How many sorters are required? The objective is to have the
minimum number of sorters but also avoid overflowing the
conveyors.
b. How many forklifts do we need?
c. Report on the sorter utilization, total number of cases shipped, and
number of cases palletized by lane.
9. Repeat Exercise 8 with a dedicated sorter in each sorting lane. Address
all the same issues.
10. Printed circuit boards arrive randomly from the preparation department.
The boards are moved in sets of five by a handtruck to the component
assembly department, where the board components are manually
assembled. Five identical assembly stations are connected by a
recirculatingconveyor.
When boards are placed onto the conveyor, they are directed to the
assembly station with the fewest boards waiting to be processed. After
the components are assembled onto the board, they are set aside and
removed for inspection at the end of the shift. The time between boards
arriving from preparation is exponentially distributed with a mean of five
seconds. The hand truck moves at a velocity of five feet per second and
the conveyor moves at a velocity of two feet per second. The conveyor
is 100 feet long. No more than 20 boards can be placed on the belt at any
one time.
At each assembly station, no more than two boards can be waiting
for assembly. If a board arrives at an assembly station and there is
no room for the board (there are already two boards waiting), itis
recirculated. The board goes around the conveyor and tries to enter the
station again. The assembly time is normally distributed with a mean of
35 seconds and standard deviation of 8 seconds. The assembly stations
are equallyspaced around the belt, and boards are placed onto the belt
four feet before the first station. After all five boards are placed onto
the belt, the handtruck waits until five boards have arrived from the
preparation area before returning for another set of boards.
Simulate until 100 boards have been assembled. Report on the
utilization of the handtruck, conveyor, and the five operators. How many
assemblies are produced at each of the five stations? (Adapted from
Hoover and Perry, 1989.)
11. In this example, we will model the assembly of circuit boards at four
identical assembly stations located along a closed loop conveyor
(100 feet long) that moves at 15 feet per minute. The boards are
assembled from kits, which are sent in totes from a load/unload station
to the assembly stations via the top loop of the conveyor. Each kit can
be assembled at any one of the four assembly stations. Completed
boards are returned in their totes from the assembly stations back to the
har01307_LB12_617-638.indd 633 20/12/10 1:57 PM

634 Part II Labs
load/unloadstation. The load/unload station is located at the left end of

the conveyor and the four assembly stations (identified by the letters A
through D) are equally spaced along the conveyor, 20 feet apart. The
time to assemble each kit is N(7, 2) minutes. If an assembly station
is busy when a tote arrives, it moves to the next station. Loading and
unloading a tote to/from the conveyor takes N(2,0.2) minutes each. Since
each tote is one foot long, at most 100 totes will fit on the conveyor
at one time. Kits arrive at the load/unload station with an average
interarrival time of five minutes (exponential). Simulate for 2,000 hours.
Answer the following questions:
a. What are the average and maximum numbers of kits in the system?
b. What is the number of assemblies produced and shipped per hour?
c. What are the average and maximum times spent by a kit in the
system?
d. What are the utilizations of the four assembly stations?
e. What percentage of kits goes around the conveyor more than once
before being assembled?
f. Should we consider installing more assembly stations? Fewer?
12. At the Assam Plywood Mill, logs arrive to the mill in truckloads of
15 each. The time between arrivals of trucks is exponentially distributed
[e(60) minutes]. The logs are precut to a standard length of 8 feet,
16 feet, or 24 feet before their arrival to the mill, with a distribution
by length of 20 percent, 50 percent, and 30 percent, respectively. The
arriving logs are unloaded into a bin, from which they are loaded one at
a time onto a conveyor. Sensor devices automatically control this loading
operation such that the next log is kicked onto the conveyor as soon as
sufficient space is available.
The conveyor moves the logs at a speed of 40 feet/minute to a
peeling machine, which debarks the logs and peels them into wood
strips. These strips are dried and glued together in subsequent phases of
the operation to produce plywood. The peeling machine accepts only
logsthat are 8 feet long. Therefore, the 16-foot and the 24-foot logs must
be cut into 8-foot sections before arriving to the peeling machine. This
sectioning is done at an automatic cutting station located 32 feet from the
start of the conveyor. Whenever the leading end of a 16-foot or 24-foot
log reaches the 40-foot mark, the conveyor is automatically stopped and
a vertical cut is made at the 32-foot mark. This operation requires 0.5
minute, after which the conveyor can be restarted.
The peeling machine is located beginning at the 48-foot mark along
the conveyor. Whenever the trailing end of a log reaches this point, the
conveyor is stopped and the log is ready to be moved from the conveyor
and loaded into the peeling machine. The loading operation requires
0.8 minute. At the end of this time the conveyor can be restarted, and the
debarking/peeling operation begins. The peeling machine processes one
8-foot log at a time at the rate of 20 cubic feet of log per minute.
har01307_LB12_617-638.indd 634 20/12/10 1:57 PM

Simulate the system for 2,000 hours to determine the throughput of

wood in cubic feet, the utilization of the peeler, and the number of feet of
conveyor occupied.
13. At Orem Casting Company, rough castings arrive for machining
according to anexponential distribution with a mean of 30 minutes.
An inspector examines the castings for cracks. The inspection time is
normally distributed (mean of 10 minutes and standard deviation of
5 minutes). Approximately 15 percent of the castings are rejected and
scrapped. Three automatic CNC milling machines are used to machine
the remaining castings. Two loaders (material handlers) load the mills in
a time that is uniformly distributed from 15 to 25 minutes. The loaders
will always try to load a casting into an empty machine before waiting
for a mill to become available. The milling operation takes a time that
varies according to a triangular distribution with a minimum, mode,
and maximum of 30, 50, and 80 minutes. The loaders also remove
the milled castings in a time that is uniformly distributed from 10 to
20 minutes. Simulate the operation of the company until 1,000 castings
are machined.
a. What is the average and maximum number of castings in the shop?
b. How many castings did the inspector reject?
c. What is the utilization of the inspector, loaders, and the CNC mills?
d. Should we increase the number of loaders? Should we decrease them?
e. Do we need more CNC mills? Less?
14. At the Washington Logging Company, a logging operation involves
cutting a tree, which is normally distributed (mean of 25 minutes and
standard deviation of 10 minutes). The weight of a tree varies uniformly
from 1,000 to 1,500 pounds,based on the density of fir. The logging
company has five trucks with a capacity of 30 tons each. The truck
will not move the trees to the sawmill until at least 15 tons of trees are
loaded. The trip to the sawmill takes two hours, and the return trip also
requires two hours. The truck takes 20 minutes to unload trees at the
sawmill. Develop a simulation model of this facility. Determine whether
the logging operation has sufficient capability to deliver all the trees cut
in onemonth. The logging operation runs eight hours each day, 20 days
a month.
15. The Washington Logging Company (Exercise 14) also logs alder trees
in addition to fir trees. Logging alder trees are also normally distributed
(mean of 35 minutes and standard deviation of 15 minutes). Alder trees
are segregated from the fir trees at the cutting site. The alder trees have
weights ranging from 1,500 to 2,000 pounds, uniformly distributed.
The trucks transport the alder and pine trees in separate trips. Develop
a simulation model of this facility. Determine whether the logging
operation has sufficient capability to deliver all the trees (alder and
fir) cut in onemonth. The logging operation runs eight hours each day,
20 days a month.
har01307_LB12_617-638.indd 635 20/12/10 1:57 PM

636 Part II Labs
FIGURE L12.19 Steel Coils Copper Coils

Crane system at the Input Input
Rancho Cucamonga Station A Station B
Coil Company.
Crane 1 Crane 2
Output Output Output

Station C Station D Station E
TABLE L12.9 Distances between Stations at the Rancho

Cucamonga Coil Company
From To Distance (feet)
Input A Output C 100

Input A Output D 150
Input A Input B 300
Input B Output E 100
16. For the simulation model in Exercise 14, vary the minimum requirement
of trees that are loaded on a truck from 15 to 20 to 25 tons to determine
the effect on throughput of this decision variable.
17. Rancho Cucamonga Coil Company is considering implementing a
crane system as shown in Figure L12.19. Two cranes operate on separate
tracks. Crane 1 serves only input station A and is used to transport steel
cables to either output station C or D. Crane 2 transports copper cables
from input station B to either output station D or E. The distances
between the input and output stations are shown in Table L12.9.
The arrival of steel cables to input station A is four rolls per hour
(exponential), with 50 percent of the rolls sent to output station C and
50 percent to output station D. The arrival rate of copper cables to input
station B is two rolls per hour (exponential), with 40 percent of the rolls
sent to output station D and 60 percent to output station E. The travel
speeds of both the cranes are 20 feet/minute without load and 10 feet/
minute with load. The pickup and drop-off times are two minutes each.
Simulate the crane system and collect statistics on the throughput from
each input station to each output station.
18. For the Rancho Cucamonga Coil Company (Exercise 17), what will be
the impact on throughput if the cranes run on a single track with output
har01307_LB12_617-638.indd 636 20/12/10 1:57 PM

TABLE L12.10 Operations and Set-up Time Distributions Data
Operation Distribution Cumulative Values (minutes)
Mill Set-up Discrete 30,1,70,2,100,3

Plane Set-up Discrete 20,1.5,50,2.5,80,3.5,100,4.5
Drill Set-up Continuous 0,0.5,30,1,60,2,100,3
Inspect Set-up Discrete 33,0.5,66,1,100,1.5
Throw Away Discrete 20,1.5,50,2.5,80,3.5,100,4.5
station D being on the common path of both the cranes? Model the
interference in reaching output station D.
19. For the Rancho Cucamonga Coil Company (Exercise 17), establish
a cost structure to evaluate and compare the single track (with
interference) versus the separate track crane systems. Use ProModel’s
cost tracking feature.
20. For the Rancho Cucamonga Coil Company (Exercise 17), evaluate the
impact of giving priority to jobs that do not require the common track—
that is, to jobs destined to output stations C and E.
21. Four operators are employed in the production facility described in
Lab 6, Exercise 15. An operator will be required to place a part on a
machine (mill, drill, planer, inspection), before the machine performs its
function, remove the part after machining or inspection, and move it to
the next operation. The part runs on the machine without the operator’s
presence.The scrap parts are placed in a box that can accumulate 10 parts
and is then thrown away by an operator. The set-up times (per batch) and
the throw away times are random variables with distributions as shown
in Table L12.10.
The system runs for three 8-hour shifts per day, 5 days per week.
Let the Type I part have higher priority over Type II parts at the drilling
machine. Let the operators go to lunch one at a time for 30 minutes each
in sequence. The lunch break begins 3.0 hours into an 8 hour shift.
There are electrical interruptions that occur independently with
times following a uniform distribution with mean ⫽ 240 minutes, and
half-range ⫽ 30 minutes. When one of these interruptions occurs, one
mill breaks down. A repair takes 20 minutes, exponentially distributed.
One of the operators performs the repair.
Simulate this system for 2,000 hours with an 8-hour warm-up of the
system.
Reference
Hoover, S. V., and Perry, R. F. Simulation: A Problem Solving Approach. Boston: Addison
Wesley, 1989, pp. B93–B95.
har01307_LB12_617-638.indd 637 20/12/10 1:57 PM

har01307_LB12_617-638.indd 638 20/12/10 1:57 PM
L A B
13 MODELING SERVICE
SYSTEMS
Be silent as to services you have rendered, but speak of favors you have received.
—Seneca (Roman dramatist, philosopher, & politician)
In this lab we will present several simulation models of service systems. In

section L13.1 we introduce you to the operations of a carwash and customer
balking. In section L13.2 we present the operations of a grocery store and
introduce you to table functions. In section L13.3 we simulate the operations of
a burger stand and the cyclic patterns of customer arrival. In section L13.4 we
introduce the concept of user distribution for the same burger stand. In section
L13.5 we present the operations of a university cafeteria. In section L13.6 we
present the operations of a call center. In section L13.7 we model the triage of a
busy county hospital. Finally, in section L13.8 we present a model of the opera-
tions of an office (DMV).
L13.1 Balking of Customers

Balking occurs when a customer looks at the size of a queue and decides to leave
instead of entering the queue because it is too long. Balking is modeled by lim-
iting the capacity or the number of entities that are allowed in the queue. Any
customer that arrives when the queue is full must go away. This means loss of
customers, loss of business, and lost opportunities.
Problem Statement
All American Car Wash is a five-stage operation that takes 2 1 minutes for
each stage of wash (Figure L13.1). The queue for the car wash facility can hold up
to three cars. The cars move through the washing stages in order, one car not being
639
har01307_LB13_639-667.indd 639 20/12/10 1:58 PM

640 Part II Labs
FIGURE L13.1
Layout of the All
American Car Wash.
FIGURE L13.2
Customer arrivals at the All American Car Wash.
FIGURE L13.3
Processing and routing tables at the All American Car Wash.
able to move until the car ahead of it moves. Cars arrive every 2.5 2 minutes for
a wash. If a car cannot get into the car wash facility, it drives across the street to
Better Car Wash. Simulate for 2,000 hours. Estimate the following:
a. How many customers does All American Car Wash lose to its
competition per hour (balking rate per hour)?
b. How many cars are served per hour?
c. What is the average time spent by a customer at the car wash facility?
The customer arrivals and the processing and routing tables are defined as shown
in Figures L13.2 and L13.3. The simulation run hours are set at 100. The total
har01307_LB13_639-667.indd 640 20/12/10 1:58 PM

FIGURE L13.4
Cars that balked.
FIGURE L13.5
Customers served.
number of cars that balked and went away to our competition across the street is
104 in 100 hours (Figure L13.4). That is about 1.04 cars per hour.
The total number of customers served is 2,377 in 100 hours (Figure L13.5).
That is about 24 cars per hour. We can also see that, on average, the customers
spent 14.45 minutes in the car wash facility.
L13.2 Table Functions

Table functions are a useful way to specify and setup relationships between inde-
pendent and dependent variables. A table function editor is provided in ProModel
to build tables. This feature is especially useful when building empirical distribu-
tions or functions into a model.
Table functions are defined by the user and return a dependent value based
on the independent value passed as the function argument. Independent values
are entered in an ascending order. The dependent value is calculated by linear in-
terpolation. For a linear function, only two endpoints are needed. For a nonlinear
function more than two reference values are specified.
Problem Statement
Customers arrive at the Save Here Grocery store with a mean time between ar-
rivals (exponential) that is a function of the time of the day as shown below in
Table L13.1. The grocery store consists of two aisles and a cashier. Once inside
the store, customers may choose to shop in one, two, or none of the aisles. The
probability of shopping aisle one is 0.75, and for aisle two it is 0.5. The number
of items selected in each aisle is described by a normal distribution with a mean
of 15 and a standard deviation of 5. The time these shoppers require to select an
item is five minutes. When all desired items have been selected, customers queue
up at the cashier to pay for them. The time to check out is a uniformly distributed
random variable that depends on the number of items purchased. The checkout
har01307_LB13_639-667.indd 641 20/12/10 1:58 PM

642 Part II Labs
TABLE L13.1 Time between Arrivals at Save Here Grocery
Time of Day Time Between Arrival (minutes)
0 50
2 40
4 25
6 20
8 20
10 20
12 25
14 30
16 30
FIGURE L13.6
Locations in Save Here
Grocery.
FIGURE L13.7
Arrivals in Save Here Grocery.
FIGURE L13.8
Processing and routing
tables at Save Here
Grocery.
time per item varies from 0.6 0.5 minutes (uniformly distributed). Simulate for
10 days (16 hours/day).
The locations, arrival of entities, processes and the layout of the grocery store
are shown in Figures L13.6, L13.7, L13.8, and L13.9, respectively.
har01307_LB13_639-667.indd 642 20/12/10 1:58 PM

FIGURE L13.9
The layout of the Save
Here Grocery.
FIGURE L13.10
Arrival cycles edit
menu.
The arrival frequency derived from the table function arrival time(HR) as
follows:
e(arrival_time(CLOCK(HR)-16*TRUNC(CLOCK(HR)/16))) min
L13.3 Arrival Cycles

Sometimes the arrival of customers to a system follows a definite cyclic pattern.
Examples are arrival of flights to an airport, arrival of buses to a bus stand or
depot, traffic on the freeway, and customers coming to a takeout restaurant.
Arrival cycles are defined in the Arrival Cycles edit table (Figure 13.10),
which is part of More Elements in the Build menu.
har01307_LB13_639-667.indd 643 20/12/10 1:58 PM

644 Part II Labs
TABLE L13.2 Customer Arrival Pattern at Newport Beach Burger
From To Percent
6:00 A.M. 6:30 A.M. 5

6:30 A.M. 8:00 A.M. 20
8:00 A.M. 11:00 A.M. 5
11:00 A.M. 1:00 P.M. 35
1:00 P.M. 5:00 P.M. 10
5:00 P.M. 7:00 P.M. 20
7:00 P.M. 9:00 P.M. 5
FIGURE L13.11
Locations at the
Newport Beach
Burger.
FIGURE L13.12
tables at the Newport
Beach Burger.
The number of entities arriving for each interval of the cycle can be defined
cumulatively or non-cumulatively and expressed as a number or as a percentage
of total entities arriving for the complete cycle. Once the arrival cycles are de-
fined, they can be assigned to an arrivals record in the Arrivals edit table.
Problem Statement
At the Newport Beach Burger stand there are three peak periods of customer ar-
rivals: breakfast, lunch, and dinner (Table L13.2). The customer arrivals taper out
before and after these peak periods. The same cycle of arrivals repeat every day. A
total of 100 customers visit the store on an average day (normal distribution with
a standard deviation of five). Upon arrival, the customers take Uniform(5 2)
minutes to order and receive food and Normal(15, 3) minutes to eat, finish busi-
ness discussions, gossip, read a newspaper, and so on. The restaurant currently has
only one employee (who takes the order, prepares the food, serves, and takes the
money). Simulate for 2,000 hours.
har01307_LB13_639-667.indd 644 20/12/10 1:58 PM

The locations, processing and routing tables, arrivals, arrival quantities are
shown in Figures L13.11, L13.12, L13.13, and L13.14, respectively. The arrival
cycles and the probability density function of arrivals at Newport Beach Burger
are shown in Figure L13.15. A snapshot of the simulation model is shown in
Figure L13.16.
FIGURE L13.13
Arrivals at the Newport
Beach Burger.
FIGURE L13.14
Arrival quantity
defined for the arrivals
table.
FIGURE L13.15
Arrival cycles and
probability density
function of arrivals
defined at Newport
Beach Burger.
FIGURE L13.16
A snapshot of the
Newport Beach Burger.
har01307_LB13_639-667.indd 645 20/12/10 1:58 PM

646 Part II Labs
L13.4 User Distribution

Sometimes it becomes necessary to use the data collected by the user directly
instead of using one of the built-in distributions. User distribution tables can be
accessed from Build → More elements (Figure L13.17). These user-defined or
empirical distributions can be either discrete or continuous. The user distributions
can be expressed as either a cumulative (cumulative density function or cdf) or a
noncumulative (probability density function or pdf) fashion.
Problem Statement
The customers at the Newport Beach Burger stand in section L13.3 arrive in
group sizes of one, two, three, or four with the probabilities shown in Table L13.3.
The mean time between arrivals is 15 minutes (exponential). The probability den-
sity function of ordering and eating times are shown in Tables L13.4 and L13.5,
respectively. Simulate for 2,000 hours.
The user distributions, group size distributions, order time distributions, and
eating time distributions are defined as shown in Figures L13.18, L13.19, L13.20,
and L13.21, respectively.
FIGURE L13.17
User distributions
menu.
TABLE L13.3 Probability Density Function

of Customer Group Sizes
Group Size Probability
1 .4
2 .3
3 .1
4 .2
har01307_LB13_639-667.indd 646 20/12/10 1:58 PM

TABLE L13.4 Probability Density Function TABLE L13.5 Probability Density Function
of Ordering Times of Eating Times
Ordering Time Probability Eating Time Probability
P(0 X 3) .0 P(0 X 10) .0

P(3 X 4) .35 P(10 X 12) .3
P(4 X 5) .35 P(12 X 14) .35
P(5 X 8) .3 P(14 X 16) .35
P(8 X) .0 P(16 X) .0
FIGURE L13.18
User distributions for
the Newport Beach
Burger.
FIGURE L13.19
Group size
distributions for
the Newport Beach
Burger.
FIGURE L13.20
Order time
distributions for
the Newport Beach
Burger.
FIGURE L13.21
Eating time
distribution for the
Newport Beach
Burger.
L13.5 Modeling a University Cafeteria

Students at the State University dormitory use the attached cafeteria for meals.
Lunch time is usually very busy and service has been deteriorating with long lines
and too much time waiting to get food. The food court was originally designed
with all the food displayed behind one long counter. The lunch line will form
in front of this counter and during busy periods will sometimes go outside the
cafeteria building.
har01307_LB13_639-667.indd 647 20/12/10 1:58 PM

648 Part II Labs
Management has analyzed this situation and suggested the following.

a. Separate the dining hall into two areas—Food Court and Dining Area.
b. Separate the food choices into five individual islands positioned around
the food court—Mexican, Hamburger, Thai, Indian, and Pizza.
This layout will discourage students to wait in one long line and instead look
around the dining hall to see what food choices are available and which area is
less crowded.
Students will form a lunch line; check in (swipe their meal card); and pickup
a tray, a dish, and other eating utensils. Students will go to the middle of the food
court to an area we call the decision square. There they will decide which food
island to visit next. Each student decides to visit from one up to five islands, most
popular choice being three islands. They can also decide to visit the same food
island more than once.
From analyzing past data on choices students made management found the
information on the popularity of different food choices as shown in Table L13.6.
Once at a specific food island, students check how busy that island is and get
food from that island only if there is space available. Otherwise, they go back to
the decision square and make another food choice. The capacities of each area are
given in Table L13.7.
TABLE L13.6 Popularity of Various Food

Choices at the Cafeteria
Mexican 30%
Hamburger 10%
Thai 25%
Indian 15%
Pizza 20%
TABLE L13.7 Capacities of Various

Locations in the Cafeteria
Location Capacities TABLE L13.8 Service Time Data in the

Cafeteria
Check In 1
Dish Rack 1 Check In Normal (0.25,0.1) min.
Decision Square 6 Dish Rack Normal (0.5,0.2) min.
Mexican 4 Mexican Normal (3,0.5) min.
Hamburger 4 Hamburger Normal (3,0.5) min.
Thai 4 Thai Normal (3,0.3) min.
Indian 4 Indian Normal (3,0.25) min.
Pizza 4 Pizza Normal (3,0.5) min.
Dining Area 50 Eating Lunch Normal (20,5) min.
Dish Dispose 4 Dispose Dish Normal (.5,.1) min.
har01307_LB13_639-667.indd 648 20/12/10 1:58 PM

Once the students have visited the number of islands they have decided
they go to the dining area. Also, students have a tendency to not spend any more
than 15 minutes in the food court (getting food) before going to the dining area.
After students have finished their lunch in the dining area they dispose their
dishes and leave. Data on various service times were collected and summarized
in Table L13.8. Moving from one area to another takes 15 seconds. A total of
200 students eat lunch every day. They arrive to the cafeteria on average every
1/2 minute (exponentially distributed) as shown in Figure L13.22. Various loca-
tions defined for this cafeteria are shown in Figure L13.23.
We define two attributes Food Choices and Time In to keep track of the choice
of food students make and also to keep track of the student’s time of arrival to the
cafeteria (Figure L13.24). Figure L13.25 shows the processing and routing of stu-
dents at the cafeteria. Note the use of an If-Then-Else logic to model that students
do not wait any more than 15 minutes in the food court to get food. Also, note the
modeling of checking for space availability at any island and deciding whether
to wait or move on to another food island. Figure L13.26 shows a snapshot of the
State University Cafeteria simulation model. Run 10 replications of this simula-
tion model.
FIGURE L13.22
Arrival of students to
the cafeteria.
FIGURE L13.23
Locations for State
University Cafeteria.
FIGURE L13.24
Attributes defined for
the State University
Cafeteria model.
har01307_LB13_639-667.indd 649 20/12/10 1:58 PM

650 Part II Labs
FIGURE L13.25
tables for the State
University Cafeteria
simulation model.
FIGURE L13.26
A snapshot of the State
University Cafeteria
simulation model.
har01307_LB13_639-667.indd 650 20/12/10 1:58 PM

From the results of the simulation we can conclude that it takes little over
3 hours to serve these 200 students. The average time a student spends in the cafete-
ria (waiting, getting food, and eating) is about an hour although almost 48 percent
of that time they are blocked. It will be interesting to explore how we can improve
the operations of this cafeteria. What are some of the alternatives you would like to
explore? How will you model them? Which alternative will you recommend? Why?
How will your recommended solution improve the performance of the cafeteria?
L13.6 Modeling a Call Center—Outsource2US

Outsource2US is a call center outsourcing company located offshore. They have
been in business over a decade and have a reputation for providing excellent ser-
vice. Outsource2US provides three categories of services—voice, email support,
and chat support. Within each category they provide the services as shown in
Table L13.9.
Customer contacts are prioritized as follows. Voice calls have higher pri-
ority than live chats. Live chats are given higher priority than emails. Cur-
rently, there are ten associates and two supervisors working in the call center.
The various types of service requests arrive with the frequency shown in
Table L13.10.
TABLE L13.9 Services Provided by Outsource2US
Percentage within
Service Category Service Category
Voice (Telephone) Online Sales 30%

Order Taking 40
Order Entry 20
Customer Service 10
Email Support Order Fulfillment 30
Tech Support 30
Query Resolution 30
Resolving Escalated Cases 10
Chat Support Online Sales 40
Tech Support 30
Customer Service 30
TABLE L13.10 Service Arrival Frequency
Service Category Frequency Interarrival Time
Voice 60 per hour Exponential (1) min.

Email Support 12 per hour Exponential (5) min.
Chat Support 12 per hour Exponential (5) min.
har01307_LB13_639-667.indd 651 20/12/10 1:58 PM

652 Part II Labs
Customers’ first point of contact is one of the associates. The time to serve a
customer depends on the type of service and is shown in Table L13.11.
After 9 minutes of servicing a telephone call it is elevated to a supervisor.
After 7 minutes of servicing a live chat it is elevated to a supervisor. Ten per-
cent of all emails (complicated technical issues) are elevated to a supervisor, on
arrival.
Figure L13.27 shows all the locations defined for the Outsource2US call cen-
ter model. Figure L13.28 shows the processing and routings tables for this call
center.
Figure L13.29 shows a snapshot of the Outsource2US call center simulation
model. Run the simulation model for 2,000 hours.
TABLE L13.11 Service Times for Various Types of Services

Provided by Outsource2US
Service Category Type Service Time
Inbound Calls Online Sales N (5,2) min.

Order Taking N (7,3) min.
Order Entry N (6,2) min.
Customer Service N (10,3) min.
Email Support Order Fulfillment N (3,1) min.
Tech Support N (5,1) min.
Query Resolution N (7,2) min.
Resolving Escalated Cases N (5,2) min.
Live Chat Online Sales N (4,2) min.
Tech Support N (5,3) min.
Customer Service N (6,3) min.
FIGURE L13.27
Locations for the
Outsource2US model.
har01307_LB13_639-667.indd 652 20/12/10 1:58 PM

FIGURE L13.28
Processing and routing tables for Outsource2US call center.
har01307_LB13_639-667.indd 653 20/12/10 1:58 PM

654 Part II Labs
FIGURE L13.29
A snapshot of the
Outsource2US call
center simulation
model.
TABLE 13.12 Time in System for Telephone Calls, Live Chats, and Emails
Average Time in Maximum Time

Type of Service System (min.) in System (min.)
Telephone Calls 7.18 24.49

Live Chats 6.22 59.55
Emails 6.81 90.06
From analyzing the results of the simulation we conclude that the average
associate is busy about 80 percent of the time while the average supervisor is
busy for about 27 percent of the time. The average and maximum times the tele-
phone calls, live chats, and emails spent in the call center are summarized in
Table L13.12. While the average times are very reasonable, the maximum times
are quite large. How could we improve the quality of the service of the call cen-
ter? What alternatives can we explore? Which alternative would you recommend?
How will it improve the performance of the call center?
L13.7 Modeling a Triage—Los Angeles County Hospital

Sick or injured patients arriving at the ED facility of the Los Angeles County
Hospital are triaged for treatment according to the seriousness of the condition or
injury. They use a simple triage and rapid treatment system as follows:
• Minor delayed care—can delay up to three hours.
• Delayed urgent care—delay up to one hour.
• Immediate care—life-threatening
• Deceased—victim is dead or mortally wounded on arrival, no care needed.
har01307_LB13_639-667.indd 654 20/12/10 1:58 PM

Patients arrive with an interarrival time that is exponentially distributed with

a mean of 60 minutes. Arriving patients wait in the triage queue.
The whole evaluation process is generally conducted on average in
three minutes with a standard deviation of one minute, normally distributed.
Once the evaluation is complete, the victims are labeled with one of the four
triage categories—minor delayed (25%), delayed urgent (40%), immediate
care (30%), and deceased (5%).
After triage patients are sent to the observation area in the ED, the time for
initial treatment depends on the category of the patient’s condition or injury and is
distributed as shown in Table L13.13. After the initial treatment patients are kept
for further observation for a length of time as shown in Table L13.13.
After observation, 5 percent immediate care patients do not survive, 90 per-
cent are admitted to the general ward, and 5 percent are released. Admitted im-
mediate care patients stay for an average of 4 1 day in the hospital. Five percent
admitted immediate care patients do not survive.
Seventy percent of delayed urgent care patients are admitted to the general
ward and 30 percent released. Admitted delayed urgent care patients stay for an
average of 3 1 day in the hospital. Three percent admitted delayed urgent care
patients do not survive.
Ninety-nine percent of minor delayed care patients are released and 1 percent
admitted to the general ward. Admitted minor delayed care patients stay an aver-
age of 2 1 day in the hospital.
The hospital has 65 beds in the general ward and 10 beds in the Obser-
vation Area in the ED. The various locations and their capacities are shown
in Figure L13.30. The various entities defined in this model are shown in
TABLE L13.13 Time for Patient Care at the ED
Triage Category Initial Treatment Time Observation Time
Minor delayed care N (30,10) min. N (30,5) min.

Delayed urgent care N (20,5) min. N (60,10) min.
Immediate care N (10,5) min. N (60,10) min.
FIGURE L13.30
Locations at the
ED facility of the
Los Angeles County
Hospital.
har01307_LB13_639-667.indd 655 20/12/10 1:58 PM

656 Part II Labs
FIGURE L13.31
Entities in the ED
facility model.
FIGURE L13.32
Processing and routing tables for the ED facility of the Los Angeles County hospital.
FIGURE L13.33
A snapshot of the
Simulation Model of
the ED facility of the
Los Angeles Country
Hospital.
Figure L13.31. Figure L13.32 shows the processing and routing tables for
this ED facility of the Los Angeles County hospital.
Figure L13.33 shows a snapshot of the simulation model ED facility of the Los
Angeles County Hospital. Run the simulation model for 24 hours, 365 days a year
(8,760 hours). How is this triage facility working? Do we have adequate capacity
har01307_LB13_639-667.indd 656 20/12/10 1:58 PM

in each area of the triage? Is the waiting time for various categories of patients
acceptable? What steps would you take to improve the performance of this facility?
L13.8 Modeling an Office (DMV)

Customers arrive to the department of motor vehicle (DMV) office for one of sev-
eral services they offer. We collected data from the past year and found 20 percent
customers come in to apply for a learning permit, 25 percent come in for driver
license renewal, 15 percent show up for a drive test, 10 percent come in to get an
ID card, and 30 percent need to get their vehicle registered.
Customers arrive with an average interarrival time of one minute that is
exponentially distributed. About 30 percent of these customers arrive with prior
appointments and the other 70 percent show up without making any appoint-
ments. Customers with appointment have higher priority over customers who
drop in without appointment, for any of these services. Customers first show up at
the reception desk (there are two receptionists) where they check if the customer
has a valid appointment. The customers are directed to the Appointment queue
or the Drop_In queue based on whether they have appointment or not. Next the
customers go to one of two associates. Table L13.14 show the sequence of process
TABLE L13.14 Process Steps and Distribution of Service Times at the DMV
Service Time
Type of Service Process Steps Distribution
Learning Permit Receptionist N (2,1) min.

Show IDdocuments N (5,1) min.
Pay cashier for permit N (4,2) min.
Get written test from permit associate N (3,2) min.
Take written test (test booth capacity is five) U (20,10) min.
Permit associate checks test for pass/fail N (5,2) min.
If pass (70%), take picture and sign N (3,2) min.
Driver License Renewal Receptionist N (2,1) min.
Show IDdocuments N (5,1) min.
Pay cashier for renewal N (4,2) min.
Take picture and sign N (3,2) min.
Drive Test Receptionist N (2,1) min.
Show documents N (3,1) min.
Take the drive test U (15,5) min.
If pass, take pictures and sign N (3,2) min.
ID Card Receptionist N (2,1) min.
Show documents N (5,1) min.
Pay cashier for ID card N (4,2) min.
Take pictures and sign N (3,2) min.
Vehicle Registration Receptionist N (2,1) min.
Show IDand vehicle insurance documents N (7,2) min.
Pay cashier for registration N (4,2) min.
Get new plate and sticker U (4,1) min.
har01307_LB13_639-667.indd 657 20/12/10 1:58 PM

658 Part II Labs
FIGURE L13.34
Locations for the DMV
model.
FIGURE L13.35
Attributes defined for
the DMV model.
FIGURE L13.36
Global variables
defined for the DMV
model.
steps they go through for these services and the distribution of service times for
these processes. Figure L13.34 shows the various locations defined for this model.
Figures L13.35 and L13.36 show the attributes and variables (global) defined in
this model.
Seventy percent of the customers who take the written test for the driving per-
mit pass and go on to get their picture taken. The other 30 percent fail the written
test and need to come back at a later date. Eighty percent of the customers who
take the drive test pass, the remaining 20 percent fail the test and need to come
back again at a later date.
har01307_LB13_639-667.indd 658 20/12/10 1:58 PM

To abide by the local fire code the maximum number of customers in the
building is kept at or below 50. When the building reaches its capacity, any ar-
riving customers are turned away and asked to come back later that day or come
back another day. The front door of the building is closed to arriving customers
after six hours of operation and any remaining customers already in the building
are processed until the last customer has been served. Figure L13.37 shows the
FIGURE L13.37
Processing and routing tables for the DMV model.
(continued)
har01307_LB13_639-667.indd 659 20/12/10 1:58 PM

660 Part II Labs
FIGURE L13.37 (Continued)
processing and routing tables for the DMV model. Figure L13.38 shows a snap-
shot of the simulation model of the DMV office.
Simulate for 30 replications (30 working days) and determine:
a. The 95 percent confidence interval for average waiting time in system
for customers with appointment.
har01307_LB13_639-667.indd 660 20/12/10 1:58 PM

FIGURE L13.38
A snapshot of the Simulation Model for the DMV Office.
b. The 95 percent confidence interval for average waiting time in system

for customers without appointment.
c. The 95 percent confidence interval for the length of the workday for the
DMV office.
d. If it helps the customer to make prior appointments?
e. If the level of manning (2 receptionists, 2 associates, one cashier,
one picture-taking clerk, one permit associate, and one sticker/plate
associate) is sufficient?
f. Which is the bottleneck in this DMV office? What can we do to improve
the situation?
g. The 95 percent confidence interval for the average number of customers
served per day.
h. A consultant has suggested combining the two associates and one cashier
positions that allows these three people to be responsible for checking
all the documents and also to receive payments. Will that improve the
functioning of the DMV office? What metrics would you use to measure
the improvements? Use a paired-t-confidence interval approach (0.05
significance level) to compare the original system and the new system.
har01307_LB13_639-667.indd 661 20/12/10 1:58 PM

662 Part II Labs
L13.9 Exercises
1. San Dimas Bank and Loan has an ATM for drive-in customers. They also
have two tellers—current account and savings account. Thirty percent
of all customers use the drive-in ATM facility. Each teller has his/her
own waiting line of customers. The current account teller can serve a
customer from the savings line if there are no current account customers
waiting for service. Among the walk in customers 20 percent are current
account holders. The ATM use time depends on the type of service (cash
withdrawal, cash or check deposit, balance enquiry, buy stamps, etc.) and
is uniformly distributed between 6 to 12 minutes. The service time at the
teller counter is as follows:
Current Account Service Time – N(5,2) min.

Savings Account Service Time – N(6,2) min.
The customer interarrival time is exponentially distributed with a

mean of 2 minutes. The ATM drive-in facility can only accommodate up
to a maximum of three vehicles at a time (including the one being served
at the ATM). If the drive-in facility is full, arriving customers park their
vehicles and walkin.
The bank closes its front door after six hours of operation each day
but continue to serve the customers remaining in the building until the
last customer has left. Simulate for 30 replications (days) and determine
the following:
a. The 95 percent confidence interval for the average number of
customers served per day.
b. The 95 percent confidence interval for the average time in system for—
ATM customer, Current account customer, Savings account customer.
c. The 95 percent confidence interval for the average number of
customers (ATM, Current account, and Savings account) in the
bank.
d. Is it better to drive-in and use the ATM or walk-in and use the tellers?
How would you make that comparison?
e. There was a suggestion to combine the two tellers and let each teller
handle any type of customers—current account or savings account.
Also, combine the two waiting lines into one line. Will that improve
customer service at the bank? Is this a good suggestion? How would
you decide that?
2. Management of Apna Bazaar, a supermarket store in India wishes to
keep the checkout lines below five people each. Right now there are
six regular checkout counters and one express counter. There are 50
shopping carts; a customer leaves if a cart is not available. The time
between arrivals of customers is exponentially distributed with a mean
of one minute. Twenty percent of customers are express. The shopping
har01307_LB13_639-667.indd 662 20/12/10 1:58 PM

TABLE L13.15 Shopping and Checkout Times for

Customers at Apna Bazaar
Type of Customer Shopping Time Checkout Time
Express Normal (6,2) min. Normal (2,1) min.

Regular Normal (12,3) min. Normal (5,2) min.
time and checkout time of express and regular customers are as shown in
Table L13.15.
Regular customers always choose the shortest checkout line.
Once a customer joins a line, they cannot switch to another line. The
supermarket is open 24 hours/day, 365 days a year. Simulate for a whole
year and determine:
a. How many checkout counters should be kept open to keep the
checkout lines below five people?
b. Find the utilization of carts and checkout counters.
c. Average time in the supermarket for both types of customers.
3. Management of Apna Bazaar (Exercise 2) conducted a study and found
that both shopping time and checkout time correlates with the number
of items purchased. The number of items purchased by a customer is
uniformly distributed U(25,20). The shopping time per item is normally
distributed with an average of 3 minutes and a standard deviation of
1 minute. The checkout time per item is normally distributed with an
average of 30 seconds and a standard deviation of 10 seconds. The total
shopping times and total checkout times are given by the following
relationships:
Total Shopping Time No. of Items Purchased * Shopping Time / Item
Total Checkout Time No. of Items Purchased * Checkout Time / Item
The express counter only handles customers with 10 or less items to
checkout. All the other information is given in Exercise 2. Simulate for a
whole year and determine:
a. How many checkout counters should be kept open to keep the
checkout lines below five people?
b. Find the utilization of carts and checkout counters.
c. Average time in the supermarket for both types of customers.
4. Pritha’s Hair Salon employs three barbers—Puja, Prerana, and Pritha
(owner). On the average, the mean time between arrivals is exponentially
distributed with a mean of five minutes. The haircut times for various
types of customers are shown in Table L13.16. Thirty-five percent of
har01307_LB13_639-667.indd 663 20/12/10 1:58 PM

664 Part II Labs
TABLE L13.16 Haircut Times for Various Types of

Customers
Percentage of
Types of Customers all Customers Haircut Time
Women 60% N (20,5) min.

Men 25 N (12,4) min.
Child 15 N (10,4) min.
customers have a favorite hairstylist: 15 percent prefer Barber Pritha

(owner), 15 percent prefer Barber Puja, and 5 percent prefer Barber
Prerana. There is a 3 percent chance that a barber will be absent any
given day. Assume that if a favorite barber is absent, the customer will
not wait. The customer not having a favorite will look at the overall
waiting line, and if the number of customers waiting for a haircut is
more than six, will leave. A customer wanting a favorite barber will not
wait if the number of customers waiting for the same barber is more
than three. Given this situation, how many barbers should Pritha employ
if she wants to keep the average waiting time for all customers below
20 minutes?
5. West Coast Federal, a drive-in bank, has one teller and space for
five waiting cars. If a customer arrives when the line is full, he or she
drives around the block and tries to get in again. Time between arrivals
is exponential with mean of 10 minutes. Time to drive around the
block is normally distributed (mean 3 minutes and standard deviation
0.6 minute). If the second attempt is unsuccessful the customer goes
away (balks). Service time is uniform at 9 3 minutes. Build a
simulation model and run it for 2,000 hours.
a. Collect statistics on time in queue, time in system, teller utilization,
number of customers served per hour, and number of customers
balked per hour.
b. Modify the model to allow a space for two cars to wait after they are
served to get back onto the street. Waiting time for traffic is exponential
with a mean of four minutes. Collect statistics similar to part a.
c. Modify the model to reflect balking customers leaving the system and
not driving around the block. Collect statistics similar to part a. How
many customers are lost per hour?
d. The bank’s operating hours are 9 A.M. until 3 P.M. The drive-in facility
is closed at 2:30 P.M. Customers remaining in line are served until
the last customer has left the bank. Modify the model to reflect these
changes.Collect statistics similar to part a.
har01307_LB13_639-667.indd 664 20/12/10 1:58 PM

FIGURE L13.39
San Dimas Mutual Bank
Layout of the San
Dimas Mutual Bank.
Walk-In Indoor ATM
Parking Lot
ATM 1 ATM 2
Drive-In
TABLE L13.17 Types of Customers and Their Service

Time Distributions
Fraction of Time at ATM

Type of Customer All Customers (indoor or drive-in)
Save Money 1/3 Normal (7,2) minutes

Spend Money 1/2 Normal (5,2) minutes
Count Money 1/6 Normal (3,1) minutes
6. San Dimas Mutual Bank has two drive-in ATM kiosks in tandem but
only one access lane (Figure L13.39). In addition, there is one indoor
ATM for customers who decide to park (30 percent of all customers) and
walk in. Customers arrive at intervals that are spaced on average five
minutes apart (exponentially distributed). ATM customers are of three
types—save money (deposit cash or check), spend money (withdraw
cash), or count money (check balance). The times for these services are
shown in Table L13.17. If both ATMs are free when a customer arrives,
the customer will use the “downstream” ATM 2. A car at ATM 1 cannot
pass a car at ATM 2 even if it has finished.
Create a simulation model of the bank. The ATM kiosks operate
24/7. Run the model until 2,000 cars have been served. Analyze the
following:
a. The average and maximum drive-in queue size.
b. The average and maximum time spent by a customer waiting in
drive-in queue.
har01307_LB13_639-667.indd 665 20/12/10 1:58 PM

666 Part II Labs
c. The average and maximum walk-in queue size.

d. The average and maximum time spent by a customer waiting in
walk-in queue.
e. The average and maximum time in the system.
f. Utilization of the three ATMs.
g. Number of customers served each hour.
7. Modify the San Dimas Mutual Bank model in Exercise 6 so that if
four or more cars are in the ATM drive-in system at the bank when the
customer arrives, the customer decides to park and walk in to use the
indoor ATM. Run the model until 2,000 cars have been served. Analyze
the following:
a. Average and maximum number of customers deciding to park and
walk in. How big should the parking lot be?
b. The average and maximum drive-in queue size.
c. The average and maximum time spent by a customer waiting in
drive-in queue.
d. The average and maximum walk-in queue size.
e. The average and maximum time spent by a customer waiting in
walk-in queue.
f. The average and maximum time in the system.
g. Utilization of the three ATMs.
h. Number of customers served each hour
8. At the West Coast Diagnostic Lab patients come in for blood test and/or
urine test. There are two classes of patients—patients who made prior
appointments and others who did not. The mean interval between
arrivals for patients with appointments is U(15 5) minutes while
that for patients without appointment is U(5 2) minutes. Patients
with appointment have higher priority for service over patients without
appointment. About 20 percent patients are fasting. Fasting patients have
higher priority over nonfasting patients within their own class. There
are two phlebotomists currently working in the lab. Drawing blood
requires N(20,5) minutes. Getting the urine sample takes N(5,2) minutes.
About 70 percent patients require blood draw, only. The remaining 30
percent require both blood and urine tests done. The laboratory is open
from 8 A.M. until 4 P.M. each day with a lunch break from 12 until 1 P.M.
Patients who show up during the lunch break are turned away. Patients
already in the building after the lunch break are continued to be served.
When the lab closes for the day (4 P.M.), patients already in the building
are continued to be served until the last patient has left the building.
Simulate for 200 working days and determine:
a. On average how many patients of each type are served per day?
Develop a 95 percent confidence interval on the average number of
patients (of each type) served per day.
har01307_LB13_639-667.indd 666 20/12/10 1:58 PM

b. What is the average waiting time for each type of patients? Develop
a 95 percent confidence interval on the average waiting time for each
type of patients.
c. What is the average time in the diagnostic lab for each type of
patients? Develop a 95 percent confidence interval on the average
time in the diagnostic lab for each type of patients.
d. What is the average number of patients in the diagnostic lab? Is
this at a reasonable level? Why or why not? Develop a 95 percent
confidence interval on the average number of patients in the
diagnostic lab.
e. What is the average utilization of the phlebotomists? Do we need to
hire additional phlebotomists? Why or why not?
har01307_LB13_639-667.indd 667 20/12/10 1:58 PM

A P P E N D I C E S

Appendix B Critical Values for Student’s t Distribution and Standard Normal
Distribution 674
Appendix C F Distribution for ␣ 0.05 675
Appendix D Critical Values for Chi-Square Distribution 676
har01307_APPA_668-673.indd 668 12/17/10 9:28 AM

APPENDIX A
CONTINUOUS AND DISCRETE DISTRIBUTIONS IN PROMODEL*
In ProModel, distribution functions are built-in functions that return observations (random
variates) from various probability distributions. Common continuous and discrete prob-
ability distributions implemented in ProModel as callable functions are summarized in this
appendix. Note that ProModel includes a User-defined function that allows users to specify
a custom discrete or continuous distribution via entries to a table if none of the standard
ProModel built-in distribution functions can adequately model the situation.
In addition to required parameters, some ProModel distribution functions have optional
parameters denoted by {< >}. The symbols <s> denote the random number generator stream
used to generate observations from the distribution. If this option is omitted, ProModel
will use stream 1. The symbols <ax> denote an optional axis shift. The exponential (E),
Gamma (G), Weibull (W), Lognormal (L), Inverse Gaussian (IG), Pearson 5 (P5), and
Pearson 6 (P6) distributions all normally have a minimum value of zero. Use an axis shift to
alter the distribution’s minimum value and at the same time shift the entire distribution. Any
time an axis shift is specified, a random number generator stream must be specified also.
Any negative value returned by a distribution that is used for a time expression in
ProModel will be automatically converted to zero.
A.1 Continuous Distributions

Distribution: Beta
Syntax: B(a, b, c, d{,<s>}) a shape value 1, b shape value 2, c lower boundary,
d upper boundary
Application: The beta distribution is a continuous distribution that has both upper and
lower finite bounds. Because many real situations can be bounded in this way, the beta
distribution can be used empirically to estimate the actual distribution before many data
are available. Even when data are available, the beta distribution should fit most data in a
reasonable fashion, although it may not be the best fit. The uniform distribution is a special
case of the beta distribution.
Beta distributions have been used to model distributions of activity time in PERT analy-
sis, porosity/void ratio of soil, phase derivatives in communication theory, size of prog-
eny in Escherichia coli, dissipation rate in breakage models, proportions in gas mixtures,
steady-state reflectivity, clutter and power of radar signals, construction duration, particle
size, tool wear, and others. Many of these uses occur because of the doubly bounded nature
of the beta distribution.
Distribution: Erlang
Syntax: ER(a, b{,<s>}) a mean value, b parameter
Application: The Erlang distribution is a continuous distribution bounded on the lower side.
It is a special case of the gamma distribution and can be reduced to the exponential distribution.
*Adapted with permission from ProModel Users Guide and Stat::Fit Users Guide.
669
har01307_APPA_668-673.indd 669 12/17/10 9:28 AM

670 Appendix A Continuous and Discrete Distributions in ProModel
The Erlang distribution has been used to specify the time required for a task. It has
been used extensively in reliability and in queuing theory, and thus in discrete event
simulation, because it can be viewed as a sum of exponentially distributed random
variables.
Distribution: Exponential
Syntax: E(a{,<s>, <ax >}) a mean
Application: The exponential distribution is a continuous distribution bounded on the
lower side. Its shape is always the same, starting at a finite value at the minimum and con-
tinuously decreasing at larger x.
The exponential distribution is frequently used to represent the time between random
occurrences, such as the time between arrivals at a specific location in a queuing model or
the time between failures in reliability models. It has also been used to represent the service
times of a specific operation.
Distribution: Gamma
Syntax: G(a, b{,<s>, <ax >}) a shape value, b scale value
Application: The gamma distribution is a continuous distribution bounded at the lower
side. It can be reduced to the exponential distribution and the Erlang distribution.
The gamma distribution has been used to represent lifetimes, lead times, personal
income data, a population about a stable equilibrium, interarrival times, and service
times.
Distribution: Inverse Gaussian

Syntax: IG(a, b{,<s>, <ax >}) a shape value, b scale value
Application: The Inverse Gaussian distribution is a continuous distribution with a
bound on the lower side. The Inverse Gaussian has been used to model the distribution of
particle size in aggregates, reliability and lifetimes, and repair time.
Distribution: Lognormal
Syntax: L(a, b{,<s>, <ax >}) a mean, b standard deviation
Application: The lognormal distribution is a continuous distribution bounded on the
lower side. By definition, the natural logarithm of a lognormal random variable is a normal
random variable.
The lognormal distribution is used in many different areas including the distribution
of particle size in naturally occurring aggregates, dust concentration in industrial atmo-
spheres, the distribution of minerals present in low concentrations, the duration of sick-
ness absence, physicians’ consulting time, lifetime distributions in reliability, distribution
of income, employee retention, and many applications modeling weight, height, and so
forth.
Distribution: Normal
Syntax: N(a, b{,<s>}) a mean, b standard deviation
har01307_APPA_668-673.indd 670 12/17/10 9:28 AM

Application: The normal distribution is an unbounded continuous distribution. It is

sometimes called a Gaussian distribution or the bell curve. Because of its property of rep-
resenting an increasing sum of small, independent errors, the normal distribution finds
many, many uses in statistics.
The normal distribution is frequently used to represent symmetrical data but suffers
from being unbounded in both directions. If the data are known to have a lower bound,
they may be better represented by suitable parameterization of the lognormal, Weibull,
or gamma distribution. If the data are known to have both upper and lower bounds, the
beta distribution can be used, although much work has been done on truncated normal
distributions.
Distribution: Pearson 5
Syntax: P5(a, b{,<s>, <ax >}) a shape value, b scale value
Application: The Pearson 5 distribution is a continuous distribution with a bound on
the lower side and is sometimes called the inverse gamma distribution due to the reciprocal
relationship between a Pearson 5 random variable and a gamma random variable.
The Pearson 5 distribution is useful for modeling time delays where some minimum delay
value is almost assured and the maximum time is unbounded and variably long, such as time
to complete a difficult task, time to respond to an emergency, time to repair a tool, and so forth.
Distribution: Pearson 6
Syntax: P6(a, b, c{,<s>, <ax >}) a shape value 1, b shape value 2, c scale value
Application: The Pearson 6 distribution is a continuous distribution bounded on the
low side. The Pearson 6 distribution is sometimes called the beta distribution of the second
kind due to the relationship of a Pearson 6 random variable to a beta random variable. It is
perhaps most often used to represent the time to accomplish a task.
Distribution: Triangular
Syntax: T(a, b, c{,<s>}) a minimum, b mode, c maximum
Application: The triangular distribution is a continuous distribution bounded on both
sides. The triangular distribution is often used when no or few data are available; it is rarely
an accurate representation of a data set (see Law 2007).
Distribution: Uniform
Syntax: U(a, b{,<s>}) a mean (minimum maximum)/2, b half range
maximum mean
Application: The uniform distribution is a continuous distribution bounded on both
sides. It is a special case of the beta distribution. Most random number generators provide
samples from the uniform distribution on (0,1) and then convert these samples to random
variates from other distributions.
The uniform distribution is used to represent a random variable with constant likelihood
of being in any interval between min and max. Note that the probability of either the min
or max value is 0; the end points do not occur. If the end points are necessary, try the sum
of two opposing right triangular distributions.
har01307_APPA_668-673.indd 671 12/17/10 9:28 AM

672 Appendix A Continuous and Discrete Distributions in ProModel
Distribution: Weibull
Syntax: W(a, b{,<s>, <ax >}) a shape value, b scale value
Application: The Weibull distribution is a continuous distribution bounded on the

lower side. The Weibull distribution reduces to the exponential distribution.
The Weibull distribution derived its popularity from its use to model the strength
of materials, and has since been used to model just about everything. In particular, the
Weibull distribution is used to represent wear-out lifetimes in reliability, wind speed,
rainfall intensity, health-related issues, germination, duration of industrial stoppages,
migratory systems, and thunderstorm data (see Johnson et al. 1994, p. 628; Shooman
1990, p. 190).
A.2 Discrete Distributions

Distribution: Binomial
Syntax: BI(a, b{,<s>}) a batch size, b probability of “success”
Application: The binomial distribution is a discrete distribution bounded by
[0, batch size]. Typically, it is used where a single trial is repeated over and over, such
as the tossing of a coin. The parameter b is the probability of the event, either heads or
tails, either occurring or not occurring. Each single trial is assumed to be independent
of all others.
The binomial distribution can be used to describe the number of defective items in a
batch; the number of people in a group of a particular type; and out of a group of employ-
ees, the number of employees who call in sick on a given day.
Distribution: Geometric
Syntax: GEO(a{,<s>}) a probability of “success”
Application: The geometric distribution is a discrete distribution bounded at 0 and un-

bounded on the high side. It is a special case of the negative binomial distribution. In
particular, it is the direct discrete analog for the continuous exponential distribution. The
geometric distribution has no history dependence.
The geometric distribution has been used for inventory demand, marketing survey re-
turns, a ticket control problem, and meteorological models (see Law 2007).
Distribution: Poisson
Syntax: P(a{,<s>}) a quantity
Application: The Poisson distribution is a discrete distribution bounded at 0 on the low

side and unbounded on the high side. The Poisson distribution is a limiting form of the
hypergeometric distribution.
The Poisson distribution finds frequent use because it represents the infrequent occur-
rence of events whose rate is constant. This includes many types of events in time and
space such as arrivals of telephone calls, defects in semiconductor manufacturing, defects
in all aspects of quality control, molecular distributions, stellar distributions, geographical
har01307_APPA_668-673.indd 672 12/17/10 9:28 AM

distributions of plants, shot noise, and so on. Note that the time between arrivals (defects)
is exponentially distributed.
References
Johnson, Norman L.; Samuel Kotz; and N. Balakrishnan. Continuous Univariate Distribu-
tions. Vol. 1. New York: John Wiley & Sons, 1994.
Law, Averill M. Simulation Modeling and Analysis, 4th ed. New York: McGraw-Hill, 2007.
Shooman, Martin L. Probabilistic Reliability: An Engineering Approach. Melbourne,
Florida: Robert E. Krieger, 1990.
har01307_APPA_668-673.indd 673 12/17/10 9:28 AM

APPENDIX B
CRITICAL VALUES FOR STUDENT’S t DISTRIBUTION AND STANDARD
NORMAL DISTRIBUTION
(Critical values for the standard normal distribution (z␣) appear in the last row with
df . That is z␣ t,␣.)
0 tdf,␣
␣ 0.40 ␣ 0.30 ␣ 0.20 ␣ 0.10 ␣ 0.05 ␣ 0.025 ␣ 0.01 ␣ 0.005
1 0.325 0.727 1.367 3.078 6.314 12.706 31.827 63.657

2 0.289 0.617 1.061 1.886 2.920 4.303 6.965 9.925
3 0.277 0.584 0.978 1.638 2.353 3.182 4.541 5.841
4 0.271 0.569 0.941 1.533 2.132 2.776 3.747 4.604
5 0.267 0.559 0.920 1.476 2.015 2.571 3.365 4.032
6 0.265 0.553 0.906 1.440 1.943 2.447 3.143 3.707
7 0.263 0.549 0.896 1.415 1.895 2.365 2.998 3.499
8 0.262 0.546 0.889 1.397 1.860 2.306 2.896 3.355
9 0.261 0.543 0.883 1.383 1.833 2.262 2.821 3.250
10 0.260 0.542 0.879 1.372 1.812 2.228 2.764 3.169
11 0.260 0.540 0.876 1.363 1.796 2.201 2.718 3.106
12 0.259 0.539 0.873 1.356 1.782 2.179 2.681 3.055
13 0.259 0.538 0.870 1.350 1.771 2.160 2.650 3.012
Degrees of Freedom (df )
14 0.258 0.537 0.868 1.345 1.761 2.145 2.624 2.977

15 0.258 0.536 0.866 1.341 1.753 2.131 2.602 2.947
16 0.258 0.535 0.865 1.337 1.746 2.120 2.583 2.921
17 0.257 0.534 0.863 1.333 1.740 2.110 2.567 2.898
18 0.257 0.534 0.862 1.330 1.734 2.101 2.552 2.878
19 0.257 0.533 0.861 1.328 1.729 2.093 2.539 2.861
20 0.257 0.533 0.860 1.325 1.725 2.086 2.528 2.845
21 0.257 0.532 0.859 1.323 1.721 2.080 2.518 2.831
22 0.256 0.532 0.858 1.321 1.717 2.074 2.508 2.819
23 0.256 0.532 0.858 1.319 1.714 2.069 2.500 2.807
24 0.256 0.531 0.857 1.316 1.708 2.060 2.485 2.797
25 0.256 0.531 0.856 1.316 1.708 2.060 2.485 2.787
26 0.256 0.531 0.856 1.315 1.706 2.056 2.479 2.779
27 0.256 0.531 0.855 1.314 1.703 2.052 2.473 2.771
28 0.256 0.530 0.855 1.313 1.701 2.048 2.467 2.763
29 0.256 0.530 0.854 1.310 1.697 2.042 2.457 2.756
30 0.256 0.530 0.854 1.310 1.697 2.042 2.457 2.750
40 0.255 0.529 0.851 1.303 1.684 2.021 2.423 2.704
60 0.254 0.527 0.848 1.296 1.671 2.000 2.390 2.660
120 0.254 0.526 0.845 1.289 1.658 1.980 2.358 2.617
0.253 0.524 0.842 1.282 1.645 1.960 2.326 2.576
674
har01307_APPB_674.indd 674 12/17/10 9:29 AM

APPENDIX C
F DISTRIBUTION FOR ␣ 0.05
har01307_APPC_675.indd 675
Numerator Degrees of Freedom [df (Treatment)]
1 2 3 4 5 6 7 8 9 10 12 14 16 20 24 30 40 50 100 200
3 10.13 9.55 9.28 9.12 9.01 8.94 8.89 8.85 8.81 8.79 8.74 8.71 8.69 8.66 8.64 8.62 8.59 8.58 8.55 8.54 8.54
4 7.71 6.94 6.59 6.39 6.26 6.16 6.09 6.04 6.00 5.96 5.91 5.87 5.84 5.80 5.77 5.75 5.72 5.70 5.66 5.65 5.63
5 6.61 5.79 5.41 5.19 5.05 4.95 4.88 4.82 4.77 4.74 4.68 4.64 4.60 4.56 4.53 4.50 4.46 4.44 4.41 4.39 4.36
6 5.99 5.14 4.76 4.53 4.39 4.28 4.21 4.15 4.10 4.06 4.00 3.96 3.92 3.87 3.84 3.81 3.77 3.75 3.71 3.69 3.67
7 5.59 4.74 4.35 4.12 3.97 3.87 3.79 3.73 3.68 3.64 3.57 3.53 3.49 3.44 3.41 3.38 3.34 3.32 3.27 3.25 3.23
8 5.32 4.46 4.07 3.84 3.69 3.58 3.50 3.44 3.39 3.35 3.28 3.24 3.20 3.15 3.12 3.08 3.04 3.02 2.97 2.95 2.93
9 5.12 4.26 3.86 3.63 3.48 3.37 3.29 3.23 3.18 3.14 3.07 3.03 2.99 2.94 2.90 2.86 2.83 2.80 2.76 2.73 2.71
10 4.96 4.10 3.71 3.48 3.33 3.22 3.14 3.07 3.02 2.98 2.91 2.86 2.83 2.77 2.74 2.70 2.66 2.64 2.59 2.56 2.54
11 4.84 3.98 3.59 3.36 3.20 3.09 3.01 2.95 2.90 2.85 2.79 2.74 2.70 2.65 2.61 2.57 2.53 2.51 2.46 2.43 2.41
12 4.75 3.89 3.49 3.26 3.11 3.00 2.91 2.85 2.80 2.75 2.69 2.64 2.60 2.54 2.51 2.47 2.43 2.40 2.35 2.32 2.30
13 4.67 3.81 3.41 3.18 3.03 2.92 2.83 2.77 2.71 2.67 2.60 2.55 2.51 2.46 2.42 2.38 2.34 2.31 2.26 2.23 2.21
14 4.60 3.74 3.34 3.11 2.96 2.85 2.76 2.70 2.65 2.60 2.53 2.48 2.44 2.39 2.35 2.31 2.27 2.24 2.19 2.16 2.13
15 4.54 3.68 3.29 3.06 2.90 2.79 2.71 2.64 2.59 2.54 2.48 2.42 2.38 2.33 2.29 2.25 2.20 2.18 2.12 2.10 2.07
16 4.49 3.63 3.24 3.01 2.85 2.74 2.66 2.59 2.54 2.49 2.42 2.37 2.33 2.28 2.24 2.19 2.15 2.12 2.07 2.04 2.01
17 4.45 3.59 3.20 2.96 2.81 2.70 2.61 2.55 2.49 2.45 2.38 2.33 2.29 2.23 2.19 2.15 2.10 2.08 2.02 1.99 1.96
18 4.41 3.55 3.16 2.93 2.77 2.66 2.58 2.51 2.46 2.41 2.34 2.29 2.25 2.19 2.15 2.11 2.06 2.04 1.98 1.95 1.92
19 4.38 3.52 3.13 2.90 2.74 2.63 2.54 2.48 2.42 2.38 2.31 2.26 2.21 2.16 2.11 2.07 2.03 2.00 1.94 1.91 1.88
20 4.35 3.49 3.10 2.87 2.71 2.60 2.51 2.45 2.39 2.35 2.28 2.23 2.18 2.12 2.08 2.04 1.99 1.97 1.91 1.88 1.84
22 4.30 3.44 3.05 2.82 2.66 2.55 2.46 2.40 2.34 2.30 2.23 2.17 2.13 2.07 2.03 1.98 1.94 1.91 1.85 1.82 1.78
24 4.26 3.40 3.01 2.78 2.62 2.51 2.42 2.36 2.30 2.25 2.18 2.13 2.09 2.03 1.98 1.94 1.89 1.86 1.80 1.77 1.73
26 4.23 3.37 2.98 2.74 2.59 2.47 2.39 2.32 2.27 2.22 2.15 2.09 2.05 1.99 1.95 1.90 1.85 1.82 1.76 1.73 1.69
28 4.20 3.34 2.95 2.71 2.56 2.45 2.36 2.29 2.24 2.19 2.12 2.06 2.02 1.96 1.91 1.87 1.82 1.79 1.73 1.69 1.66
30 4.17 3.32 2.92 2.69 2.53 2.42 2.33 2.27 2.21 2.16 2.09 2.04 1.99 1.93 1.89 1.84 1.79 1.76 1.70 1.66 1.62
35 4.12 3.27 2.87 2.64 2.49 2.37 2.29 2.22 2.16 2.11 2.04 1.99 1.94 1.88 1.83 1.79 1.74 1.70 1.63 1.60 1.56
Denominator Degrees of Freedom [df (Error)]

40 4.08 3.23 2.84 2.61 2.45 2.34 2.25 2.18 2.12 2.08 2.00 1.95 1.90 1.84 1.79 1.74 1.69 1.66 1.59 1.55 1.51
45 4.06 3.20 2.81 2.58 2.42 2.31 2.22 2.15 2.10 2.05 1.97 1.92 1.87 1.81 1.76 1.71 1.66 1.63 1.55 1.51 1.47
50 4.03 3.18 2.79 2.56 2.40 2.29 2.20 2.13 2.07 2.03 1.95 1.89 1.85 1.78 1.74 1.69 1.63 1.60 1.52 1.48 1.44
60 4.00 3.15 2.76 2.53 2.37 2.25 2.17 2.10 2.04 1.99 1.92 1.86 1.82 1.75 1.70 1.65 1.59 1.56 1.48 1.44 1.39
70 3.98 3.13 2.74 2.50 2.35 2.23 2.14 2.07 2.02 1.97 1.89 1.84 1.79 1.72 1.67 1.62 1.57 1.53 1.45 1.40 1.35
80 3.96 3.11 2.72 2.49 2.33 2.21 2.13 2.06 2.00 1.95 1.88 1.82 1.77 1.70 1.65 1.60 1.54 1.51 1.43 1.38 1.33
100 3.94 3.09 2.70 2.46 2.31 2.19 2.10 2.03 1.97 1.93 1.85 1.79 1.75 1.68 1.63 1.57 1.52 1.48 1.39 1.34 1.28
200 3.89 3.04 2.65 2.42 2.26 2.14 2.06 1.98 1.93 1.88 1.80 1.74 1.69 1.62 1.57 1.52 1.46 1.41 1.32 1.26 1.19
1.04 3.00 2.61 2.37 2.21 2.10 2.01 1.94 1.88 1.83 1.75 1.69 1.64 1.57 1.52 1.46 1.40 1.35 1.25 1.17 1.03
675
12/17/10 9:29 AM
APPENDIX D
CRITICAL VALUES FOR CHI-SQUARE DISTRIBUTION
␹ 2df,␣
␣ ⫽ 0.4 ␣ ⫽ 0.3 ␣ ⫽ 0.2 ␣ ⫽ 0.1 ␣ ⫽ 0.05 ␣ ⫽ 0.025 ␣ ⫽ 0.01 ␣ ⫽ 0.005
1 0.708 1.074 1.642 2.706 3.841 5.024 6.635 7.879

2 1.833 2.408 3.219 4.605 5.991 7.378 9.210 10.597
3 2.946 3.665 4.642 6.251 7.815 9.348 11.345 12.838
4 4.045 4.878 5.989 7.779 9.488 11.143 13.277 14.860
5 5.132 6.064 7.289 9.236 11.070 12.832 15.086 16.750
6 6.211 7.231 8.558 10.645 12.592 14.449 16.812 18.548
7 7.283 8.383 9.803 12.017 14.067 16.013 18.475 20.278
8 8.351 9.524 11.030 13.362 15.507 17.535 20.090 21.955
9 9.414 10.656 12.242 14.684 16.919 19.023 21.666 23.589
10 10.473 11.781 13.442 15.987 18.307 20.483 23.209 25.188
11 11.530 12.899 14.631 17.275 19.675 21.920 24.725 26.757
12 12.584 14.011 15.812 18.549 21.026 23.337 26.217 28.300
13 13.636 15.119 16.985 19.812 22.362 24.736 27.688 29.819
Degrees of Freedom (df)
14 14.685 16.222 18.151 21.064 23.685 26.119 29.141 31.319

15 15.733 17.322 19.311 22.307 24.996 27.488 30.578 32.801
16 16.780 18.418 20.465 23.542 26.296 28.845 32.000 34.267
17 17.824 19.511 21.615 24.769 27.587 30.191 33.409 35.718
18 18.868 20.601 22.760 25.989 28.869 31.526 34.805 37.156
19 19.910 21.689 23.900 27.204 30.144 32.852 36.191 38.582
20 20.951 22.775 25.038 28.412 31.410 34.170 37.566 39.997
21 21.992 23.858 26.171 29.615 32.671 35.479 38.932 41.401
22 23.031 24.939 27.301 30.813 33.924 36.781 40.289 42.796
23 24.069 26.018 28.429 32.007 35.172 38.076 41.638 44.181
24 25.106 27.096 29.553 33.196 36.415 39.364 42.980 45.558
25 26.143 28.172 30.675 34.382 37.652 40.646 44.314 46.928
26 27.179 29.246 31.795 35.563 38.885 41.923 45.642 48.290
27 28.214 30.319 32.912 36.741 40.113 43.195 46.963 49.645
28 29.249 31.391 34.027 37.916 41.337 44.461 48.278 50.994
29 30.283 32.461 35.139 39.087 42.557 45.722 49.588 52.335
30 31.316 33.530 36.250 40.256 43.773 46.979 50.892 53.672
40 41.622 44.165 47.269 51.805 55.758 59.342 63.691 66.766
60 62.135 65.226 68.972 74.397 79.082 83.298 88.379 91.952
120 123.289 127.616 132.806 140.233 146.567 152.211 158.950 163.648
676
har01307_APPD_676.indd 676 12/17/10 9:29 AM

I N D E X
Abandonment; see Reneging Automated storage/retrieval systems

Activity/Activities (AS/RS), 331
classification of, 36 configuring, 332–334
definition of, 36 modeling, 334
AGVS; see Automatic guided vehicles systems
Animation 5, 104 Batch mean method; see Interval batching
Animation processor, 105 Batch sizing, 308–309
Analysis of variance (ANOVA), 134, 259 move batch, 348
Arrivals process batch, 309
classifications, 175 production batch, 308
event-triggered, 177 transfer batch, 308
fluctuating, 176 Behavior, 163
periodic, 175 Beta distribution, 669
scheduled, 176 Binomial distribution, 139, 672
“As-is” model, 118 Bonferroni approach, 253
Assumptions Bottleneck, 301
documenting, 152
making, 124 Calling population, 52
Attribute, 163 Carousels, 335
Autocorrelated output data, 223 configuration, 335
Autocorrelation plot, 128 modeling, 335
AS/RS; see Automated storage/retrieval systems Central limit theorem, 224
Automatic guided vehicles systems (AGVS), 171 Chi-square multisample test, 144
controlling, 338 Common random number, 266, 282
designing, 339 example, 268–269
modeling, 339 random number stream, 266
path, 171 synchronization, 268
677
har01307_index_677-686.indd 677 12/17/10 9:49 AM

678 Index
Comparing systems, 243–272 Descriptive statistics, 212

hypothesis testing, 244–247 Design decisions, 6, 303–304, 343
more than two alternative designs, 253–264 Detail
two alternative designs, 250–253 incidental, 121
paired-t confidence interval, 250 level of detail, 118–119
Welch confidence interval, 248 Deterministic simulation, 70
variance reduction techniques, 265 Difference equations, 88
benefit, 270 Differential equations, 88
common random numbers, 266 Direct search techniques, 276
Conditional events, 88 Discrete-event simulation, 87–89,
Confidence interval, 216–218 assumptions, 89
assumptions, 222 conditional events, 89
paired-t, 250 example, 89–105
Welch, 248 scheduled events, 88
Continuous distributions, 151–152, 669–673 simulation clock, 92
Continuous simulation, 88 Discrete probability distributions, 151–152, 672–673
Constraints, theory of, 300 Distribution fitting, 134–148
Continuous process systems, 319 in the absence of data, 148–149
Control empirical distribution, 135
definition of, 37 frequency distribution, 135
examples of, 37 theoretical distribution, 135
Conveyors Downtimes and repairs, 181
modeling, 328 Drum-buffer-rope (DBR) control, 309, 311
network, 330 Dynamic simulation, 58
single-section, 329
operational characteristics, 327 EA; see Evolution algorithm, 277–281
types, 325–327 Empirical distribution, 135
CONWIP, 311 Emulation, 315, 164
Cost modeling, 189 Entity/Entities
Cranes and hoists aggregating, 166
management, 340 arrivals, 175–177
modeling, 340 attributes, 92, 165, 475–479
CRN; see Common random number, 266 definition of, 36, 165
discrete, 36
Data high-rate processing, 167
categorization, 118–120 modeling of, 165–167
collecting of, 121–124 movement, 177
documentation and approval, 152 operations, 172–175
source, 120–121 resolution, 166
Data analysis software, 123 to include in modeling, 166
Data collection and analysis, 115–159 types of, 36
guidelines, 116–118 Entity flow diagram, 122
Data documentation and approval, 152 Erlang distribution, 669–670
examples of, 152–155 Error
Decision variables, 43, 303 type I, 245
Description of operation, 132 type II, 245
har01307_index_677-686.indd 678 12/17/10 9:49 AM

Index 679
Event-triggered arrivals, 177 Inferential statistics, 212

Events Initial seed values, 213–214, 267
conditional, 88 Interdependencies, 38, 127
scheduled, 88 Internal performance criteria, 347–348
Evolutionary algorithm, 277–281 Interval batching, 234
combining with simulation, 278 Interval estimate, 216–218
global optimum, 280 International Journal of Modeling and Simulation,
local optimal, 280 18–22
types of, 278–280 Inventory control, 312–313
steps of, 278
Exponential distribution, 141–142, 669, Just in time (JIT), 299, 310
670–672
External performance criteria, 347
Kanban, 301, 311
Kolmogorov-Smirnov two sample test, 134
Factorial design and optimization, 264
Fitting theoretical distribution to data,
142–144 Laffer curve, 164
Flow control (service), 352 Large scale modeling, 188–189
Fluctuating arrivals, 176 Layout
Frequency distribution, 135 plant (manufacturing), 306–308
continuous, 136 service system, 351
histogram, 136 Linear conguential generators (LCG), 64
discrete, 135–136 mixed, 67
Functional validity, 202 multiplicative, 67
Little’s Law, 53
Gamma distribution, 669–671 Location/locations
Geometric distribution, 672 definition, 165
Global optimum, 280 resolution, 168
Global variables, 186 selection criteria, 171
to include in modeling, 167–168
Half-width, 217 Local optimal, 280
Hand calculation, 50 Lognormal distribution, 670
Heuristic optimization methods, 276 LSD test, 262
Histogram, 136
Hoist; see Cranes and hoists Machine load and unload time, 316
Homogeneity test, 132 Machine setup, 316
HP simulation case, 207–209 Macros, 575
Hypothesis testing, 243–247 Management decisions, 349
Manufacturing automation, 300
Incidental details, 123 Manufacturing modeling techniques,
Industrial vehicles, 330–331 316–319
characteristics, 330 continuous process systems, 319
modeling, 331 machine load and unload time, 316
Independent observations, 234–239 machine setup, 316
batch means technique, 234–239 rework and scrap, 317
replications, 234 transfer machines, 317
har01307_index_677-686.indd 679 12/17/10 9:49 AM

680 Index
Manufacturing systems determining, 203–205

applications of simulation in, 309–315 example, 205–209
automation, 305–306 functional validity, 202
batch sizing, 308–309 maintaining, 205
emulation, 315–316 Model verification, 194–195
inventory control, 312–313 definition of, 205
method analysis, 305–306 establishing a standard, 198
plant layout, 306–308 preventive measures, 197
production control, 309–312 techniques, 198–201
production scheduling, 314–315 semantic errors, 196
real-time control, 315–316 syntax errors, 196
supply chain management, 313–314 Model verification and validation, 193–209
characteristics, 315 importance, 194
decision range, 303 reasons for neglect, 194–193
definition, 299 practices that facilitate, 195–196
simulation, 303–304 Modeling conveyors
terminology, 301–303 network, 330
Material handling single section, 329
classification of, 324–325 Modeling interface module, 102–103
principles of, 323–324 Modeling manufacturing systems, 299–322
Mean, 126 Modeling processor, 103
Mean time between failures, 198 Module, Simulation software
Mean time to repair, 198 modeling interface, 103–104
Method selection (service), 351–352 output interface, 105
Model simulation interface, 104
“as-is,” 118 Monte Carlo methods, 5, 83
building; see Model building Move batch, 308
manufacturing systems; see Manufacturing systems, Movement, 177
service systems; see Service systems MRP, 310, 315
“to-be, 118 MTBF, 183, 484
verification and validation; see Model verification MTTR, 183, 484
and validation Multiple comparison test, 262
Model building, 161–192 protected LSD test, 262
complexity and utility, 164
cost, 189 Nonhomogenous data, 133
definition of, 164 Nonlinear optimization technique, 277
large scale, 188–189 Nonsequential routing, 172
object-based, 163 Nonstationary, 133, 223
object classification, 163 Nonterminating simulations, 226–227
operational elements, 171 determine the run length, 239–240
paradigm, 162–163 determine the warm-up period,
rare occurrence, 187–188 229–233
structural elements, 165 obtain the sample observation, 212–213
Model validation, 193–196, 202–209 Normal distribution, 140–141, 671
definition of, 194 Numerical data, 119
har01307_index_677-686.indd 680 12/17/10 9:49 AM

Index 681
Object-based modeling, 163 Point estimates, 215–216

Object function, 274 Poisson distribution, 672
Object-oriented programming (OOP), 163 Population, 245–250
Operation, description of, 122 Preemption, 177
Operational data, 118–119 Priority, 177
Operational efficiency, 281 Probability distributions, 309
Operational elements, 171 Process batch (manufacturing), 350
accessing location and resource, 178 Process design (service), 307
downtimes and repairs, 181 Process layout (manufacturing), 307
entity and resource movement, 178 Processor
entity arrivals, 175 animation, 104
entity operations, 172 modeling, 103
resource scheduling, 179 output, 105
routing, 171 simulation, 104
use of programming logic, 185 Product layout, 307
Optimization Production batch, 308
direct search techniques, 276 Production control, 309–312
evolutionary algorithm, 277 drum-buffer-rope (DBR) control,
example, 284–286 309, 311
general procedure, 282–284 pull control, 309, 310–311
heuristics methods, 276 push control, 309, 310
nonlinear techniques, 277 Production lot, 308
steps, 275 Production scheduling, 314–315
strategic and tactical issues, 281–294 ProModel, overview, 103–107
system, 44–46 Protected LSD test, 262
value of, 273 Pull control, 310–311
Optimization, factorial design and, 264 Push control, 310
Optimization, system, 44–46
Operations research, 51–54 Queuing theory, 52–54
Output interface module, 105
Output processor, 105 Random behavior, 59
Overlapped production, 309 Random number
generation, 63
Paired-t confidence interval, 250–252 generator, 63
Part family layout, 307 initial seed values, 213–214
Path networks, 171 pseudo, 64
Paths stream, 268
definition, 165 Random number stream, 268
modeling of, 171 seed number, 269
Pearson 5 distribution, 669, 671 synchronization, 268
Pearson 6 distribution, 669, 671 Random variates, 68, 59
Performance measures, 347–348 Rare occurrence modeling, 187
Periodic arrivals, 175–176 Real-time control, 315
Plan-do-check-act (PDCA), 300 Reneging, 346
Plant layout, 306–308 Repairs, downtimes and, 181
har01307_index_677-686.indd 681 12/17/10 9:49 AM

682 Index
Replication, 8, 213–214 modeling, 345–359

Replications, number of, 218–222, performance measures, 347
228–229 types of, 352
Replications, via observations, 234 use of simulation in, 348
Resource/Resources SimRunner, 274
availability of, 180–181 Simulation
classification of, 36 advantages over alternatives, 7–8
consumable, 170 application, 11
definition of, 165, 169 characteristics, 7
scheduling, 179–180 definition of, 5
to include in modeling, 170 economic justification of, 18–22
transport, 170 example, 91
travel time of, 170 modeling; see Model building
Resource scheduling, 179–180 nonterminating, 226–227
Response surface, 263 objectives 15, 47
Response variables, 44 optimization; see Optimization
Rework and scrap, 317 output analysis; see Statistical analysis of
Robots simulation output
control, 341 procedure, 15
definition of, 341 replications, 213–214
modeling, 342 selection criteria, 11–12
Route stop; see Location spreadsheet, 74
Routing surge of popularity, 10
decision rules, 172 system; see System
nonsequential, 172 terminating, 226
recirculation, 172 types, 57–61
Run, 8 static, 58
Run length, determine, 239–249 stochastic, 58
user prerequisite, 13–14
Sample, 213 Simulation interface module, 104
Sample size, 212, 218 Simulation optimization; see
Scatter plot, 128 Optimization
Scheduled arrivals, 176 Simulation processor, 104
Scheduled events, 88 Simulation software
Scientific method, 9 module
Scrap, rework and, 317 modeling interface, 103–104
Seed value, 213–234 output interface, 105
Selection rules, task, 179 simulation interface, 104
Semantic errors, 196 processor
Sensitivity analysis, 125 animation, 104
Service systems modeling, 103–104
application of simulation, 346 output, 105
characteristics, 345, 346–347 simulation, 104
definition of, 345 Single-section conveyor,
example, 355–359 modeling, 44
har01307_index_677-686.indd 682 12/17/10 9:49 AM

Index 683
Special computational tools, 54 variables, 43–45

Spreadsheets, 50–51 decision, 43
St. John Hospital case, 206–207 response, 44
Staff planning (service), 351 state, 44
Stat::Fit, 143 System complexity, 38– 40
Static simulation, 58 interdependencies, 38
Statistical accumulators, 92–93 variability, 39
simple-average, 93, 86 Systems approach, 46–48
time-average, 93 Systems analysis techniques
Statistical terminology, 126–127 hand calculation, 48–50
Statistical analysis of numerical data, operations research, 51–54
125–134 special computational tools, 54
data analysis software, 125 spreadsheets, 50–51
Stat::Fit, 143
tests for identically distributed data, Task selection rules, 179
(homogeneity), 132 Terminating simulations, 223–224,
tests for independence, 127 225–229
autocorrelation plot, 128 determine the number of replication, 228
scatter plot, 128 selecting the initial model state, 228
Statistical analysis of simulation output, selecting the terminating event, 228
211–240 Theoretical distributions, 136–142
autocorrelation, 223 binomial distribution, 138, 672
batch means method, 234 exponential distribution, 141, 670
central limit theorem, 224 normal distribution, 140–141, 670–671
independent observations, 222 triangular distribution, 140, 671
number of replications, 218–221 uniform distribution, 139, 671
performance estimation, 214–218 Theory of Constraints, 300, 311
interval estimate, 216–218 Time variant, 133
point estimate, 215–216 “To-be” model, 118
Standard deviation, 127 Transfer batch, 308
Steady state, 226 Transfer machines, 317–318
Stochastic simulation, 58 Transient state, 225–226
Structural data, 138 Triangular distribution, 140, 671
Structural elements, 165 Type I error, 245
Supply chain management, 313–314 Type II error, 245
Synchronizing random
numbers, 268 Uniform distribution, 139–140, 671
Syntax errors, 196
System Variability, 39–40
comparing; see Comparing systems Variables, 43–45
complexity, 38 decisional, 43
definition of, 34 global, 186
elements, 35 local, 186
optimization, 44–45 response, 44
performance metrics, 41–43 state, 92
har01307_index_677-686.indd 683 12/17/10 9:49 AM

684 Index
Variance, 126 Welch confidence interval, 248

Variance reduction techniques, 218, 265 Welch moving-average method, 232
benefit, 270 Winter Simulation Conference, 22
common random numbers, 266 WIP, 310–311
Warm-up period, 229–233 Yield, 302

Weibull distribution, 672
har01307_index_677-686.indd 684 12/17/10 9:49 AM

FV609-Simulation Using ProModel - Charles Harrell

Uploaded by

Copyright:

Available Formats

FV609-Simulation Using ProModel - Charles Harrell

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

FV609-Simulation Using ProModel - Charles Harrell

Uploaded by

Copyright:

Available Formats

T H I R D E D I T I O N

Dr. Charles Harrell

Dr. Biman K. Ghosh, Project Leader

Dr. Royce O. Bowden, Jr.

har01307_fm_i-xviii.indd i 12/21/10 8:25 AM

SIMULATION USING PROMODEL, THIRD EDITION

Library of Congress Cataloging-in-Publication Data

har01307_fm_i-xviii.indd ii 12/21/10 8:25 AM

har01307_fm_i-xviii.indd iii 12/21/10 8:25 AM

har01307_fm_i-xviii.indd iv 12/21/10 8:25 AM

har01307_fm_i-xviii.indd v 12/21/10 8:25 AM

PART I 2.8 The Systems Approach 46

har01307_fm_i-xviii.indd vi 12/21/10 8:25 AM

3.5 Simple Spreadsheet Simulation 74 4.8 Summary 112

har01307_fm_i-xviii.indd vii 12/21/10 8:25 AM

5.10 Bounded versus Boundless 7 Model Verification and

har01307_fm_i-xviii.indd viii 12/21/10 8:25 AM

8.4.2 Nonterminating More Than Two Alternative

har01307_fm_i-xviii.indd ix 12/21/10 8:25 AM

10.7.2 Simulation Model and 11.7 Summary 320

har01307_fm_i-xviii.indd x 12/21/10 8:25 AM

12.11 Summary 343 L1.4 Run-Time Menus and

har01307_fm_i-xviii.indd xi 12/21/10 8:25 AM

L4.7 Batching Multiple Entities of L6.2 Cycle Time 479

har01307_fm_i-xviii.indd xii 12/21/10 8:25 AM

8 Simulation Output Analysis 527 L10.2.2 Multiterm Objective

har01307_fm_i-xviii.indd xiii 12/21/10 8:25 AM

L13.5 Modeling a University Appendix B Critical Values for Student’s t

har01307_fm_i-xviii.indd xiv 12/21/10 8:25 AM

Simulation is a modeling and analysis technique used to evaluate and improve

har01307_fm_i-xviii.indd xv 12/21/10 8:25 AM

problem-solving situations. In simulation courses we have taught over the years,

Electronic Textbook Options

har01307_fm_i-xviii.indd xvi 12/21/10 8:25 AM

McGraw-Hill Higher Education and

har01307_fm_i-xviii.indd xvii 12/21/10 8:25 AM

No work of this magnitude is performed in a vacuum, independently of the help

har01307_fm_i-xviii.indd xviii 12/21/10 8:25 AM

har01307_ch01_001-032.indd 1 12/17/10 9:30 AM

har01307_ch01_001-032.indd 3 12/17/10 9:30 AM

har01307_ch01_001-032.indd 4 12/17/10 9:30 AM

• What are the qualifications for doing simulation?

1.2 What Is Simulation?

Simulation is the imitation of a dynamic system

In practice, simulation is usually performed using commercial simulation

har01307_ch01_001-032.indd 5 12/17/10 9:30 AM

model performance. These methods are sometimes referred to as Monte Carlo

1.3 Why Simulate?

Trial-and-error approaches are

With the emphasis today on time-based competition, traditional trial-and-error

har01307_ch01_001-032.indd 6 12/17/10 9:30 AM

of trial-and-error approaches in designing manufacturing systems, Solberg (1988)

har01307_ch01_001-032.indd 7 12/17/10 9:30 AM

other analysis techniques. Simulation gives systems planners unlimited freedom

1.4 Doing Simulation

har01307_ch01_001-032.indd 8 12/17/10 9:30 AM

har01307_ch01_001-032.indd 9 12/17/10 9:30 AM

1.5 Use of Simulation

har01307_ch01_001-032.indd 10 12/17/10 9:30 AM

On a smaller scale, simulation is being used to provide interactive, computer-