Autonomous Guided Vehicles: Hamed Fazlollahtabar Mohammad Saidi-Mehrabad

Studies in Systems, Decision and Control 20
Hamed Fazlollahtabar
Mohammad Saidi-Mehrabad
Autonomous
Guided
Vehicles
Methods and Models for Optimal Path
Planning
Studies in Systems, Decision and Control
Volume 20
Series editor
Janusz Kacprzyk, Polish Academy of Sciences, Warsaw, Poland
e-mail: kacprzyk@ibspan.waw.pl
About this Series
The series "Studies in Systems, Decision and Control" (SSDC) covers both new
developments and advances, as well as the state of the art, in the various areas of
broadly perceived systems, decision making and control- quickly, up to date and
with a high quality. The intent is to cover the theory, applications, and perspec-
tives on the state of the art and future developments relevant to systems, decision
making, control, complex processes and related areas, as embedded in the fields of
engineering, computer science, physics, economics, social and life sciences, as well
as the paradigms and methodologies behind them. The series contains monographs,
textbooks, lecture notes and edited volumes in systems, decision making and con-
trol spanning the areas of Cyber-Physical Systems, Autonomous Systems, Sensor
Networks, Control Systems, Energy Systems, Automotive Systems, Biological Sys-
tems, Vehicular Networking and Connected Vehicles, Aerospace Systems, Automa-
tion, Manufacturing, Smart Grids, Nonlinear Systems, Power Systems, Robotics,
Social Systems, Economic Systems and other. Of particular value to both the con-
tributors and the readership are the short publication timeframe and the world-wide
distribution and exposure which enable both a wide and rapid dissemination of
research output.
More information about this series at http://www.springer.com/series/13304

Hamed Fazlollahtabar · Mohammad Saidi-Mehrabad
Autonomous Guided Vehicles

Methods and Models for Optimal
Path Planning
ABC
Hamed Fazlollahtabar Mohammad Saidi-Mehrabad
Faculty of Industrial Engineering Faculty of Industrial Engineering
Iran University of Science and Technology Center of Excellence for Advance
Tehran Manufacturing
Iran Iran University of Science and Technology
Tehran
Iran
ISSN 2198-4182 ISSN 2198-4190 (electronic)

Studies in Systems, Decision and Control
ISBN 978-3-319-14746-8 ISBN 978-3-319-14747-5 (eBook)
DOI 10.1007/978-3-319-14747-5
Library of Congress Control Number: 2014958756
Springer Cham Heidelberg New York Dordrecht London

c Springer International Publishing Switzerland 2015
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, express or implied, with respect to the material contained herein or for any
errors or omissions that may have been made.
Printed on acid-free paper
Springer International Publishing AG Switzerland is part of Springer Science+Business Media

(www.springer.com)
Preface
The Automated Guided Vehicle (AGV) is an automated guided cart that follows a
guided path. This equipment is widely used in industrial fields and places of
physical distribution. We have developed a new type of AGV that has additional
functions such as following the motion of people and avoiding obstacles on the
course it is traveling on. Therefore, it can work with service personnel even
though there is no pre-determined guided path. In general, an AGV is a driverless
transport system used for horizontal movement of materials. AGVs are especially
used for the internal and external transport of materials. Moreover, vision or lasers
can be used for determining the movement of an AGV.
Since their introduction in 1955, AGVs have found widespread industrial
applications. AGVs are now found in all types of industries, with the only
restrictions on their use mainly resulting from the dimensions of the goods to be
transported or spatial considerations. Many applications of AGVs are technically
feasible, but the purchase and implementation of such systems is usually based on
economic considerations. The uses of AGVs can be divided into four main areas
of application:
1) supply and disposal at storage and production areas;
2) production-integrated application of AGV trucks as assembly platforms;
3) retrieval, especially in wholesale trade; and
4) supply and disposal in special areas, such as hospitals and offices. In all of
these settings.
AGVs have been found to reduce the damage to inventory, make production
scheduling more flexible, and reduce staffing needs. But, as with any other major
capital decision, implementation of these systems must be undertaken cautiously.
The AGVs were traditionally employed in manufacturing systems, but have
recently extended their popularity to many other industrial applications, such as
goods transportation in warehouses and container transshipment systems at
container terminals. Industrial transportation systems using AGVs are used in
warehouses and manufactures.
An AGV system is a fully automated industrial transport system that makes use
of numerous AGVs. An AGV is a battery powered, computer controlled,
VI Preface
unmanned vehicle that is capable of transporting goods in an industrial

environment. The AGVs are restricted to follow pre-determined paths, on which
they navigate by using sensors and stationary beacons. Transports, which consist
of picking up a load on a certain location and transporting it to a certain
destination, are generated by an external system and have to be assigned to the
AGVs. Traditionally, one computer system (central server) is in charge of
numerous complex and time-consuming tasks such as routing, collision avoidance
or deadlock avoidance.
The AGVs themselves have little autonomy. Although this planning is efficient,
it lacks flexibility. In highly dynamic systems, where the situation changes
frequently, problems are experienced. A new and innovative architecture has been
developed that additional qualities, like flexibility and openness, to cope with the
highly dynamic environments.
AGV usage is growing. One reason is that as manufacturers strive to become
more competitive, they are adopting flexible manufacturing systems (FMS). These
systems integrate automated material handling systems, robots, numerically
controlled machine tools, and automated inspection stations. Flexible
manufacturing systems offer high capital utilization and reduced direct labor costs.
They also reduce work-in-process inventories and make it possible to work with
shorter lead times. Because the systems are flexible, they are more responsive to
changes in production requirements. These systems offer high product quality and
increased productivity.
In this book different models and methods for optimizing AGV’s path are
developed. The chapters are useful for academic courses in graduate studies. The
models are classified into single and multiple AGVs. Also, models under
uncertain conditions including fuzzy and statistical approaches are proposed.
The book chapters are organized as follows. In the first chapter an overview of
the models and methods of scheduling and routing problems for AGVs is
presented. In chapter 2, material flow optimization using AGV is discussed. In
chapter 3, a manufacturing system equipped with AGV is modeled using a
nonlinear stochastic mathematical programming. In chapter 4, reliability of AGVs
is considered and optimized. In chapter 5, path planning under uncertainty is
explained and modeled. In chapter 6, AGV routing under stochastic run time is
modeled by cross entropy concept. In chapter 7, expert system is designed for
AGV optimal path planning considering multiple criteria. In chapter 8, intelligent
Preface VII
agent is developed for path planning of AGV. In chapter 9, delay optimization for
multiple AGVs is developed and modeled. In chapter 10, Markovian modeling is
proposed for evaluation of multiple AGV system. In chapter 11, producer
behavior in an AGV equipped manufacturing system is analyzed. In the last
chapter risk evaluation for AGVs in a manufacturing system is described.
January 2015 Hamed Fazlollahtabar

Tehran, Iran Mohammad Saidi-Mehrabad
Acknowledgments
We would like to express our gratitude to the many people who saw us through
this book; to all those who provided support, talked things over, read, wrote,
offered comments, allowed us to quote their remarks and assisted in the editing,
proofreading and design.
We would like to thank Iran University of Science and Technology for enabling
us to publish this book. Above all we want to thank our families, who supported
and encouraged us in spite of all the time it took us away from them. It was a long
and difficult journey for them.
We would like to thank Prof. Nezam Mahdavi-Amiri for helping us in the
process of selection and editing. Thanks to our publisher who encouraged us.
Thanks to Prof. Abdolreza Sheikholeslami - without you this book would never
find its way to the academician.
Last and not least: We beg forgiveness of all those who have been with us over
the course of the years and whose names we have failed to mention.
H. Fazlollahtabar
M. Saidi-Mehrabad
Contents
1 Models for AGVs’ Scheduling and Routing ......................................... 1

1.1 Summary ......................................................................................... 1
1.2 Scheduling Models .......................................................................... 2
1.3 Routing Models ............................................................................... 4
1.4 Conclusions ..................................................................................... 12
References ................................................................................................ 13
2 Analytical Material Flow Model for AGV System .............................. 17

2.1 Summary ......................................................................................... 17
2.2 Statement of the Problem ................................................................ 19
2.3 Mathematical Model ....................................................................... 20
2.4 Conclusions ..................................................................................... 24
References ................................................................................................ 25
3 Nonlinear Stochastic Model for AGV System ...................................... 27

3.1 Summary ......................................................................................... 27
3.4 Conclusions ..................................................................................... 37
References ................................................................................................ 38
4 Reliability Model for AGV .................................................................... 41

4.1 Summary ......................................................................................... 41
4.4 Conclusions ..................................................................................... 53
References ................................................................................................ 54
5 Uncertain Optimal Path for AGV System ............................................ 57

5.1 Summary ......................................................................................... 57
XII Contents
5.3 Mathematical Modeling .................................................................. 60

5.4 Conclusions ..................................................................................... 75
References ................................................................................................ 75
6 Cross Entropy Model for AGV Routing Time ..................................... 79

6.1 Summary ......................................................................................... 79
6.4 Conclusions ..................................................................................... 90
References ................................................................................................ 90
7 Neuro-Fuzzy-Regression Expert System for AGV Optimal Path ...... 93

7.1 Summary ......................................................................................... 93
7.4 Conclusions ..................................................................................... 112
References ................................................................................................ 112
8 Optimal Path for AGV System with Intelligent Agents ...................... 117
8.1 Summary ......................................................................................... 117
8.4 Conclusions ..................................................................................... 130
References ................................................................................................ 130
9 Earliness/Tardiness for a Multiple AGV System ................................. 133

9.1 Summary ......................................................................................... 133
9.4 Conclusions ..................................................................................... 144
References ................................................................................................ 144
10 Markovian Model for Multiple AGV System ...................................... 147

10.1 Summary ......................................................................................... 147
10.4 Conclusions ..................................................................................... 166
References ................................................................................................ 166
Contents XIII
11 Producer's Behavior Analysis for AGV System .................................. 169

11.1 Summary ......................................................................................... 169
11.4 Conclusions ..................................................................................... 184
References ................................................................................................ 185
12 Risk for Multiple AGV System ............................................................. 189

12.1 Summary ......................................................................................... 189
12.4 Conclusions ..................................................................................... 202
References ................................................................................................ 203
Chapter 1
Models for AGVs’ Scheduling and Routing
1.1 Summary
An automated guided vehicle (AGV) is a driverless material handling system used

for horizontal movement of materials. AGVs were introduced in 1955 (Muller,
1983). The use of AGVs has grown enormously since their introduction. The
number of areas of application and variation in types has increased significantly.
AGVs can be used in inside and outside environments, such as manufacturing,
distribution, transshipment and (external) transportation areas. At manufacturing
areas, AGVs are used to transport all types of materials related to the
manufacturing process. According to Gotting (2000) over 20,000 AGVs were used
in industrial applications. The author states that the usage of AGVs will pay off
for environments with repeating transportation patterns. Examples of these
environments are distribution, transshipment and transportation systems.
Warehouses and cross docking centers are examples of distribution areas. AGVs
are used in these areas for the internal transport of, for example, pallets between
the various departments, such as receiving, storage, sorting and shipment areas. At
transshipment systems, such as container terminals, AGVs take care of the
transport of products between the various modes of transport. Gotting (2000)
presented an overview of available technology for automation in container
terminals. Furthermore, navigation and vehicle guidance systems applicable in
various indoor/outdoor environments are described. Haefner and Bieschke (1998)
stated that AGV systems can provide benefits to both the port and its customers by
executing transportation requests between vessels and inland transportation.
Namely, in non-automated terminals this transportation process is one of the
least efficient and most costly processes. AGVs can also be used in the outdoor
transportation process. An example of such a transportation system is an
underground automated transportation system with AGVs travelling in tubes
between companies and an airport (see Van der Heijden et al., 2002 a,b). In such
systems, we notice a high traffic density and long tube driving times. It has even
been studied if AGVs can be used as a communication system between work
stations (see Maughan and Lewis, 2000). Clearly, the specifications of AGVs
differ per environment. To transport a container, the capacity of an AGV should at
© Springer International Publishing Switzerland 2015 1

H. Fazlollahtabar and M. Saidi-Mehrabad, Autonomous Guided Vehicles,
Studies in Systems, Decision and Control 20, DOI: 10.1007/978-3-319-14747-5_1
2 1 Models for AGVs’ Scheduling and Routing
least be equal to 40 tons. Less capacity is required for the transport of pallets at
warehouses. Furthermore, at container terminals self-lifting automated guided
vehicles (ALVs) are used. For this type of AGV no other equipment is required to
transfer a load to the vehicle. Vis and Harika (2004) and Yang et al. (2004)
discussed this new type of AGVs in more detail.
In this chapter, we will discuss literature concerning the usage of AGVs in
manufacturing and the new areas of application, namely distribution,
transshipment and transportation systems. The most important differences
between traditional and new areas of application are the number of AGVs used,
the number of transportation requests, the occupancy degree of AGVs, the
distances to be travelled and the number of pick-up and delivery points where
transportation requests become available. At manufacturing systems, a small
number of AGVs with relatively low occupancy degrees are used to transport a
small number of requests over short distances between a few pick-up and delivery
points. For continuous mass transport in these systems conveyors are used instead
of large numbers of AGVs (see Gotting, 2000). In contrast to manufacturing
systems, large numbers of AGVs (up to 400; Van der Heijden et al., 2002a) were
used to execute a large number of repeating transportation tasks at container
terminals and external transportation systems. Furthermore, operational
conditions, such as weather conditions and spatial dimensions, found in outside
environments (container terminals and external transportation systems) differ from
the operational conditions in inside areas (manufacturing and distribution
systems).
1.2 Scheduling Models
The literature discussed so far on scheduling of AGVs hardly considers (side)

constraints such as capacity constraints of the machines where transportation jobs
become available, schedules of other types of equipment and limited parking
space for vehicles. These side constraints become more and more important in real
life situations with large AGV systems. To take these constraints into account
more attention should be paid to integrated scheduling of different types of
material handling equipment, which also meet space and capacity requirements.
1.2.1 Artificial Intelligence

The increased use of FMS to efficiently provide customers with diversified
products has created a significant set of operational challenges. Although
extensive research has been conducted on design and operational problems of
automated manufacturing systems, many problems remain unsolved. In particular,
the scheduling task, the control problem during the operation, is of importance
owing to the dynamic nature of the FMS such as flexible parts, tools and AGV
routings. The FMS scheduling problem has been tackled by various traditional
1.2 Scheduling Models 3
optimization techniques. While these methods can give an optimal solution to

small-scale problems, they are often inefficient when applied to larger-scale
problems. Jerald et al. (2005) designed different scheduling mechanisms to
generate optimum scheduling; these included non-traditional approaches such as
genetic algorithm (GA), simulated annealing (SA) algorithm, memetic algorithm
(MA) and particle swarm algorithm (PSA) by considering multiple objectives, i.e.,
minimizing the idle time of the machine and minimizing the total penalty cost for
not meeting the deadline concurrently. The MA presented was essentially a GA
with an element of SA. The results of the different optimization algorithms (MA,
GA, SA, and PSA) were compared and conclusions were presented.
Solving multi-objective scientific and engineering problems is, generally, a
very difficult goal. In these optimization problems, the objectives often conflict
across a high-dimensional problem space and require extensive computational
resources. Sarvana Sankar et al. (2006) developed a migration model of
parallelization for a GA based multi-objective evolutionary algorithm (MOEA).
The MOEA generated a near-optimal schedule by simultaneously achieving two
contradicting objectives of an FMS. The parallel implementation of the migration
model showed a speedup in computation time and needed less objective function
evaluations when compared to a single-population algorithm. So, even for a
single-processor computer, implementing the parallel algorithm in a serial manner
(pseudo-parallel) delivers better results. Two versions of the migration model
were constructed and the performance of two parallel GAs was compared for their
effectiveness in bringing genetic diversity and minimizing the total number of
functional evaluations.
1.2.2 Simulation
Gaur et al. (2003) studied the problem of scheduling an AGV in a flexible
manufacturing system while minimizing completion times. A vehicle needs to
visit each site only after its release time and before its due time. Sabuncuoglu
(1998) used simulation to test various AGV scheduling rules.
Fazlollahtabar et al. (2012) concerned with applying tandem automated guided
vehicle (TAGV) configurations as material handling devices and optimizing the
production time considering the effective time parameters in a flexible automated
manufacturing system (FAMS) using Monte Carlo simulation. Due to different
configurations of TAGVs in an FAMS, the material handling activities are
performed. With respect to various stochastic time parameters and the TAGV
defects during material handling processes, sample data were collected and their
corresponding probability distributions were fitted. Using the probability
distributions, they modelled the TAGV material handling problem via Monte
Carlo simulation. The effectiveness of the proposed model was illustrated in a
case study.
1.3 Routing Models
If the dispatching decision is carried out, a route and schedule should be planned
for the AGV to move the job from its origin to its destination within the AGV
network. A route implies the path which should be taken by the AGV when
making a pick-up or delivery. The related schedule gives arrival and departure
times of the AGV at each part, pick-up and delivery point and intersection during
the route to ensure collision free routing. The selection of a certain route and
schedule is effective on the performance of the system. The longer it takes to
transport a job, the fewer the jobs that can be handled within a certain time.
Therefore, one of the objectives of the routing of AGVs is to minimize
transportation times. Algorithms have to be developed to solve the routing
problem. Two categories of algorithms can be distinguished, namely static and
dynamic algorithms.
Analogies between these problems from transportation literature and routing
and scheduling problems for AGVs in automated guided vehicle systems are clear.
A number of loads at various locations have to be transported by vehicles at a
certain start time or at a certain moment within a time window. However, the use
of the described models from transportation literature is not always possible.
These models do not take into account congestion in the system. Furthermore,
most models are not developed to deal with real time response to dynamically
changing transportation requests. Therefore, attention is paid in the literature to
developing non-conflicting routes for AGVs. With a non-conflicting route, an
AGV arrives as early as possible at the destination without conflicting with other
AGVs.
In AGV routing with static algorithms the route from node i to node j is
determined in advance and is always used if a load has to be transported from i to
j. In this way, a simple assumption is to choose the route with the shortest distance
from i to j. However, these static algorithms are not able to adapt to changes in the
system and traffic conditions.
In dynamic routing, the routing decision is made based on real-time
information and, as a result, various routes between i and j can be chosen. Static
routing problems in AGV systems are related to vehicle routing problems (VRP)
studied in transportation literature. In the vehicle routing problem a set of n clients
with known demands need to be served by a fleet of m vehicles with limited
capacity. The vehicles are all housed at one depot. The route of each vehicle starts
and ends at this depot. m least costs (length) routes have to be planned such that
each customer is served exactly once and that the total demand of the customers
served by each vehicle does not exceed the capacity of each vehicle. The objective
is to minimize the total distance of all m routes under previously mentioned
conditions. This is an NP-hard problem to solve. The vehicle routing problem has
been studied extensively in literature. Bodin et al. (1983), Laporte (1992) and
Fisher (1995) provided an overview of literature in this area. A more recent paper
observing this problem is from Kelly and Xu (1999). They proposed a set
1.3 Routing Models 5
partitioning based heuristic. In a systematic way fragments of routes are combined

to obtain high-quality solutions. Vehicle scheduling problems can be seen as
routing problems with additional constraints concerning times at which certain
activities (e.g. delivery of a load) have to be executed. Vehicle activities have to
be sequenced both in time and space. An overview of methods to solve vehicle
scheduling problems has been given in Bodin et al. (1983).
Numerous studies on the vehicle routing problem with time windows have been
executed. Branch and bound methods (Kolen et al., 1987), insertion heuristics
(Solomon, 1987), extensions of vehicle routing problem heuristics (Solomon et al.,
1988), Lagrangian relaxation (Fisher et al., 1997; Kohl and Madsen, 1997),
constrained shortest path relaxation (Desrochers et al., 1992; Kohl et al., 1999)
and set covering formulations (Bramel and Simchi-Levi, 1997) can be used to find
solutions to the problem. Desrochers et al. (1988) provided an overview of
solution methods. More recently, Cordeau et al. (2002) presented a survey of
approximation and optimal approaches to solve the vehicle routing problem with
time windows.
1.3.1 Artificial Intelligent

Owing to the lack of robust design and control algorithms, most current
applications of AGV systems employ simple control methods despite the fact that
the system is far from efficient. Petri nets have evolved into a powerful tool for
modelling complex manufacturing systems. One of the advantages of the use of
Petri nets is that analysis, simulation and on-line control can all be done on the
same model once the model is built. By also using petri nets, conflict free routes
can be determined (see Zeng et al., 1991). The purpose of Hsieh and Lin (1991)
was to establish the research fundamentals in the field of the Petri-net modelling
of an AGVS. The main contribution of the chapter was to define basic traffic-
control nets which can be used directly to model an AGVS without too much
thinking. Some basic AGV Petri-net control elements were described and
illustrated in detail. Also, difficulties in the use of bidirectional flows were
discussed, and it has been determined how Petri nets can be used to solve these
problems.
The main problem for an AGV dispatching system is to assign vehicles to
transport demands which optimize some predetermined objectives of a
manufacturing shop. Singh and Tiwari (2004) presented a framework for an AGV
dispatching system based on an object oriented approach using the unified
modeling language (UML), and the development of a dispatching algorithm to
facilitate a human controller to dispatch efficiently a fleet of AGVs in response to
calls from any shop floor (or machine) operator. The main reason for their work
was to model an AGV dispatching system as well as to develop a dispatching
algorithm which can record details of the AGV position and movement and ensure
their allocation of new orders. The provision of both immediate and pre-booked
orders for an AGV was also incorporated in the proposed AGVs dispatching
system. The underlying AGV dispatching system and algorithm were capable of
dispatching a vehicle automatically to handle a call at the required time. In order
to overcome difficulties associated with tackling immediate orders, pre-booked
orders, and processing of information related to AGVs, a comprehensive
dispatching algorithm was developed which aims to minimize lateness, traveling
time and distance of empty vehicles in a simulated jobshop scenario.
Kizil et al. (2006) evaluated the effects of various dispatching rules on the
operation and performance of cellular manufacturing systems (CMS). When the
study of a CMS considers the automated material handling, it is crucial to reduce
the gridlock probability (i.e., the probability of an unsuccessful load transfer
attempt occurring in the interface point between the intercell and intracell
handling system). Preventing an unsuccessful load transfer is critical for the
operation of the entire system as a blockage between the AGV and the overloaded
cell results in a total system shutdown. The gridlock probability was influenced by
the dispatching rule used to schedule the load transfers in the system. Therefore, in
order to reduce this probability it was necessary to use a dispatching rule that will
decrease the number of waiting loads in the transfer spurs. The main objective of
the paper was to identify a dispatching rule that maintains the system operational
at all times. A group of dispatching rules, including the first come first served,
shortest imminent operation, longest imminent operation, most remaining
operations, shortest processing time, shortest remaining process time, and a newly
developed rule proposed by the authors, loads with the minimum number of
processing first, were tested and evaluated with respect to whether the capacity of
the transfer spurs of the cells was exceeded. The paper presented a simulation
model of a cellular manufacturing system, which was used to further explore the
effects of the dispatching rules on the system performance.
AGVs are the most flexible means to transport materials among workstations of
a flexible manufacturing system. Complex issues associated with the design of
AGV control of these systems are conflict-free shortest path, minimum time
motion planning and deadlock avoidance. Srivastava et al. (2008) presented an
intelligent agent-based framework to overcome the inefficacies associated with the
aforementioned issues. Proposed approach described the operational control of
AGVs by integrating different activities such as path generation, journey time
enumeration, collision and deadlock identification, waiting node location and its
time estimation, and decision making on the selection of the conflict-free shortest
feasible path. It represented efficient algorithms and rules associated with each
agent for finding the conflict-free minimum time motion planning of AGVs,
which were needed to navigate unidirectional and bidirectional flow path network.
A collaborative architecture of AGV agent and its different modules were also
presented. Three complex experimental scenarios were simulated to test the
robustness of the proposed approach. It was shown that the proposed agent-based
controller was capable of generating optimal, collision- and deadlock-free path
with less computational efforts.
The objective of Aized (2009) was to model and maximize performance of an
integrated AGVS, which is embedded in a pull type multi-product, multi-stage and
multi-line FMS. The author examined the impact of guide-path flexibility on

system performance through the development of three different guide-path
configurations which range from dedicated to flexible relationships between
AGVs and machine/assembly station resources. The system was modelled using
coloured Petri net method (CPN) and the simulation results lead to identify the
resource redundancy which can be rectified to achieve lower overall cost of the
system through the development of flexible guide-path configurations. The study
was extended to seek global near-optimal conditions for each guide-path
configuration using response surface method, which yields improvements in
system throughput and cycle time along with a decrease in the numbers of AGVs.
Material handling in manufacturing systems is becoming easier as the
automated machine technology is improved. Nowadays, most of the research aims
at increasing the flexibility and improving the performance of the AGV. Yahyaei
et al. (2010) designed and made AGV in the Industrial Control Laboratory in
Royce Lab at the University of Manchester Institute of Science and Technology.
For controlling the navigation of the AGV, a newly developed controller
integrated fuzzy logic with programmable logic controller was used. By using
integrated fuzzy logic controller with programmable logic controller (IFLPLC),
the flexibility of AGV was increased and they achieved great advantages. Since
that AGV used programmable logic controller and fuzzy logic controllers
together, it proved usefulness for factories which implement FMS. Online
maintenance and sending the commands to other machines from AGV and so on
were the advantages that can be used in FMS.
Fazlollahtabar and Mahdavi-Amiri (2012) proposed an approach for finding an
optimal path in a flexible jobshop manufacturing system considering two criteria
of time and cost. A network was configured in which the nodes are considered to
be the shops with arcs representing the paths among the shops. An automated
guided vehicle functioned as a material handling device through the
manufacturing network. The expert system for cost estimation was based on fuzzy
rule backpropagation network to configure the rules for estimating the cost under
uncertainty. A multiple linear regression model was applied to analyze the rules
and find the effective rules for cost estimation. The objective was to find a path
minimizing an aggregate weighted unscaled time and cost criteria. A fuzzy
dynamic programming approach was presented for computing a shortest path in
the network. Then, a comprehensive economic and reliability analysis was worked
out on the obtained paths to find the optimal producer’s behavior.
Singh and Tiwari (2002) presented an intelligent agent framework to find a
conflict-free shortest-time path for an AGV travelling in a bi- or unidirectional
network.
Fazlollahtabar and Mahdavi-Amiri (2013) proposed an approach for finding an
optimal path in a flexible jobshop manufacturing system considering two criteria
of time and cost. With rise in demands, advancement in technology and increase
in production capacity, the need for more shops persists. Therefore, a flexible
jobshop system has more than one shop with the same duty. The difference among
shops with the same duty is in their machines with various specifications.
A network was configured in which the nodes were considered to be the shops
with arcs representing the paths among the shops. An AGV functioned as a
material handling device through the manufacturing network. To account for
uncertainty, the authors considered time to be a triangular fuzzy number and
applied an expert system to infer cost. The objective was to find a path minimizing
both the time and cost criteria, aggregately. Since time and cost have different
scales, a normalization procedure was proposed to remove the scales. The model
being biobjective, the analytical hierarchy process weighing method was applied
to construct a single objective. Finally, a dynamic programming approach was
presented for computing a shortest path in the network. The efficiency of the
proposed approach was illustrated by a numerical example.
1.3.2 Simulation
Software aids for simulation are very important to practitioners of simulation. The
widespread availability of inexpensive computing power now allows computer
assistance in each stage of simulation activities such as input data analysis,
modelling, programming, output analysis and so on. Therefore Ashayeri and
Gelders (1987) described an interactive microcomputer GPSS simulation program
generator for automated material handling systems. The program was written in
Pascal and consisted of several modules to capture data, build the model, and
generate the corresponding GPSS simulation program for automated guided
vehicle systems as well as surge systems. The application of the program to a real
life project was used to highlight practical advantages of the proposed approach.
Automated Guidance Vehicles' guidance techniques make use, in most
vehicles, of a painted strip or a cable buried in the floor. Free ranging vehicles are
now also available in limited types but at prohibitive cost. Katz and Bright (1992)
presented a type of guidance technique whereby the vehicle follows a path created
by the light emitted from suspended fluorescent lights. The method has been
tested on a multi-directional vehicle. The results were encouraging when
compared with other available guidance techniques. The possible integration of
this guidance method with a navigation method could make AGVs more attractive
to most of the manufacturing environments. The design, implementation and
experimental results were outlined and described.
A simulation-based cost model was presented by Kasilingam and Gobal (1996)
for determining the number of AGVs needed to meet the material handling
requirements in a manufacturing system. The estimation of the number of vehicles
was based on the sum of the idle-time costs of vehicles and machines, and the cost
of waiting time of parts. While an increase in the number of vehicles reduces the
waiting time of parts and the idle time of machines, it increases the idle time of
vehicles. The application was illustrated using a hypothetical manufacturing
system.
A TAGV system is obtained by partitioning all workstations into multiple

zones and assigning a single vehicle to each zone. Kim et al. (2003) proposed an
analytical model to design a tandem AGVS with multi-load AGVs. Using
simulations, the performance of the proposed model was shown by comparing it
with a conventional multi-load AGVS.
Kim and Jae (2003) presented an object-oriented simulation modelling
environment, AgvTalk, to provide flexible modelling capabilities for the
simulation of many alternative AGV systems. The hierarchical features and
modularity of AgvTalk created possibilities for the extension and reuse of
simulation object components. The detailed behaviour of each object in the AGV
system can also be modelled easily and exactly in AgvTalk because there were no
limiting modeling constructs. The modelling capabilities of AgvTalk were
demonstrated by designing and simulating a conceptually different configuration
AGV system, known as the tandem configuration. For the tandem and
conventional AGV systems, the characteristics and design methodology in
AgvTalk were described. Simulations between the two systems were compared
using AgvTalk in a job shop environment.
Um et al. (2009) presented the simulation design and analysis of an FMS with
an AGV system. To maximize the operating performance of FMS with AGVs,
many parameters must be considered, including the number, velocity, and
dispatching rule of AGV, part types, scheduling, and buffer sizes. Of the various
critical factors, they considered the following three: (1) minimizing the
congestion; (2) minimizing the vehicle utilization; and (3) maximizing the
throughput. The authors considered the systematic analysis methods that
combined a simulation-based analytic and optimization technique that was Multi-
Objective Non-Linear Programming (MONLP) and Evolution Strategy (ES).
MONLP determined the design parameters of the system through multi-factorial
and regression analyses. ES was used to verify each parameter for simulation-
based optimization. A validation test for the two methods was conducted. This
method-based approach towards design yielded the correct experimental results,
ensures confidence in the specification of design parameters and supports a robust
framework for analysis.
Kim et al. (2009) proposed a simple blocking prevention method for a path-
based automated material handling system (AMHS) such as a semiconductor
fabrication line. In their research, blocking means the situation in which a vehicle
stands and waits without doing anything because the vehicle in front of it is in the
process of loading or unloading. Since a typical bay type path-based AMHS has a
single path in each bay and no sidetracks, the blocking issue is inevitable in a large
complex system with many vehicles. The proposed method was based on the
swapping of load assignments between retrieval vehicles on the same path. The
simulation study on an example semiconductor line showed that the proposed
method improves the AMHS productivity under various vehicle dispatching rules.
Guan and Dai (2009) developed a flexible, efficient, and deadlock-free
dispatching method for automated guided vehicle systems. For this purpose, a
deadlock-free multi-attribute dispatching method with dynamically adjustable
weights (AWMA) was proposed. Traveling distance, input, and output buffer
statuses were selected as dispatching attributes according to the efficiency and
deadlock avoidance requirement. The weight for each attribute was dynamically
adjusted according to the processing load and transportation load of the system.
To ensure the system to be deadlock-free, a deadlock avoidance policy based on
remaining capacity concept was introduced. It worked by temporarily forbidding
critical tasks according to the system state, which will otherwise cause system
deadlock. The AWMA method was formed by integrating the deadlock avoidance
policy into the multi-attribute dispatching procedure. To validate the effectiveness
of the proposed method, several simulation experiments were carried out to
compare three commonly used dispatching methods with the proposed one under
different system settings. The simulation results indicated that the deadlock
avoidance policy can guarantee the system to be deadlock-free and that the
proposed method was efficient.
In the majority of small and medium sized enterprises, the direct costs of
material handling cannot be clearly measured. There are several reasons for this,
including the large number of product types, complexity of their production cycle,
and continuous change in markets. Therefore, production managers require
flexible tools to create a suitable material handling system model which explicitly
and rapidly calculates the indices required as these are traditionally neglected or
laboriously approximated, (i.e., time and cost in material flow inside the factory,
storage area requirements, and MH utilization percentage). Gamberi et al. (2009)
proposed an integrated approach to analyzing and controlling material handling
operations in an industrial manufacturing plant from a “full quantitative” point of
view. The model presented united quite different fields of research into a unique
methodology. The material handling model rapidly and automatically provided
production managers with extensive and significant information. As a result,
integrated layout flow analysis interrelated systematic layout planning with
operational research algorithms and visual interactive simulation, using a complete
software platform to implement them. This integrated layout flow analysis
approach focused on determining the space requirement for manufacturing
department buffers, the transportation system requirements, the performance
indices, and the time and cost of material flows spent in the layout and in MH
traffic jams.
Flexible material handling systems (FMHS) have been widely used to enhance
productivity involved with product proliferation, and thus far, only fixed-track
material handling systems such as Eton systems in the apparel industry are
commonly used. Dai et al. (2009) explored the potential advantages of a FMHS
using free-ranging automated-guided vehicles with a local positioning system for
the apparel industry. First, the free-ranging FMHS (FRMHS) for the apparel
industry has been designed. Then, through Monte Carlo simulation and analytical
models, the performance in terms of manufacturing system effectiveness,
workstation utilization, and the total transportation distance of the FRMHS were
compared with those of the fixed-track system. Based on their analysis, the current
proposed FRMHS can have significant advantages over the fixed-track system.
Kuttolamadom et al. (2010) dealt with the path tracking and stability of motion
of automated guided vehicle systems and wheeled mobile robots. A two degree-
of-freedom dynamic model was developed to represent the plane motion of the
vehicle. This model along with the instantaneous posture errors (position and
orientation errors) of the vehicle were used to formulate their path-tracking
problem in state space format. Implementation of stability criterion and
application of the actual physical limits of the pertinent parameters of the system
were the strategies used to design the controller. It was shown that implementation
of a control strategy based on the vehicles’ position and orientation errors gives
satisfactory results in vehicles’ path tracking. The feasibility of the approach and
the performance of the controlled system were demonstrated by using a prototype
vehicle.
Joseph and Sridharan (2011) focused on a simulation-based experimental study
of the interaction among routing flexibility, sequencing flexibility and part
sequencing rules in a typical FMS. Two scenarios were considered for
experimentation. Three routing flexibility levels, five sequencing flexibility levels
and four scheduling rules for part sequencing decision were considered for
detailed investigation. The performance of the FMS was evaluated using various
measures related to flow time and tardiness of parts. The simulation results were
subjected to statistical analysis. The analysis of results reveals that deterioration in
system performance can be minimized substantially by incorporating either
routing flexibility or sequencing flexibility or both. However, the benefits of either
of these flexibilities diminish at higher flexibility levels. Part sequencing rules
such as earliest due date and earliest operation due date provide better
performance for all the measures at higher flexibility levels.
Routing flexibility is a major contributor of the flexibility of an FMS. Joseph
and Sridharan (2011) focused on the evaluation of the routing flexibility of an
FMS with the dynamic arrival of part types for processing in the system. A typical
FMS configuration was chosen for detailed study and analysis. The system was set
at five different levels of routing flexibility. Operations of part types can be
processed on alternative machines depending upon the level of routing flexibility
present in the system. Two cases have been considered with respect to the
processing times of operations on alternative machines. A discrete-event
simulation model has been developed to describe the operation of the chosen
FMS. The performance of the system under various levels of routing flexibility
was analyzed using measures such as mean flow time, mean tardiness, percentage
of tardy parts, mean utilization of machines, mean utilization of automatic-guided
vehicles, and mean queue length at machines. The routing flexibility for producing
individual part types has been evaluated in terms of measures such as routing
efficiency, routing versatility, routing variety and routing flexibility. The routing
flexibility of the system has been evaluated using these measures. The flexibility
levels were ranked based on the routing flexibility measure for the system. The
ranking thus obtained has been validated with that derived using fuzzy logic
approach.
Nowadays, so as to adapt to the global market, where competition is getting

tougher, firms producing through the modern production approach need to bring
not the only performance of the system designed both during the research and
development phase and the production phase but also the performance of the
product to be developed as well as the process to be improved to the highest level.
The Taguchi method is an experimental design technique seeking to minimize the
effect of uncontrollable factors, using orthogonal arrays. It can also be designed as
a set of plans showing the way data are collected through experiments.
Experiments are carried out using factors defined at different levels and a solution
model generated in ARENA 3.0 program using SIMAN, which is a simulation
language. Many experimental investigations reveal that the speed and capacity of
automated-guided vehicle, the capacities of local depots, and the mean time
between shipping from the main depot are the major influential parameters that
affect the performance criteria of the storage system. For the evaluation of
experiment results and effects of related factors, variance analysis and signal/noise
ratio were used and the experiments were carried out in MINITAB15 according to
Taguchi L16 scheme. The purpose of Subulan and Cakmakci (2012) was to prove
that experimental design was an utilizable method not only for product
development and process improvement but it can also be used effectively in the
design of material handling–transfer systems and performance optimization of
automation technologies, which were to be integrated to the firms.
Seifert et al. (1998) introduced a dynamic vehicle routing strategy based on
hierarchical simulation. At each time a route decision has to be made, simulations
were performed to evaluate a set of possible routes. The route with the smallest
estimated travel time was chosen.
Except for finding conflict free routes, attention should be paid to the presence
of interruptions in the system. Interruptions might occur due to, for example,
vehicle breakdowns, objects on AGV paths and manual intervention. As a result of
interruptions, AGVs may be blocked and routes cannot be finished. Therefore, if
an AGV encounters an interruption it has to be rerouted in such a way that no
conflicts with other AGVs occur. Narasimhan et al. (1999) used simulation to
analyze rerouting of AGVs.
1.4 Conclusions
Summarizing, in the context of manufacturing areas, static and dynamic

algorithms have been developed to solve the routing of vehicles. Network models,
queuing networks, simulation and intelligent routing techniques are used to route
AGVs conflict free through the network. The routing of AGVs through
distribution, transshipment and transportation systems is hardly studied.
References 13
References
Aized, T.: Modelling and performance maximization of an integrated automated guided
vehicle system using coloured Petri net and response surface methods. Computers &
Industrial Engineering 57, 822–831 (2009)
Ashayeri, J., Gelders, L.F.: Interactive GPSS-PC Program Generator for Automated
Material Handling Systems. International Journal of Advanced Manufacturing
Technology 2(4), 63–77 (1987)
Bodin, L.D., Golden, B.L., Assad, A.A., Ball, M.O.: Routing and scheduling of vehicles
and crews: The state of the art. Computers and Operations Research 10(2), 63–211
(1983)
Bramel, J., Simchi-Levi, D.: On the effectiveness of set covering formulations for the
vehicle routing problem with time windows. Operations Research 45(2), 295–301
(1997)
Dai, J.B., Lee, N.K.S., Cheung, W.S.: Performance analysis of flexible material handling
systems for the apparel industry. International Journal of Advanced Manufacturing
Technology 44, 1219–1229 (2009)
Desrochers, M., Desrosiers, J., Solomon, M.: A new optimization algorithm for the vehicle
routing problem with time windows. Operations Research 40(2), 342–354 (1992)
Desrochers, M., Lenstra, J.K., Savelsbergh, M.W.P., Soumis, F.: Vehicle routing with time
windows: Optimization and approximation. In: Golden, B.L., Assad, A.A. (eds.)
Vehicle Routing: Methods and Studies. Studies in Management Science and Systems,
pp. 65–84 (1988)
Fazlollahtabar, H., Mahdavi-Amiri, N.: Producer’s behavior analysis in an uncertain
bicriteria AGV-based flexible jobshop manufacturing system with expert system. The
International Journal of Advanced Manufacturing Technology (2012),
doi:10.1007/s00170-012-4283-0.
Fazlollahtabar, H., Mahdavi-Amiri, N.: An optimal path in a bi-criteria AGV-based flexible
jobshop manufacturing system having uncertain parameters. International Journal of
Industrial and Systems Engineering 13(1), 27–55 (2013)
Fazlollahtabar, H., Es’haghzadeh, A., Hajmohammadi, H., Taheri-Ahangar, A.: A Monte
Carlo simulation to estimate TAGV production time in a stochastic flexible automated
manufacturing system: a case study. International Journal of Industrial and Systems
Engineering 12(3), 243–258 (2012)
Fazlollahtabar, H., Rezaie, B., Kalantari, H.: Mathematical programming approach to
optimize material flow in an AGV-based flexible jobshop manufacturing system with
performance analysis. The International Journal of Advanced Manufacturing
Technology 51(9-12), 1149–1158 (2010)
Fisher, M.: Vehicle routing. In: Ball, M.O., Magnanti, C.L., Monma, C.L., Nemhauser,
G.L. (eds.) Network Routing, pp. 1–33. Elsevier, Amsterdam (1995)
Fisher, M.L., Jörnsten, K.O., Madsen, O.B.G.: Vehicle routing with time windows: Two
optimization algorithms. Operations Research 45(3), 488–492 (1997)
Gamberi, M., Manzini, R., Regattieri, A.: An new approach for the automatic analysis and
control of material handling systems: integrated layout flow analysis (ILFA).
International Journal of Advanced Manufacturing Technology 41, 156–167 (2009)
Gans, N., Van Ryzin, G.: Dynamic vehicle dispatching: Optimal heavy traffic performance
and practical insights. Operations Research 47(5), 675–692 (1999)
Gaur, D.R., Gupta, A., Krishnamurti, R.: A 5/3-approximation algorithm for scheduling
vehicles on a path with release and handling times. Information Processing Letters 86,
87–91 (2003)
Guan, X., Dai, X.: Deadlock-free multi-attribute dispatching method for AGV systems.
Haefner, L.E., Bieschke, M.S.: ITS opportunities in port operations. In: Transportation
Conference Proceedings, pp. 131–134 (1998)
Hsieh, S., Lin, K.: Building AGV Traffic-Control Models With Place-Transition Nets.
Jerald, J., Asokan, P., Prabaharan, G., Saravanan, R.: Scheduling optimization of flexible
manufacturing systems using particle swarm optimization algorithm. International
Journal of Advanced Manufacturing Technology 25, 964–971 (2005)
Joseph, O.A., Sridharan, R.: Effects of routing flexibility, sequencing flexibility and
scheduling decision rules on the performance of a flexible manufacturing system.
Joseph, O.A., Sridharan, R.: Evaluation of routing flexibility of a flexible manufacturing
system using simulation modelling and analysis. International Journal of Advanced
Manufacturing Technology 56, 273–289 (2011)
Kasilingam, R.G., Gobal, S.L.: Vehicle Requirements Model for Automated Guided
Vehicle Systems. International Journal of Advanced Manufacturing Technology 12,
276–279 (1996)
Katz, Z., Bright, G.: A Guidance Technique for an Automated Guided Vehicle.
Kelly, J.P., Xu, J.: A set-partitioning-based heuristic for the vehicle routing problem.
Journal on Computing 11(2), 161–172 (1999)
Kim, B., Shin, J., Chae, J.: Simple blocking prevention for bay type path-based automated
material handling systems. International Journal of Advanced Manufacturing
Technology 44, 809–816 (2009)
Kim, C.W., Tanchoco, J.M.A.: Conflict-free shortesttime bidirectional AGV routing.
International Journal of Production Research 29(12), 2377–2391 (1991)
Kim, K., Jae, M.: An object-oriented simulation and extension for tandem AGV systems.
Kim, K.S., Chung, B.D., Jae, M.: A design for a tandem AGVS with multi-load AGVs.
Kizil, M., Ozbayrak, M., Papadopoulou, T.C.: Evaluation of dispatching rules for cellular
manufacturing. International Journal of Advanced Manufacturing Technology 28, 985–
992 (2006)
Kohl, N., Madsen, O.B.G.: An optimization algorithm for the vehicle routing problem with
time windows based on Lagrangian relaxation. Operations Research 45(3), 395–406
(1997)
Kohl, N., Desrosiers, J., Madsen, O.B.G., Solomon, M.M., Soumis, F.: 2-path cuts for the
vehicle routing problem with time windows. Transportation Science 33(1), 101–116
(1999)
Kolen, A.W.J., Rinnooy Kan, A.H.G., Trienekens, H.W.J.M.: Vehicle routing with time
windows. Operations Research 35(2), 266–273 (1987)
Kuttolamadom, M., Mehrabi, M.G., Weaver, J.: Design of a stable controller for accurate
path tracking of automated guided vehicles systems. International Journal of Advanced
Laporte, G.: The vehicle routing problem: An overview of exact and approximate
algorithms. European Journal of Operational Research 59, 345–358 (1992)
References 15
Maughan, F.G., Lewis, H.J.: AGV controlled FMS. International Journal of Production
Research 38(17), 4445–4453 (2000)
Muller, T.: Automated Guided Vehicles. IFS (Publications) Ltd./Springer-Verlag,
UK/Berlin (1983)
Narasimhan, R., Batta, R., Karwan, M.H.: Routing automated guided vehicles in the
presence of interruptions. International Journal of Production Research 37(3), 653–681
(1999)
Sabuncuoglu, I.: A study of scheduling rules of flexible manufacturing systems: A
simulation approach. International Journal of Production Research 36(2), 527–546
(1998)
Seifert, R.W., Kay, M.G., Wilson, J.R.: Evaluation of AGV routing strategies using
hierarchical simulation. International Journal of Production Research 36(7), 1961–1976
(1998)
Singh, S.P., Tiwari, M.K.: Object oriented modelling and development of a dispatching
algorithm for automated guided vehicles. International Journal of Advanced
Singh, S.P., Tiwari, M.K.: Intelligent agent framework to determine the optimal conflict-
free path for an automated guided vehicles system. International Journal of Production
Research 40(16), 4195–4223 (2002)
Solomon, M.M.: Algorithms for the vehicle routing and scheduling problems with time
window constraints. Operations Research 35(2), 254–265 (1987)
Solomon, M.M., Desrosiers, J.: Time window constrained routing and scheduling
problems. Transportation Science 22(1), 1–13 (1988)
Solomon, M.M., Baker, E.K., Schaffer, J.R.: Vehicle routing and scheduling problems with
time window constraints: Efficient implementations of solution improvement
procedures. In: Golden, B.L., Assad, A.A. (eds.) Vehicle Routing: Methods and
Studies. Studies in Management Science and Systems, pp. 85–104 (1988)
Srivastava, S.C., Choudhary, A.K., Kumar, S., Tiwari, M.K.: Development of an intelligent
agent-based AGV controller for a flexible manufacturing system. International Journal
of Advanced Manufacturing Technology 36, 780–797 (2008)
Subulan, K., Cakmakci, M.: A feasibility study using simulation-based optimization and
Taguchi experimental design method for material handling—transfer system in the
automobile industry. International Journal of Advanced Manufacturing Technology 59,
433–444 (2012)
Van der Heijden, M., Ebben, M., Gademann, N., Van Harten, A.: Scheduling vehicles in
automated transportation systems. OR Spectrum 24, 31–58 (2002b)
Van der Heijden, M.C., Van Harten, A., Ebben, M.J.R., Saanen, Y.A., Valentin, E.C.,
Verbraeck, A.: Using simulation to design an automated underground system for
transporting freight around Schiphol airport. Interfaces 32(4), 1–19 (2002a)
Vis, I.F.A., Harika, I.: Comparison of vehicle types at an automated container terminal. OR
Spectrum 26, 117–143 (2004)
Yahyaei, M., Jam, J.E., Hosnavi, R.: Controlling the navigation of automatic guided vehicle
(AGV) using integrated fuzzy logic controller with programmable logic controller
(IFLPLC). International Journal of Advanced Manufacturing Technology 47, 795–807
(2010)
Yang, C.H., Choi, Y.S., Ha, T.Y.: Simulation-based performance evaluation of transport
vehicles at automated container terminals. OR Spectrum 26, 149–170 (2004)
Zeng, L., Wang, H.P., Jin, S.: Conflict detection of automated guided vehicles: A petri net
approach. International Journal of Production Research 29(5), 865–879 (1991)
Chapter 2
Analytical Material Flow Model for AGV
System
2.1 Summary
Manufacturing automation has become increasingly important as the need to

remain productive increases. In manufacturing of a product, many processes may
be needed. For example, semiconductor manufacturing may include over 400
processing steps involving more than 100 different tools. Furthermore, the process
route can include a high level of reentrance in which the same tool or tool types
are used. An important aspect in manufacturing automation is material handling.
To facilitate material handling, automated transport systems are employed.
A special attention has been given during the last few years to the problem of
production control in manufacturing systems (Akella et al., 1990; Gershwin,
1989). Flow models are often used in the intermediate level to represent the
inventory balance equation of the production system. The aim is to compute an
average production rate for each product type in order to meet the demand with
minimum surplus or backlogging costs. During the optimization procedure
capacity changes due to machine failures must be taken into account. In
(Gershwin, 1989 ), changes in the system’s capacity due to machine failure are
described as a function of the state of the system and since machine breakdowns
cannot be predicted, capacity is considered as a stochastic set. This means that for
the dynamic system representing the flow of parts, a sudden change in the
system’s state due to machine failure is transformed to a change (or a jump) for
the capacity constraint.
Conventional automated transport systems are used in, for example, a
semiconductor manufacturing plant typically comprises a plurality of bays. Each
bay includes a plurality of tools which are used to process wafers. Transport
vehicles, such as overhead transport (OHT) are provided for automatic transfer of
wafers contained in a carrier. A stocker which temporarily stores carriers is
provided for each bay. A transport system controller controls the movement of
carriers within a bay or between bays. In recent years, the automated material
handling system has rapidly developed as an efficient manufacturing system.
The AGV (automated guided vehicle) system plays an especially significant role

18 2 Analytical Material Flow Model for AGV System
and has become more widely used in modern manufacturing environments

due to its flexibility and precision. With an AGV system, we can easily
respond to changes in production volume, product mix, product routing and so on
(Ho and Liao, 2009).
However, due to the continuously increasing size and complexity of the modern
manufacturing system, controlling this system has become more difficult.
Consequently, various types of research have been conducted to solve this
problem. Bozer and Srinivasan (1992) defined the conventional AGV system and
proposed tandem configuration based on the ‘‘divide and conquer’’ principle.
Multiple vehicles can exist in a traditional AGV system and each vehicle can pick-
up, deliver and drop off a load at any workstation. A tandem AGV system is
obtained by partitioning all workstations into multiple zones, assuming a single
vehicle to each zone. As a result, any potential blocking, congestion or deadlock
does not occur. And the performance of the tandem AGV system has been
demonstrated (Bozer and Srinivassan, 1992; Laporte et al., 2006; Asef-Vaziri
et al., 2001; Kaspi et al., 2002).
The emergence of high performance automated manufacturing systems (AMSs)
has lead to the need for methods of modeling these types of system in order to
maximize throughput, flexibility and competitiveness. AMSs belong to the
domain of discrete event dynamic systems (DEDS) in which the evolution of the
system depends on the complex interactions of various discrete events such as
the arrival of raw materials, departure of finished goods, failure of equipment etc.
The state of DEDS changes only at these discrete points in time. Over the last
decade several models have been presented to describe DEDS and these can be
grouped into two distinct area (Rajagopalan et al., 2004).
Qualitative models are concerned with the logical aspects of system evolution
such as controllability, stability and the existence of deadlocks in system
operation, etc. This category also includes Petri Nets, extended state machines and
finitely recursive processes (Castillo et al., 2001).
Quantitative models are concerned with the quantitative system performance in
terms of throughput and lead time. This category also includes discrete event
simulation, min-max algebra, Markov Chains, stochastic Petri nets, queues, and
queuing networks (Berman et al., 2009) . Quantitative models are a general term
including performance modeling which is the area of interest to this article. Within
the life cycle of an AMS various decisions are made concerning implementation,
design and operation of the system. Typical decisions at the planning stage include
number and type of machines, number of material handling devices, number of
buffers, size of pallet pool and number of fixtures, best possible layout, tool
storage capacity, evaluate candidate AMS configurations, part type selection,
machine grouping, batching and balancing decisions, and scheduling policies
(Huang et al., 2009)
During the operational phase of an AMS, performance modeling can be used to
assist decisions about how to react in the event of a breakdown, removal or
addition of resources and parts, optimal scheduling in the event of machine failure
2.2 Statement of the Problem 19
or sudden changes in the product or its demand and in the avoidance of unstable
situations such as deadlocks (Demange et al., 2009).
Performance modeling is also used in the design stage of the system. It is used
in decisions such as whether to use central versus local storage, push production
versus pull production, shared versus distributed resources, the effect of
flexibility, etc. Performance predictions obtained using faithful models, can be
used to convince customers or investors and also give the designer another
perspective on the design enabling better designs (Aized, 2009).
The performance of an AMS can be measured by a set of generic measures.
These are manufacturing lead time, work in progress, throughput, machine
utilization, capacity, flexibility, performance, and quality (Gen et al., 2009). Using
performance measuring these values can be evaluated and used to compare AMS
performances.
Performance modeling has become a very important part of automated
manufacturing system design and is equally important for maintaining the system
at its peak of ability. The manufacturing methods in use by companies has
changed dramatically in recent years with the use of advanced robotics and
computer control to optimize production, this has lead to reduced prices and
higher quality of product. The production lines can only get better with more
modeling and investment and this is best achieved with the use of performance
modeling.
2.2 Statement of the Problem
Here, we consider a jobshop layout which employs an AGV for material handling.
The AGV carries raw material, semi-produced and final products in batch sizes.
Because of the increase in demands, advance in technology, and rise in the
production capacity more shops than the existing shops are required. The new
shops are associated with higher technology machines. Therefore, more than one
shop with the same performance is evolved. The difference among shops with the
same performance is machines with various specifications that effect the
production time/ cost and productivity. As a result, the system could be a flexible
jobshop model where multi shops of the same performance exist and each
operation is possible to be processed on any type of machine. The sequences of
jobs are specified and the jobs are independent.
To evaluate the performance of the proposed AMS, we assess the material flow
between any two shops of different types. In the proposed model the aim is to
optimize the material flow, i.e. finding a set of shops which minimize the material
flow throughout the system. Here, flow is considered as the distance which the
AGV moves to satisfy the production plan and demand. The proposed model is
presented schematically in Figure 1.
Fig. 1 The Schematic presentation of the Proposed Model
2.3 Mathematical Model
In this section the mathematical model of the proposed flexible jobshop problem is
represented. The indices, parameters, and decision variables are as follows:
Indices:
m Index for shops, m=1,2, … , M.

k Index for shops, k=1,2, … , K.
n Index for shop type mth , n=1,2, … , N.
h Index for shop type kth, h=1,2, … , H.
i Index for products, i=1,2, … , I.
p Index for job position, p=1,2, … , P.
2.3 Mathematical Model 21
Parameters:
Cipmn Completion time of product ith in position pth in shop nth of type
mth
P.Timn Processing time of shop nth of type mth for product ith
T.Timn kh Transferring time from shop nth of type mth to shop hth of type
kthfor product ith
VAGV Velocity of AGV
f imnkh Flow (distance) for product ith between shop nth of type mth and
shop hth of type kth
Wimn Waiting time for product ith in shop nth of type mth
C.Ti Cycle Time for product ith

T Total working time in each day
⎧ 0 o.w
z ipm = ⎨ th th th
⎩1 if shop m is allocated for product i in position p
Decision variable:
⎧0 o.w
z ipmn = ⎨ th th th th
⎩1 if shop n of type m is chosen for product i in position p
Objective Function:
H K N M P I
Min ∑∑∑∑∑∑ z
h =1 k =1 n =1 m =1 p =1 i =1
ipm .zi ( p −1) kh .T .Timn kh (1)
m≠k
S.T :
( )
P N M H K N M P
Cipmn = ∑∑∑ z ipmn ( P.Timn + Wimn ) + ∑∑∑∑∑ z ipmn .z i ( p −1) mn .T .Timn k h , ∀i, p, m, n, (2)
p =1 n =1 m =1 h =1 k =1 n =1 m =1 p =1
Cipmn − Cip −1kh ≥ P.Timn + Wimn + T .Timn kh , ∀i, p, m, n, k , h, (3)
∑z
n =1
ipmn = z ipm , ∀i, p, m, (4)
z ipmn .Cipmn ≤ C.Ti , ∀i, p, m, n, (5)
Cipmn ≤ T , (6)
f imn kh
T .T mn kh = , ∀i,k,h,m,n, (7)
V AGV
z ipmn ∈ {0,1} (8)
Equation (1), is the objective function of the proposed problem which

minimizes the material flow. The output of the objective function is the types of
the shops which minimize the total material flow. Equation (2), indicates the
computation of completion time for each product. Equation (3), certifies that the
differences between completion time of product ith in position pth in shop nth of
type mth is larger than or equal to the addition of the processing time and waiting
time of shop nth of type mth and the transferring time between shop nth of type mth
and shop hth of type kth. Equation (4), warranties that if any shop in any position
for any product is allocated then the corresponding shop type is also chosen.
Equation (5), guarantees that if a shop is chosen then the corresponding
completion time for each product is lower than or equal to the cycle time.
Equation (6), certifies that the completion time for any product in any position in
any shop is lower than or equal to the total working time in each day. Equation
(7), indicates the transferring time between two shops is related directly to the
flow of each product. Equation (8), presents the values of the decision variables.
After finding the shops which optimize the material flow, now we analyze the
performance of the system considering these shops.
2.3.1 Regression Model for Performance Analysis

Here, we consider the output of the mathematical program in last section and
analyze the performance of the obtained shops to satisfy the demands.We want to
estimate the coefficients which imply the capability of the obtained shops to
satisfy the demand of the manufacturing system. We should investigate how much
a shop is significant on the demand of the system. One way to survey the effect of
independent variables on dependent variable is multiple linear regression model.
Assume that Dt is the demand in time t and Xj is the maximum production rate of
shop jth. Therefore, we consider the following equation,
Dt = β 0 + β1 X 1 + ... + β j X j + ε j , t = 1,...,T ; j = 1,..., J . (9)

where β0 is the intercept, β j s are the coefficients for Xjs and ε is the error term.
The aim is first determining the value of the coefficients to see whether they are
lower than 1 or not, and second identifying the β j s which are not important on
demand (Dt). For the first objective, if the value of any coefficient is higher than 1,
then it indicates that the corresponding shop can not satisfy the allocated demands
with the current working time and lead to have extra working time. For the second
objective, we apply two-sided hypothesis testing as follows;
H 0 : β j = 0;
(10)
H 1 : β j ≠ 0.
The test of the null hypothesis H0 against the two sided alternative proceeds as
in the three steps. The first is to compute the standard error of β j , SE ( β j ) . The
standard error of βj is an estimator of σβ j
, the standard deviation of the
sampling distribution of βj.

m
1
× ∑ (r j − r ) 2 ε 2j
1 m − 2 j =1
σ β2 = × 2
, (11)
m ⎡1 m ⎤
j
⎢ ∑ (r j − r ) ⎥
2
⎣ m j =1 ⎦
SE ( β j ) = σ β2 j . (12)
Although the formula for σ β2 j

is complicated, in applications the standard error
is computed by regression software. The second step is to compute the t-statistic,
βj −0
t= (13)
SE ( β j )
The third step is to compute the p-value, the probability of observing a value of
βj at least as different from zero as the estimate actually computed ( β j ),
act
assuming that the null hypothesis is correct. Stated mathematically,

⎡ βj −0 β act −0⎤
[
p − value = PrH 0 β j − 0 > β act
j ]
− 0 ⇒ PrH 0 ⎢
β
>
j
β
,
⎥ ⇒ PrH 0 ( t > t act ) (14)
⎢⎣ SE ( j ) SE ( j) ⎥
⎦
where PrH 0 denotes the probability computed under the null hypothesis, the second
equality follows by dividing by SE ( β j ) , and t act is the value of the t-statistic

actually computed. Because β j is approximately normally distributed in large
samples, under the null hypothesis the t-statistic is approximately distributed as a
standard normal random variable, so in large samples,
p − value = Pr( Z > t act ) = 2Φ (− t act ) , (15)
A small value of the p-value, say less than 5%, provides evidence against the
null hypothesis in the sense that the chance of obtaining a value of βj by pure
random variation from one sample to the next is less than 5% if, in fact, the null
hypothesis is correct. If so, the null hypothesis is rejected at the 5% significance
level. Simply we can reject the null hypothesis at the 5% significance level if,
t act > 1.96 . Therefore, if the hypothesis is accepted then we find that the
corresponding coefficient is not important and has no effect on the demand.
After describing the aspects of the proposed model for this type of flexible
jobshop automated manufacturing system, a numerical illustration is provided in
the next section.
2.4 Conclusions
In this chapter a new model in a flexible jobshop automated manufacturing
systems has been proposed. The innovation is in the multi shops of the same
performance, but different specification. The proposed flexible jobshop model is
associated with an AGV for material handling. The mathematical model is
identifying the optimal material flow amongst the shops with respect to job
sequence, cycle time, and AGV capability constraints. Also a performance
analysis has been worked out on the selected shops to consider their capability for
satisfying demands. The numerical illustrations indicate the applicability and
efficiency of the proposed approaches to investigate the proposed problem.
References 25
References
Industrial Engineering (2009) (article in press, corrected proof),
doi:10.1016/j.cie.2009.02.009
Akella, R., Krogh, B.H., Singh, M.R.: Efficient computation of coordinatingncontrols in
hierarchical structures for failure-prone multi-cell flexible assembly systems. IEEE
Trans. Robotics and Automation 6, 659–672 (1990)
Asef-Vaziri, A., Dessouky, M., Sriskandarajah, C.: A loop material flow system design for
automated guided vehicles. Int. J. Flex. Manuf. Sys. 13, 33–48 (2001)
Berman, S., Schechtman, E., Edan, Y.: Evaluation of automatic guided vehicle systems.
Robotics and Computer-Integrated Manufacturing 25(3), 522–528 (2009)
Bozer, Y.A., Srinivasan, M.M.: Tandem configuration for automated guided vehicle
systems and the analysis of single vehicle loops. IIE Trans. 23, 72–82 (1991)
Bozer, Y.A., Srinivasan, M.M.: Tandem AGV systems: a partitioning algorithm and
performance comparison with conventional AGV systems. Eur. J. Operat. Res. 63, 173–
191 (1992)
Castillo, I., Reyes, S.A., Peters, B.A.: Modeling and analysis of tandem AGV systems
using generalized stochastic Petri nets. Journal of Manufacturing Systems 20(4), 236–
249 (2001)
Demange, M., Ekim, T., de Werra, D.: A tutorial on the use of graph coloring for some
problems in robotics. European Journal of Operational Research 192(1), 41–55 (2009)
Gen, M., Lin, L., Zhang, H.: Evolutionary techniques for optimization problems in
integrated manufacturing system: State-of-the-art-survey. Computers & Industrial
Engineering 56(3), 779–808 (2009)
Gershwin, S.B.: Hierarchical flow control: a framework for scheduling and planning
discrete events in manufacturing systems. Proceedings of the IEEE 77, 195–209 (1989)
Ho, Y.C., Liao, T.W.: Zone design and control for vehicle collision prevention and load
balancing in a zone control AGV system. Computers & Industrial Engineering 56(1),
417–432 (2009)
Huang, Y., Liang, C., Yang, Y.: The optimum route problem by genetic algorithm for
loading/unloading of yard crane. Computers & Industrial Engineering 56(3), 993–1001 (2009),
Kaspi, M., Kesselman, U., Tanchoco, J.M.A.: Optimal solution for the flow path design
problem of a balanced unidirectional AGV system. Int. J. Prod. Res. 40, 349–401 (2002)
Laporte, G., Zanjirani Farahani, R., Miandoabchi, E.: Designing an efficient method for
tandem AGV network design problem using tabu search. Applied Mathematics and
Computation (2006)
Rajagopalan, S., Heragu, S.S., Taylor, G.D.: A Lagrangian relaxation approach to solving
the integrated pick-up/drop-off point and AGV flow path design problem. Appl. Math.
Model. 28, 735–750 (2004)
Chapter 3
Nonlinear Stochastic Model for AGV System
3.1 Summary
Advanced automated manufacturing systems are widely used in industrial

companies where productivity objectives have to be met. These systems often
being costly, they must be designed to be as efficient as possible. Here, we focus
on automated manufacturing systems in a job shop layout considering automated
guided vehicle as a material handling resource. The key issue in manufacturing
operations is how to produce high quality products at low costs in such a way that
the diversified demand are met. Hence, modern manufacturing companies should
become as responsive as possible in order to satisfy customer demands.
Material handling accounts for 30–75% of the total cost of a product, and
efficient material handling can result in reducing the manufacturing system
operations cost by 15–30% (Sule, 1994). These points underscore the importance
of material handling costs reduction as a key element in improving the cost
structure of a product. The determination of a material handling system involves
both the selection of suitable material handling equipment and the assignment of
material handling operations to each individual piece of equipment. Hence,
material handling system selection can be defined as the selection of material
handling equipment to perform material handling operations within a working area
considering all aspects of the products to be handled (Rau and Liu, 2006).The
material handling system plays a crucial role in automated manufacturing systems.
When inadequately designed, the material handling system indeed can adversely
affect the overall performance of the system and lead to substantial losses in
productivity and competitiveness, and to unacceptably long lead times. Thus, to
avoid such pitfalls, material handling system design must be integrated into the
overall design of the manufacturing system centering on the selection of machines
and the allocation of operations to the machines (Gholipour-Kanani at al., 2011).
Automated manufacturing systems (AMSs), equipped with several CNC
machines and an AGV-based material handling system, are designed and
implemented to secure automation and gain efficiency of production. To achieve
these benefits, the planning in the AMS decision making process is critical
because the planning decision has an influence on the subsequent decision

28 3 Nonlinear Stochastic Model for AGV System
processes such as scheduling, dispatching, etc. Planning in automated

manufacturing systems can be characterized to be online and short-term in order to
respond to frequently changing production orders. Given a production order, the
manufacturing planning function is responsible for establishing a plan by
decomposing the production task into a set of subtasks (Azadeh and Anvari,
2009).Subtasks are classified into two types for material processing and delivery.
To pursue a global optimization of material processing and transportation cost, a
manufacturing plan should consider the two costs simultaneously. Moreover, to
obtain a better solution for subsequent decision making processes, not only total
workloads are evenly distributed among the workstations, but workloads are well
balanced among AGV’s network segments.
The manufacturing industry is presently being much affected by the structural
changes caused by the internal and external factors in an enterprise. The market
conditions are becoming more dynamic, more global, and more customer-driven.
The manufacturing performance is no longer driven merely by the product price;
instead, other competitive factors such as automation, quality, delivery, and
customer service have become equally important. The demand of the customer for
tailored product has caused a shorter product life, reduced batch quantities and
increased product varieties.
Manufacturing firms need to give prominence to issues such as reduction of
manufacturing lead time and have flexibility to adapt to changes in the market.
The improvement in productivity and reduction of costs of goods and services has
become a key element for maintaining the market share.Operations management
in an AMS is more complex than that of the conventional manufacturing systems.
Managing an AMS demands for more decisions for its effective performance as
compared to a transfer line or a job shop production system. The optimal selection
of machines and tools and the assignment of part operations to the selected
machines turn out to be difficult tasks for the production planner. This is due to
the versatile machine tools capable of performing many different operations
resulting in many alternative routes for a part type, and due to system capability
for processing of the parts concurrently (Stecke, 1983).
Chan (1999) discussed the effects of universal loading station along with
operational control rules. Rajagopalan (1986) presented a formulation and
heuristic solution for the part grouping and tool loading in AMS. Ram et al. (1986)
developed a model and a solution procedure for the machine loading and tool
allocation problem in an AMS. Lashkari et al. (1987) extended the formulation of
the operation allocation problem to include the aspects of refixturing and limited
tool availability. Kusiak (1985), Stecke and Morin (1985), Singhal et al. (1987)
and van Looveren et al. (1993) addressed the interrelationship of various decisions
and their hierarchies in AMS. Kim and Yano (1993) presented a number of
heuristic approaches for loading problems in automated manufacturing systems.
Kim and Yano (1994) also presented a model for the loading problem in AMS
with unequal workload targets across machine groups, and demonstrated how an
existing branch-and-bound algorithm for the workload balancing objective may be
used to solve the model. Stecke and Raman (1994) presented a queuing network
production planning model to determine the optimal machine workload

assignments in an AMS. Liang (1994) proposed a two stage approach to the joint
problem of part selection, machine loading, and machine speed selection problem
in AMS. In the first stage, the mathematical model solves the part selection and
machine loading problem, whereas in the second stage, it determines the optimal
cutting speed for all job–tool–machine combinations.
A wide spectrum of multi-objective loading problems by combining two or
more criteria has been addressed in literature (Swarnkar and Tiwari, 2004; Chan et
al., 2004). However, considering the multi-objective problems, some of the
objectives turn out to be contradictory in various situations, while for others, they
may equally be applicable. When the machines are not pooled into groups, the
loading objective used most commonly is balancing the workload on all the
machines. It is established that this objective maximizes the expected production
(Stecke and Morin, 1985). The mixed-integer programming (MIP) approach for
solving the machine loading problem is shown to be computationally infeasible
even for deterministic formulation.
Here, considering the stochastic nature of the proposed problem parameters, a
stochastic programming approach is applied to optimize production time and cost
in an automated manufacturing system.
In our proposed model, it is assumed that a manufacturing system equipped with

an AGV for material handling purposes is designed in a job shop layout. The
AGV would receive the raw material and start visiting the shops to complete the
production cycle. In each shop, one part of the product is processed and at the
final shop the complete product is ready to be carried to the warehouse.
The configuration of the proposed model is represented in Figure 1.
Fig. 1 Configuration of the proposed model

The AGV should wait until the processing of a job is finished in a shop and
then move the semi-produced product to the next shop. Therefore, due to the
processing conditions in a shop, the waiting time in a shop is not deterministic and
is therefore supposed to be stochastic. While a job is processing on a machine, the
machine may break down. Due to the stochastic nature of the failure, the
breakdown cost is not known. Also, the semi-produced products carried by AGV
may be damaged during traveling between shops, but the rate is not deterministic
and thus is considered to be stochastic.
Considering these assumptions, we are confronted with a stochastic model. The
distributions of the stochastic items are assumed to be normal or can be
appropriately estimated by normal distributions. In the next section, the
mathematical model of the proposed problem is given.
Note that some parameters are stochastic, based on the nature of the problem.
Therefore, the stochastic parameters are presented using the corresponding
expected value and variance. As a result, the following nonlinear deterministic
mathematical model is configured. As mentioned before, the aim is to optimize
production time and cost simultaneously. Hence, the model is multi-objective. The
notations needed for the model are indicated below.
Indices:
i Number of shops; i=1,2,…,I
j Type of machines; j=1,2,…,J.
Notations:
Cd Defect cost
Co Operational cost of AGV
Cti Tool cost in the ith shop
Cb j Breakdown cost for jth machine
tp j Job processing time on jth machine
twij Waiting time in ith shop for jth machine
tm Material handling time

tc Cycle time
β Upper limit for defect rate
B1 Available budget for AGV
B2 Available budget for tool

Di Demand for ith shop
Mi Number of machines in ith shop
N Number of jobs
E(.) Expected value of a parameter
Var(.) Variance of a parameter
ZP Standard normal value for percentile P.
Decision variables:
Xij Number of produced products in ith shop by jth machine
Objective function 1 (cost minimization):

J I J I J I
Min (∑∑ E (C b j ) × X ij ) + Z P (∑∑ Var (C b j ) × X ij2 ) + (C o × (∑∑ C ti × X ij )) +
j =1 i =1 j =1 i =1 j =1 i =1
(1)
J I J I
( E (C d ) × ∑ ∑ X ij ) + Z P Var (C d ) × (∑∑ X ij2 ) ,
j =1 i =1 j =1 i =1
Objective function 2 (time minimization):

J I J I J I
Min (tm × (∑∑ t p j × X ij )) + ∑∑ E (twij ) × X ij ) + Z P ∑∑Var(t wij ) × X ij2 ) , (2)
j =1 i =1 j =1 i =1 j =1 i =1
Constraints:
J I J I J I
( N × (∑∑ t p j × X ij )) + (∑∑ E (t wij ) × X ij ) + Z P ∑∑ Var (t wij ) × X ij2 ) + (3)
j =1 i =1 j =1 i =1 j =1 i =1
J I J I
(t m × ∑∑ X ij ) ≤ (t c × ∑∑ X ij ),
j =1 i =1 j =1 i =1
J I
Co × ∑∑ X ij ≤ B1 , (4)
j =1 i =1
J I J I
( E (Cd ) × ∑∑ X ij ) + Z P Var(Cd ) × (∑∑ X ij2 ) ≤ β , (5)
j =1 i =1 j =1 i =1
J I Cti × X ij
∑∑
j =1 i =1 Mi
≤ B2 , (6)
∑X
j =1
ij ≥ Di , i=1,2,…,I, (7)
X ij ≥ 0, integer , i=1,2,…,I, j=1,2,…,J. (8)
The formula in (1) is the first objective function, the total cost of production,
which is to be minimized. Formula (2) is the second objective function, the total
time of manufacturing, which is also to be minimized. Inequality (3) confines the
total production time to the cycle time. Inequality (4) considers the limitation of
operational budget for AGVs. Inequality (5) represents a constraint for desirable
defect rate. Inequality (6) indicates that the available budget for tools in the system
is limited. The Inequalities (7) certify that the demands of the shops are covered.
The constraints (8) ensure the nonnegativity and integrality of the variables.
Since the proposed model is nonlinear, we apply a linear approximation method
for optimization.
3.3.1 Successive Linear Programming Approach

Successive Linear Programming (SLP) algorithms solve nonlinear optimization
problems via a sequence of linear programs. They have been widely used,
particularly in the oil and chemical industries, beginning with their introduction by
Griffith and Stewart of Shell Development Company in 1961 (Griffith and
Stewart, 1961). As the name implies, the method uses linear programming as a
search technique. A starting point is selected as an initial estimate of the solution,
and the nonlinear model and constraints are linearized at this point to give a linear
problem which can be solved by the simplex method or its variations. The solution
of the linear programming problem is used to construct a new estimate of the
solution of the original problem, and the process is continued until a stopping
criterion is met. As shown by Reklaitis et al. (1983), this procedure works, with no
need for safeguards, when problems are mildly nonlinear. However, it is necessary
to bound the steps taken in the iterations to insure that the model solution estimate
improves, the values of the independent variables remain in the feasible region
and the process converges to an optimal point. These safeguards are bounds on the
independent variables specified in advance for solving the linear programming
problem. The net result is that the bounds are additional constraints. If the bounds
are set too small, then the process proceeds slowly toward a solution. If they are
set too large, then infeasible solutions may be generated. Consequently,
precautionary steps are incorporated into computer programs to expand the bounds
when slow progress is detected and shrink them when the solution estimates do
not seem to be improving much in successive iterations, while being far from a
stationary point. Now, the general nonlinear optimization problem can be written
as:
Optimize: y(x), (9)
s.t. fi(x) < bi, i = 1, 2, ...,m,

lj ≤ xj ≤ uj , j = 1, 2, ...,n,
where upper and lower limits are shown specifically on the independent variables.
In a general step k, the model y(x) and the constraints fi(x) are linearized at xk to
give:
n
Optimize: z = ∑ c j Δx j + y ( x k ) , (10)
j =1
n
s.t. ∑a
j =1
ij Δx j + f i ( x k ) ≤ bi , i=1,2,…,m,
l j ≤ x kj + Δx j ≤ u j , j=1,2,…,n,
where,
∂y ( x k ) ∂f i ( x k )
cj = , aij = .
∂x j ∂x j
The problem (10) is now a linear programming one, but the values of Δxj can take
on either positive or negative values. To have the linear program the standard
nonnegativity imposition on the variables Δxj, a change of variables was made by
Griffith and Stewart (1961) as follows:
Δx j = Δx +j − Δx −j , (11)
where,
⎧Δx j , if Δx j ≥ 0
Δx +j = ⎨
⎩0 , if Δx j < 0
⎧− Δx j , if Δx j ≤ 0
Δx −j = ⎨
⎩ 0, if Δx j > 0.
Substituting (11) into (10), the linear programming problem has then the form:
n n
Optimize: ∑ c j Δx +j − ∑ c j Δx −j + y ( x k ) ,
j =1 j =1
(12)
n n
s.t. ∑ aij Δx +j − ∑ aij Δx −j ≤ bi − f i ( x k ) , i=1,2,…,m,
j =1 j =1
+ −
Δx − Δx ≤ (u j − x kj ) ,
j j
Δx +j − Δx −j ≥ (l j − x kj ) , j=1,2,…,n,
Δx +j , Δx −j ≥ 0, j=1,2,…,n.
The upper and lower limit bounds on the variables are specified by ( u j − x kj ) and
(lj − x kj ). The value of the next solution estimate for linearization is then defined
to be x kj +1 = x kj + Δx +j − Δx −j , where the Δxj+ and Δxj¯ are obtained as the
solution of (12). The procedure is started by specifying a starting point x0(k=0).
We note that the values of the bounds uj and lj may affect the rate of convergence
of the algorithm. To obtain an optimization algorithm using the SLP, we need
some preliminaries.
Let .* and . be two arbitrary norms in ℜ n . It is well known that there are
positive constants c1and c2 such that for all x ∈ ℜ n ,
c1 x * ≤ x ≤ c2 x * . (13)
Note that for every x ∈ X and r > 0 (r is called the trust region radius), the
following inclusions hold:
{h x + h ∈ X , h ≤ c r}⊆ {h x + h ∈ X , h
1 *
} { }
≤ r ⊆ h x + h ∈ X , h ≤ c2 r .(14)
For x ∈ X and r > 0 , consider the following problem:
LP ( x, r ) : min ∇ T f ( x)h
s.t. x + h ∈ X , (15)
h * ≤ r,
where, ∇f ( x ) is a column vector composed of the partial derivations of f at x.
Clearly, when .* is a polyhedral norm and X is a polyhedral set, LP(x,r) is a

linear programming problem. Denote H*(x,r) and v*(x,r) as the solution set and the
optimal value of LP(x,r), respectively. The inclusions (14) imply that
v( x, c2 r ) ≤ v* ( x, r ) ≤ v( x, c1r ) , (16)
for every x ∈ X and r ≥ 0 .
Lemma 1. For every x ∈ X and r > 0 , we have:

x ∈ X stat ⇔ v* ( x, r ) = 0 ,
where, X stat is a stationary set.
In order to justify the technique being used to update the trust region radius, we
need a simple technical result as given next.
Lemma 2. For every δ > 0 and x ∈ X \ X stat , there exist τ and ρ , with
0 < τ < ρ , such that the following properties hold:
(i) ∀r ∈ [ρ , ∞ ), ∀h ∈ H * ( x, r ), f ( x ) − f ( x + h) ≤ δ r 2 .
2
(ii) ∀r ∈ [0, τ ] , ∀h ∈ H * ( x, r ), f ( x) − f ( x + h) ≥ δ
r2.
2
Proof let δ > 0 and x ∈ X \ X stat .
Proof for (i). Define ρ := 2δ −1 ( f ( x) − f opt ) so that for all r ∈ [ρ , ∞) ,

δ δ
f ( x ) − f ( x + h) ≤ f ( x ) − f opt = ρ2 ≤ r2 . (17)
2 2
−1
Proof for (ii). Let a := 2c1 ( Lc 2 + δ ) τ > 0 with
2 2
, so that there exists
−
a0 ( x, c1τ ) ≤ a . Take an arbitrary r ∈ (0, r ] and h ∈ H * ( x, r ) . Then,

− −
a0 ( x, c1 r ) ≤ a0 ( x, c1τ ) ≤ a , (18)
−
and
1 c12 2
v( x, c1 r ) ≤ − (c1r ) ≤ − r .
2
(19)
a0 ( x, c1r ) a
−
It follows from (16) and (18) that
c12 2
∇f ( x), h = v* ( x, r ) ≤ v( x, c1r ) ≤ − r . (20)
a
−
Therefore, by (17),
c12 2 Lc 22 c12 2 Lc 22 2 1 ⎛ 2c 2 ⎞ (21)
r = − ⎜ 1 − Lc 22 ⎟ r 2
L 2 2
f ( x + h ) − f ( x ) ≤ ∇ f ( x ), h + h ≤− r + h ≤− r +
2 a 2 *
a 2 2⎜ a ⎟
− − ⎝ − ⎠
δ
=− r 2.
2
We now state the main algorithm. The parameter k counts the number of
iterations, while μk represents the number of subproblems solved at the kth
iteration.
Algorithm SLP: Successive Linear Programming for Nonlinear Optimization
Parameters: δ , r* > 0, 0 < θ < 1.

Input: Choose x 1 ∈ X .
Step 0: (Initialization): Set k:=0, rk :=r*, kstop:= ∞ .
Step 1: Set k:=k+1, μ := 1 , r:=rk-1. Solve the problem LP(xk,r) to
obtain a solution h* ( x k , r ) ∈ H ( x k , r ) . Set

⎧
⎪ 0, if v* ( x k , r ) = ∇f ( x k ), h* ( x k , r ) = 0,
⎪⎪ δ
v := ⎨− 1, if v* ( x k , r ) ≠ 0 and f ( x k ) − f ( x k , h* ( x k , r )) ≥ r 2 ,
⎪ 2
⎪ 1, if v* ( x k , r ) ≠ 0 and f ( x k ) − f ( x k , h* ( x k , r )) < δ r 2 .
⎪⎩ 2
k stop
If v=0 then set kstop=k, x = x k and return.
Step 2: Set μ := μ + 1 and r := rθ v . Solve the problem LP(xk,r) to
obtain a solution h* ( x k , r ) ∈ H ( x k , r ) .
3.4 Conclusions 37
δ
Step 3: If ( v = −1 and f ( x k ) − f ( x k , h* ( x k , r )) < r 2 ) or (v=1 and
2
δ
f ( x ) − f ( x , h* ( x , r )) ≥
k k k
r ) then go to Step 4, else go to
2
2
Step 2.
Step 4: If v = − 1 then set rk = θr , else set rk = r . Let μ k = μ and
x k +1 = x k + h* ( x k , rk ) and go to Step 1.
Note that is Step 1, the parameter v can take on one of three values. If v=0, then a
stationary point has been found and the algorithm terminates. If v=1, then
sufficient decrease was not achieved and thus the subproblem needs to be resolved
with a smaller value of r. If v = − 1 , then sufficient decrease was achieved, but
the subproblem is solved again, this time with a larger value of r. Both of these
resolves take place in Step 2. If the algorithm terminates after a finite number of
kstop
iterations, then we can define the iterate x as the approximate solution for the
problem. We apply the proposed SLP algorithm to solve some numerical
experiments.
3.4 Conclusions
A new model for automated manufacturing systems was proposed. Material

handling was performed via AGV and the manufacturing system was considered
to be a jobshop. The novelty of our approach is in our consideration of the
stochastic nature of parameters and our use of stochastic programming. The
mathematical model is a nonlinear bi-objective one, which considers both time
and cost minimization. We first applied a successive linear programming
technique for optimizing the nonlinear model. Numerical test results showed the
linear optimization approach to be slow for large size problems, and thus
presented a genetic algorithm (GA) for solving large size problems. We finally
analyzed the differences of the solutions obtained by LINGO and GA using test of
hypothesis. We concluded that there was a significant difference between
LINGO's and GA's obtained solutions and that their performance was not the same
due to variety in the problem size. For small and medium sized problems, LINGO
was shown to be efficient enough to provide accurate solutions but for larger sized
problems, we observed that the GA was successfully effective in finding good
solutions.
References
Azadeh, A., Anvari, M.: Implementation of Multivariate Methods as Decision Making
Models for Optimization of Operator Allocation by Computer Simulation in CMS.
Journal of the Chinese Institute of Industrial Engineers 26(4), 316–325 (2009)
Chan, F.T.S.: Evaluations of operational control rules in scheduling a flexible
manufacturing system. Robot Computer Integrated Manufacturing 15, 121–132 (1999)
Chan, F.T.S., Swarnkar, R., Tiwari, M.K.: A random search approach to the machine
loading problem of an FMS. Presented at 19th IEEE International Symposium on
Intelligent Control, Taipei, Taiwan, pp. 152–161 (2004)
Eren, T.: A multicriteria flowshop scheduling problem with setup times. Journal of
Materials Processing Technology 186(1-3), 60–65 (2007)
Eren, T., Guner, E.: A bicriteria flowshop scheduling problem with setup times. Applied
Mathematics and Computation 183(2), 1292–1300 (2006)
Eren, T., Guner, E.: Setup times with a learning effect in flowshop scheduling problem.
Journal of the Faculty of Engineering and Architecture of Gazi University 22(2), 353–
362 (2007)
Gholipour-Kanani, Y., Tavakkoli-Moghaddam, R., Khorrami, A.: Solving a multi-criteria
group scheduling problem for a cellular manufacturing system by scatter search. Journal
of the Chinese Institute of Industrial Engineers 28(3), 192–205 (2011)
Griffith, R.E., Stewart, R.A.: A nonlinear programming technique for the optimization of
continuous processing systems. Management Science 7, 379–395 (1961)
Hsu, Y.S., Lin, B.N.T.: Minimizing of maximum lateness under linear deterioration.
Omega 31, 459–469 (2003)
Kim, Y.D., Yano, C.A.: A heuristic approach for loading problems in flexible
manufacturing systems. IIE Trans. 25, 26–39 (1993)
Kim, Y.D., Yano, C.A.: A new branch and bunch algorithm for loading problems in
flexible manufacturing system. Int. J. Flexible Manuf. Systems 6, 361–382 (1994)
Kusiak, A.: Loading models in flexible manufacturing systems. In: Raouf, A., Ahmed, S.H.
(eds.) Manufacturing Research and Technology-1, pp. 78–93. Elsevier, Amsterdam
(1985)
Lashkari, R.S., Dutta, S.P., Padhye, A.M.: A new formulation of the operation allocation
problem in flexible manufacturing systems: mathematical modelling and computational
experience. Int. J. Prod. Res. 25, 1367–1383 (1987)
Liang, M.: Integrating machining speed part selection and machine loading decisions in
flexible manufacturing systems. Comput. Ind. Eng. 26, 599–608 (1994)
Rajagopalan, S.: Formulation and heuristic solution for part grouping and tool loading in
flexible manufacturing systems. In: Stecke, K., Suri, R. (eds.) Proceedings of the Second
ORSA/TIMS Conference on Flexible Manufacturing Systems, pp. 311–320. Elsevier,
Amsterdam (1986)
Ram, B., Sarin, C.C.S.: A new branch and bound algorithm for loading problems in flexible
manufacturing system. Int. J. Prod. Res. 25, 1081–1094 (1986)
Rau, H., Liu, C.-K.: The Optimal Combination of Postponement Operations in a Supply
Chain. Journal of the Chinese Institute of Industrial Engineers 23(3), 253–261 (2006)
Reklaitis, G.V., Ravindran, A., Ragsdell, K.M.: Engineering Optimization, Methods and
Applications. John Wiley and Sons, Inc., New York (1983)
References 39
Shirazi, B., Fazlollahtabar, H., Sahebjamnia, N.: Minimizing Arbitrary Earliness/Tardiness

Penalties with Common Due Date in Single-Machine Scheduling Problem Using a
Tabu-Geno-Simulated Annealing. Materials and Manufacturing Processes 25(6), 515–
525 (2010)
Singhal, K., Fine, C.H., Meredith, J.R., Suri, R.: Research and models for automated
manufacturing. Interfaces 17, 5–14 (1987)
Stecke, K.E.: Formulation and solution of nonlinear integer production planning problem
for flexible manufacturing system. Manage. Sci. 29, 273–288 (1983)
Stecke, K.E., Morin, T.L.: The optimality of balancing workload in certain types of flexible
manufacturing system. Eur. J. Oper. Res. 20, 68–82 (1985)
Stecke, K.E., Raman, N.: Production planning decision in flexible manufacturing systems
with random material flow. IIE Trans. 26, 2–17 (1994)
Sule, D.R.: Manufacturing Facilities: Location, Planning, and Design. PWS, Boston (1994)
Swarnkar, R., Tiwari, M.K.: Modelling machine loading problem of FMSs and its solution
methodology using a hybrid tabu search and simulated annealing based heuristic
approach. Robot Comput. Integrated Manuf. 20, 199–209 (2004)
Van Looveren, A.I., Gelders, J.L.F., van Wassenhove, L.N.: A review of FMS planning
models. In: Kusiak, A. (ed.) Modelling and Design of Flexible Manufacturing System,
pp. 3–31. Elsevier, Amsterdam (1986)
Chapter 4
Reliability Model for AGV
4.1 Summary
The Material Handling System (MHS) in a manufacturing setting plays an

important role in the performance of the entire system. Inadequately designed
MHSs can interfere with the overall performance of the manufacturing system and
lead to substantial losses in productivity and competitiveness, and to unacceptably
long lead times. Among the advanced technologies available for MHSs,
Automated Guided Vehicles (AGVs) have found increasing applications because
of their capability to transport a variety of part types from point to point without
human intervention.
Today’s automated MHSs are technologically advanced and increasingly
complex. Uncertainty is an inevitable consequence of the complexities generated
by technological advancements. Jain et al. (2013) show that most of the automated
manufacturing studies have used single-item measures. They argue that single-
item measures are appropriate for relatively simple manufacturing systems. As the
different dimensions of automated MHSs are complex in nature, single-item
measures are inappropriate. Choosing adequate and relevant performance
measures is critical in accurately analyzing MHSs (Beamon, 1998). A company
also needs to address uncertainties in the manufacturing system to survive and
compete in such an uncertain environment (Jain et al. 2013).
AGVs are the most flexible means for transporting pieces among workstations
in an automated manufacturing system. An AGV is a driverless and programmed
vehicle used to transfer the load from one part of a manufacturing facility to
another part (Maniya and Bhatt, 2011). Maxwell and Muckstadt (1982) first
recognized the importance of AGV-based MHS design. They developed an
optimization model that minimized the total travel time and determined the
maximum number of AGVs needed to efficiently transfer material from one shop
to another. AGVs increase efficiency, reduce costs, and improve flexibility by
automating an MHS. Time-based performance measures are often used to evaluate
MHSs with AGVs. However, moving materials from one part of the
manufacturing floor to another part utilize time and incur costs. In this study, we
consider both time and cost measures in an optimization model and evaluate an
MHS with AGVs.

42 4 Reliability Model for AGV
The use of AGVs increases flexibility and has a significant impact on the
overall performance and reliability of MHSs (Sarker and Gurav, 2005). As AGVs
become larger and more complex, the traditional design requires more attention to
issues such as control, cost, time, reliability, flexibility, etc. A number of different
MHS design and evaluation methods (e.g., simulation, optimization, and the
genetic algorithm) have been proposed in the literature. Simulation is an
acceptable method for analyzing manufacturing systems. However, simulation is
often challenging and time consuming (Law and Kelton, 2000; Kuo et al., 2007),
particularly, when it is used for modeling complex manufacturing systems such as
MHSs with AGVs.
The problem of scheduling AGVs in an automated MHS has been studied
extensively. Abdelmaguid et al. (2004) addressed the problem of simultaneous
scheduling of machines and AGVs with the objective of minimizing the
makespan. This problem is composed of two interrelated decision problems: the
scheduling of machines, and the scheduling of AGVs. They showed that each
problem is an NP-complete problem and a simultaneous consideration of the two
problems results in a more complicated NP-complete problem. They proposed a
hybrid genetic-algorithm/heuristic coding scheme to solve the problem. Deroussi
et al. (2008) also studied this problem and proposed a solution based on vehicles
rather than machines. Each solution was evaluated using a discrete event
approach. Gnanavel Babu et al. (2010) studied this problem further and proposed a
meta-heuristic differential evolution algorithm for solving it. They introduced an
iterative algorithm that anticipated the complete set of flow requirements for a
given machine schedule and made vehicle assignments accordingly. Le-Anh and
De Koster (2006) have compiled a comprehensive review of the AGV design and
control models and methods in the literature.
Farling et al. (2008) used a simulation model to compare the performance of
three AGV configurations under a variety of experimental conditions. They
showed that system size, load/unload time, and machine failure rate factors have
significant impacts on the operation and reliability of MHSs. Smith (1993) defined
reliability as the probability that an item will perform a required function, under
stated conditions, for a specific period of time. A reliability measure is a metric for
quantifying this probability. A number of different reliability measures (i.e.,
availability, unavailability, failure rate, and mean time between failures) have
been proposed in the literature. For degradable systems, such as MHSs, the
performance of the system during a specific period of time can be described by
different levels of performance as a function of machine failures (Beamon, 1998).
Miriyala and Viswanadham (1989) developed several measures and algorithms for
evaluating part-based reliability and system-based reliability for automated MHSs.
Beamon (1995) proposed an analytical model for designing guide paths for
automated MHSs as a function of reliability and quantified the reliability of the
handling components.
In order to ensure an acceptable service level for each machine in each shop,
we adopt and further extend the concept of reliability proposed by Ball and Lin
(1993) in the model. We define reliability as the probability that the system is
operational until time t. A failure is when a machine in a shop breaks down.
4.1 Summary 43
A desired level of reliability can be achieved by limiting the failure probabilities.

This approach for handling reliability is called the chance constraints method and
was initially proposed by Charnes and Cooper (1959) in the context of
mathematical programming. The use of chance constraints in the vehicle routing
problem was illustrated in Stewart and Golden (1983). Carbone (1974) used
chance constraints for selecting multiple facilities under normally distributed
demand. The model minimized an upper bound on the total demand-distance
while ensuring that the constraints were satisfied with a specified chance or
probability. Shiode and Drezner (2003) used a similar approach in a competitive
location problem on a tree network.
In many real-life applications, the parameters in a manufacturing system may
have varying values. This value variation may result from machine breakdowns,
lack of training, unexpected delays, non-qualified operators, or complex tasks,
among others (Özcan, 2010). Stochastic programming can provide an effective
means for incorporating uncertainty in real-life MHSs (Birge and Louveaux,
1997). Stochastic programming has been used frequently in the design and control
of MHSs with AGVs (Sayarshad and Tavakkoli-Moghaddam, 2010).
The presence of uncertainties in automated material handling and AGV systems
has also motivated researchers to explore compensator design such as neural
networks (Pamosoaji et al., 2013). Kuo et al. (2007) argues that simulation is very
time consuming for large MHSs and the process of collecting adequate sample data
places limitations on any analysis. They proposed to overcome this problem by
developing a neural network simulation metamodel that required only a comparably
small training data set. Artificial neural networks are composed of interconnected
adaptive elements which are intended to respond to stimuli in a manner not unlike
the human nervous system (Kohonen, 1988). McCulloch and Pitts (1943) introduced
one of the first artificial neuron models in 1943. The main feature of their neuron
model is that a weighted sum of input signals is compared to a threshold to
determine the neuron output. Unlike biological networks, the parameters of their
networks had to be designed, as no training method was available. However, the
perceived connection between biology and digital computers generated a great deal
of interest. In the late 1950s, Rosenblatt (1985) and several other researchers
developed a class of neural networks called perceptrons. The neurons in these
networks were similar to those of McCulloch and Pitts (1943). Rosenblatt’s (1985)
key contribution was the introduction of a learning rule for training perceptron
networks to solve pattern recognition problems.
In this chapter, we consider a manufacturing system with the following
physical characteristics: (1) the manufacturing system is a job shop, and (2)
single-load AGVs perform the material handling job in the shop. We use a
stochastic programming framework and propose a bi-objective optimization model
(which has not been used in the MHS evaluation studies) to determine the optimal
production time and cost in a manufacturing system with an automated MHS and
AGVs. The contribution of this chapter is fivefold: we consider (1) a stochastic
programming problem and decompose the optimization process into manageable
steps and integrate the results to arrive at a solution consistent with organizational
goals and objectives; (2) stochastic parameters in multi-objective optimization
models in general and the proposed bi-objective optimization model in particular;

(3) machine reliability as an important component in the proposed optimization
model; (4) machine maintenance through the breakdown rate for a realistic
representation of the system; and (5) the elements of uncertainty within the
proposed structured framework by using a perceptron neural network to weigh the
two objectives in the proposed bi-objective optimization model.

In this problem, we consider a manufacturing system equipped with an AGV for
material handling in a job shop environment. The AGV moves a part-type from
one shop to another to complete a production cycle. In each shop, the part-type is
partially processed. The part-types completed in the final shop are moved to the
warehouse. For example, let us consider a manufacturer of cylinder block in the
automotive industry. The manufacturer of cylinder block can install an AGV
system to supply the lines with parts and to transfer cylinder blocks between five
shops. Generally, the sequence of operations on a typical machining line for
cylinder blocks involves five shops as shown in Figure 1:
This machining line for cylinder blocks is comprised of the following 16
processes performed in each of the five job shops:
• Shop 1: Assuring a uniform wall thickness for the cylinder bores - qualifying
(Process-1); rough mill pan and head faces (Process-2); and rough machining
cylinder bores (Process 3)
• Shop 2: Milling bearing cap width and slots (Process 4) and finishing mill pan
and bearing cap width (Process 5)
• Shop 3: Drilling oil holes - compound angles (Process 6); drilling, reaming,
tapping - Left and right, pan and head faces (Process 7); assembly of bearing
caps (Process 8); finish front and rear end (Process 9); and drilling, reaming,
tapping - end faces (Process 10)
• Shop 4: Line boring crankbore (Process 11); finish tappet bores (Process 12);
assembly cam liners, finish line boring (Process 13); finish cylinder bores
(Process 14); and finish mill/grind head face (Process 15)
• Shop 5: Hone and grade (Process 16)
The AGV waiting time in the job shop is not known precisely. It is also
possible for a machine to breakdown while processing a job. The number of
breakdowns (and the breakdown cost) is also not known precisely. The
distributions for the stochastic parameters in the problem are assumed to be
normal or can be estimated by the normal distribution. Additionally, the machines'
reliability is also considered as a stochastic parameter following the exponential
distribution. We present the mathematical details of the proposed model in the
next section.
Fig. 1 Cylinder block machining line example

In this section, we propose a mathematical model for a simultaneous optimization
of the production times and costs.
4.3.1 Reliability Component

We assume that the reliability of each machine-type is independent according to
Exponential processes. Also, J is the total number of machine types (i.e., milling,
drilling, turning, welding, etc.) We discuss the reliability based model as follows:
⎧⎛ J
⎞
⎪⎜ 1 − ∏(1 − R j (t )) ⎟ , when machines in each shop are in parallel
⎪⎝
⎪ j =1 ⎠ (1)
R(t ) system =⎨
⎪ ⎛ J ⎞
⎪ ⎜ ∏R j (t ) ⎟ , when machines in each shop are in series
⎩⎪ ⎝ j =1 ⎠
where R j (t ) is the probability that the machine-type j works for a period of t time
units.
As stated earlier, the machine-types in each shop are parallel and the shops are
organized in series. Therefore, the reliability of the system can be measured as
follows:
⎛ J ⎞
⎜1 − ∏ (1 − R j (t )) ⎟ ≥ α , (2)
⎜ ⎟
⎝ j =1 ⎠
where α is the lower bound for a desirable system reliability during the time
period t. As previously assumed the reliability of each machine-type is
independent and can be measured according to the following exponential
distribution:
−t
θj
R j (t ) = e , (3)
where θ j is the exponential parameter for machine-type breakdown. Then,
⎛ J
−t
⎞
⎜1 − θj ⎟
⎜ ∏
(1 − e ) ≥α (4)
⎟
⎝ j =1
⎠
In order to obtain a higher level of reliability, more cost is incurred to the system.
Hence, a cost function ( C j (t ) ) is defined to keep machine-type jreliable for the
time period t. The following represents the cost for the entire system:
J
∑C
j =1
j (t ) (5)
In order to validate the stated cost function, we consider the costs as losses and use
the minimum expected loss (or minimum risk) associated with the system. We
represent these losses with a quadratic loss function which is mathematically more
tractable than other loss functions because of its symmetric property (i.e., an error
above the target causes the same loss as the same magnitude of error below the
target). If the target value of the pre-planned exponential parameter is f, then a
quadratic loss (cost) function is
2
C j (t ) = L f − θ j (6)
Where L is a constant and its value could be set to 1 if the constant makes no
difference to a decision.
We replace our proposed stochastic parameters with a combination of the
expected value and the variance for that parameter and the following nonlinear
deterministic mathematical model is derived. As indicated earlier, our goal is to
simultaneously optimize production time and cost in the following bi-objective
model:
Indices:
i Number of shops i=1,2,…,I
j Number of machines j=1,2,…,J
Parameters:
C Stochastic cost per defective unit
d
Co AGV operational cost per unit produced
Cti Tool cost for shop i
tpj Job processing time on machine j
twij Stochastic waiting time of machine j in shop i
tm Material handling time

tc Cycle time
β Total daily cost of the defective items
B1 Total AGV budget
B2 Total tools budget
B3 Total machines budget
Di Total Demand for shop i
Mi Number of machines in shop i
N Number of jobs
ZP Standard normal Z value for percentile P
Decision variables:
Xij Number of units produced in shop i by machine j
θ j
Exponential reliability function of machine j
Objective function 1 (cost minimization):

⎛ J I ⎞ ⎛ J I ⎞ ⎛⎛ I ⎞ J I ⎞
Min ⎜ ∑∑ E ( C j (t ) ) . X ij ⎟ + Z P ⎜ ∑∑Var ( C j (t ) ) X ij2 ⎟ + ⎜ ⎜ ∑ Cti + Co ⎟ .∑∑ X ij ⎟ + (7)
⎝ j =1 i =1 ⎠ ⎝ j =1 i =1 ⎠ ⎝ ⎝ i =1 ⎠ j =1 i =1 ⎠
⎛ J I
⎞ ⎛ J I 2⎞
⎜ E (Cd ).∑∑ X ij ⎟ + Z P Var (Cd ). ⎜ ∑∑ X ij ⎟
⎝ j =1 i =1 ⎠ ⎝ j =1 i =1 ⎠
Objective function 2 (time minimization):

⎛⎛ J ⎞ J I ⎞ ⎛ J I ⎞ ⎛ J I 2⎞
⎜⎜ ⎜ ∑ t p j + tm ⎟ .∑∑ X ij ⎟⎟ + ⎜ ∑∑ E (t wij ). X ij ⎟ + Z P ⎜ ∑∑ Var (t wij ). X ij ⎟
Min (8)
⎝ ⎝ j =1 ⎠ j =1 i =1 ⎠ ⎝ j =1 i =1 ⎠ ⎝ j =1 i =1 ⎠
Subject to:
⎛⎛ J ⎞ ⎛ J I ⎞⎞ ⎛ J I ⎞ ⎛ J I 2⎞
⎜⎜ ⎜ N ∑ t p j ⎟ . ⎜ ∑∑ X ij ⎟ ⎟⎟ + ⎜ ∑∑ E (twij ). X ij ⎟ + Z P ⎜ ∑∑Var (twij ). X ij ⎟ +
⎝ ⎝ j =1 ⎠ ⎝ j =1 i =1 ⎠ ⎠ ⎝ j =1 i =1 ⎠ ⎝ j =1 i =1 ⎠ (9)
⎛ J I
⎞ ⎛ J I
⎞
⎜ tm .∑∑ X ij ⎟ ≤ ⎜ tc .∑∑ X ij ⎟
⎝ j =1 i =1 ⎠ ⎝ j =1 i =1 ⎠
J I
C o .∑∑ X ij ≤ B1 , (10)
j =1 i =1
⎛ J I ⎞ ⎛ J I 2⎞
⎜ E ( C d ∑∑
). X ij ⎟ + Z P Var ( C d ⎜ ∑∑ X ij ⎟ ≤ β ,
). (11)
⎝ j =1 i =1 ⎠ ⎝ j =1 i =1 ⎠
⎛ J
−t
⎞
⎜1 − θj ⎟
⎜ ∏
(1 − e ) ≥ α, (12)
⎟
⎝ j =1
⎠
J I Cti . X ij
∑∑
j =1 i =1 Mi
≤ B2 , (13)
⎡ J ⎤
⎢ ∑ C j (t ) ⎥ . Rij (t ) system ≤ B3 , (14)
⎣ j =1 ⎦
J
∑X
j =1
ij ≥ Di , i=1,2,…,I, (15)
X ij ,θ j ≥ 0 , i=1,2,…,I; j=1,2,…,J. (16)

Note that, E() and Var() are the expected value and the variance of the
stochastic parameters, respectively. Equation (7) is the first objective function
intended to minimize the total cost of production. Equation (8) is the second
objective function and is intended to minimize the total time of production.
Equation (9) shows that the total production time is limited to the cycle time.
Equation (10) indicates the limitation of the operational budget for the
AGVs. Equation (11) represents a constraint for the acceptable defect rate.
Equation (12) shows that the reliability of the system is restricted to a lower-
bound α . Equation (13) indicates that the total available budgets for the required
tools in each shop are limited to a pre-specified upper-bound value. Equation (14)
indicates that the total available budget for the reliability of the machines is
limited. Equation (15) indicates that the demand at each shop must be satisfied.
Equation (16) enforces the non-negativity of the variables. Finally, θj represents
the stochastic breakdown of machine j. θj is one of our decision variables, and
our goal is to obtain the θj values that can ensure no machine breakdown.
4.3.2 Perceptron Neural Network

In this study we apply perceptron to weight our time and cost objectives. To
conceptualize a perceptron structure for our objectives consider a single-neuron
perceptron with two-input and one neuron, as shown in Figure 2.
p1
w11
∑
p2 w12
a = hardlim ( w (pT ) + b )
Fig. 2 Two input/single-output perceptron

In this figure, p1 and p2 are the inputs and w11 and w12 are their corresponding
weights, respectively. Also, w(Tp ) is the weighted input corresponding to the
target output. The output of this network is determined by:
a = hardlim(n) = hardlim(w(pT ) + b) = hardlim(w11. p1 + w12. p2 + b) . (17)
The decision boundary is determined by the input vectors for which the net
input n is zero; that is,
n = w(pT ) + b = w11. p1 + w12. p2 + b = 0 . (18)
The weights are the learning rates used in the proposed mathematical model.
The summation of the learning weights are used as a single learning rate ξ . In
order to configure our perceptron network, we consider the two factors of machine
and operation as the inputs for time and the two factors of material and operator as
the inputs for cost. The goal here is to find the final weights of the time and cost
outputs. The perceptron network is run using the data observed in the
manufacturing system and the weights of W1 (for time) and W2 (for cost) are
obtained after the convergence. The perceptron computations are implemented in
the MATLAB 7.0 package. These weights are used to unify the objective function.
The weights are 0.2 for time and 0.1 for cost.
The mathematical model gives us the θ j s considering a confidence level as a
reliability of the job-shop system. Now, we investigate the θj s for the shops
separately by data collection. The aim is to analyze the θ j s obtained from data
collection which does not have any interactions with other shops, in comparison
with the θ j s gained from the mathematical model in the last section. We collect
data for a specific working time t=24 (minute). Then for t>24, our data are type I
censored data. Assume x1:n , x 2:n , … , x r:n are the r censored data in a specific
shop, then the estimated θ ( θˆ ) using maximum likelihood estimation (MLE) is as

follows:
Assume that X follows an exponential distribution,
X ≈ EXP (θ ) , (23)
The cumulative distribution function is,

−x
F ( x) = 1 − e θ , (24)
The probability density function is,

−x
1
f ( x) = eθ , (25)
θ
Then for all xi:n , the ordered statistic is,
n! ⎛ ⎞ r 1 − xi:n
− t0
g ( x1:n , x2:n ,..., xr:n ) = ⎜1 − (1 − e θ ) ⎟.∏ e θ , (26)
( n − r )! ⎜⎝ ⎟ i =1 θ
⎠
The likelihood function is,
r
∑ x i :n + ( n − r ) t 0
i =1
n! −
L (θ ) = .θ − r .e θ
, (27)
( n − r )!
The logarithm of both sides of (27) gives,
r
n! ∑ xi:n + (n − r )t 0
ln( L(θ )) = ln − r ln θ − i =1 , (28)
(n − r )! θ2
The partial derivative of (28) with respect to θ is,
r
∂ ln( L (θ )) r ∑x i:n + ( n − r )t 0
=− + i =1
, (29)
∂θ θ θ
Then if we set (31) equal to zero, we have:
r
∑x i:n + ( n − r )t 0
θˆMLE = i =1
. (30)
r
wheret0 is the end time of observation (t=24) and r is the number of failures.
While setting the manufacturing system with the exact value of θj obtained from
the mathematical model is difficult due to mechanical specifications’ changes
during the manufacturing process, we propose a confidence interval for θ . Here,
2rθˆ
using ≈ χ 2 (2r ) as a pivot (2r is degree of freedom), we can set a
θ
2rθˆ
confidence interval for θ . To set a confidence interval using ≈ χ 2 ( 2r ) ,
θ
we obtain:
⎡ 2rθˆ ⎤
P ⎢ χ 2 α ( 2r ) < < χ α2 ( 2r ) ⎥ = 1 − α , (31)
⎣ 1− 2 θ 2 ⎦
By inverting (31) and multiplying all fractions by 2rθˆ we obtain:
⎡ ⎤
⎢ 2rθˆ 2 rθˆ ⎥
P⎢ 2 <θ < 2 =1−α , (32)
χ α ( 2r ) χ α (2r ) ⎥
⎢ 1− ⎥
⎣ 2 2 ⎦
where (34) is a two sided ( 1 − α )% confidence interval based on type I censored
data. In our case, we collected data for a special working period. We were
supposed to collect 30 observations but a failure occurred (type I censored data) in
the 20th observation. Therefore, using (32) we can obtain the estimated θ .
235 + (30 − 20 )10

θˆMLE = = 16 .75 .
20
Clearly the estimated θ is different from the one obtained from the
mathematical model. Here, we configure the confidence interval for θ using the
estimated value with a 0.95% confidence level (equation 33).
⎡ 2 × 20 × 16.75 2 × 20 × 16.75 ⎤
P⎢ <θ < ⎥ = 1 − 0.05 ; we obtain
⎣ χ 0.025 ( 40) χ 02.975 ( 40) ⎦
2
[11.3,27.46] with a 95% confidence interval for θ .

The results show that the output produced in the integrated case was higher
than the outputs produced by separate cost and time objective function
considerations. In addition, the proposed model resulted in an increase in the
availability of the manufacturing system through the improvements in the
reliability of the machine types and the shops. These improvements help to
maintain the system at a desirable level of reliability and to prevent sudden
breakdowns.
4.4 Conclusions 53
4.4 Conclusions
AGV systems complement the operation in manufacturing systems by providing
integrated automated material handling that capitalizes on the system’s flexibility.
Previous research considering AGVs systems has focused primarily on complex
control strategies in MHSs. In this study, we focused on the time and cost
measures in an optimization model used to evaluate an MHS with AGVs. The
automated manufacturing system considered in this study has the following
physical characteristics: (1) the manufacturing system is a job shop, and (2)
single-load AGVs perform the material handling job in the shop.
We took into account the reliability of the manufacturing system because of the
need for steadiness and stability in the system. Reliability was included in the
model as a cost function. Furthermore, we considered bi-objective stochastic
programming to optimize the time and cost objectives because of the uncertainties
inherent in the optimization parameters in real-world problems. Finally, we used
perceptron neural networks to transform the bi-objective optimization model into a
single objective model.
The MHS proposed in this study could potentially be extended (or revised) to
improve the effectiveness of a wide variety of decision tools in productivity
improvement. For example, current trends in scheduling systems provide the
production scheduler with powerful tools which can be used to optimize real-time
workloads in various stages of production. The MHS could potentially be
integrated within such a real-time scheduling system. The proposed method could
also be potentially useful for general applications of business process
improvement which strive to improve workflows within and between functional
groups. These approaches often utilize systematic methods to process a large
amount of imprecise and complex information to redesign critical business
processes.
The limitations of this study and the futures research directions can be
summarized as follows:
• Although we used optimization in this study, alternatively, a simulation
method or a heuristic algorithm could be used in lieu of the proposed model.
Buzacott and Yao (1986) presented a comprehensive review of the analytical
models developed for the design and control of automated and flexible
manufacturing systems. They advocated analytical methods over the
simulation models because analytical methods provide a better insight into the
system performance. Ho et al. (2000, p. 490), Lee et al. (2006, p. 1828), Kuo
et al. (2007, p. 1002), and Crombecq et al. (2011, p. 683) have also confirmed
that although computer simulations are often the first choice for modeling
systems and even for optimization purposes, the simulation of complex
systems with multiple input and output parameters can be both prohibitively
expensive and time consuming. On the other hand, heuristics seek good
feasible solutions to optimization problems where the complexity of the
problem or the available time for its solution does not allow exact algorithms.
Although the main measure of success for exact algorithms is time efficiency,
we must often evaluate the quality of solutions when an exact optimum is not
available (Gnanavel Babu et al. 2010).
• Although we did not experience this limitation in our study, the time
complexity of the learning phase in some complex perceptron neural
networks could be high depending on the heuristic used for calculating the
weights and the halting condition. In such complex perceptron neural
networks, a large number of passes may be required. Each pass involves
computation of the outcome for every training set point followed by
modification of the weights.
• Although in this study we usedthe time and cost measures in an optimization
model, performance measures may be categorized on the basis of: time, cost,
quality, and flexibility measures. Consideration of quality and flexibility is a
natural extension of the model proposed in this study.
• Although in this study we considered single-load AGVs, researchers (e.g., Ho
et al. 2004; Lee et al. 1996; Ozden, 1988) have already proven that multiple-
load AGV systems have many advantages over single-load ones. Another
natural extension of this research is using multiple-load AGVs in the
optimization model.
References
Abdelmaguid, T.F., Nassef, A.O., Kamal, B.A., Hassan, M.F.: A hybrid GA/heuristic
approach to the simultaneous scheduling of machines and automated guided vehicles.
Ball, M.O., Lin, F.L.: A reliability model applied to emergency service vehicle location.
Operations Research 41(1), 18–36 (1993)
Beamon, B.M.: Reliability in design of fixed-path material handling systems. University of
Cincinnati Working Paper, Cincinnati (1995)
Beamon, B.M.: Performance, reliability, and performability of material handling systems.
Birge, J., Louveaux, F.: Introduction to Stochastic Programming. Springer, Berlin (1997)
Buzacott, J.A., Yao, D.D.: Flexible manufacturing systems: A review of analytical models.
Management Science 32(7), 890–905 (1986)
Carbone, R.: Public facilities location under stochastic demand. Infor. 12(3), 261–270
(1974)
Charnes, A., Cooper, W.: Chance-constraint programming. Management Science 6(1), 73–
79 (1959)
References 55
Crombecq, K., Laermans, E., Dhaene, T.: Efficient space-filling and non-collapsing
sequential design strategies for simulation-based modeling. European Journal of
Operational Research 214(3), 683–696 (2011)
Deroussi, L., Gourgand, M., Tchernev, N.: A simple metaheuristic approach to the
simultaneous scheduling of machines and automated guided vehicles. International
Journal of Production Research 46(8), 2143–2164 (2008)
Edwards, W.: How to use multiple attribute utility measurement for social decision-making.
IEEE Transactions on Systems Man and Cybernetics 7(5), 326–340 (1977)
Farling, B.E., Mosier, C.T., Mahmoodi, F.: Analysis of automated guided vehicle
configurations in flexible manufacturing systems. International Journal of Production
Research 39(18), 4239–4260 (2001)
Gnanavel Babu, A., Jerald, J., Noorul Haq, A., Muthu Luxmi, V., Vigneswaralu, T.P.:
Scheduling of machines and automated guided vehicles in FMS using differential
evolution. International Journal of Production Research 48(16), 4683–4699 (2010)
Ho, Y.-C., Sieh, P.F.: A machine-to-loop assignment and layout design methodology for
tandem AGV systems with multiple-load vehicles. International Journal of Production
Research 42(4), 801–832 (2004)
Ho, Y.-C., Cassandras, C.G., Chen, C.-H., Dai, L.: Ordinal optimisation and simulation.
Journal of the Operational Research Society 51(4), 490–500 (2000)
Jain, A., Jain, P.K., Chan, F.T.S., Singh, S.: A review on manufacturing flexibility.
Keeney, R.L., Raiffa, H.: Decisions with Multiple Objectives: Preferences and Value
Trade-offs. Wiley, New York (1976)
Khosravi, A., Nahavandi, S., Creighton, D., Atiya, A.F.: Comprehensive review of neural
network-based prediction intervals and new advances. IEEE Transactions on Neural
Networks 22(9), 1341–1356 (2011)
Kleindorfer, P.R., Kunreuther, H.C., Schoemaker, P.J.H.: Decision Sciences: An Integrative
Perspective. Cambridge University Press, New York (1993)
Kohonen, T.: An introduction to neural computing. Artificial Neural Networks 1(1), 3–16
(1988)
Kros, J.F., Lin, M., Brown, M.L.: Effects of the neural network s-Sigmoid function on
KDD in the presence of imprecise data. Computers and Operations Research 33(11),
3136–3149 (2006)
Kuo, Y., Yang, T., Peters, B.A., Chang, I.: Simulation metamodel development using
uniform design and neural networks for automated material handling systems in
semiconductor wafer fabrication. Simulation Modelling Practice and Theory 15(8),
1002–1015 (2007)
Law, A.M., Kelton, W.D.: Simulation Modeling and Analysis, 2nd edn. McGraw-Hill,
Singapore (2000)
Le-Anh, T., De Koster, M.B.M.: A review of design and control of automated guided
vehicle systems. European Journal of Operational Research 171(1), 1–23 (2006)
Lee, L.H., Chew, E.P., Manikam, P.: A general framework on the simulation-based
optimization under fixed computing budget. European Journal of Operational
Research 174(3), 1828–1841 (2006)
Lee, J., Tangjarukij, M., Zhu, Z.: Load selection of automated guided vehicles in flexible
manufacturing systems. International Journal of Production Research 34(12), 3383–
3400 (1996)
Maniya, K.D., Bhatt, M.G.: A multi-attribute selection of automated guided vehicle using
the AHP/M-GRA technique. International Journal of Production Research 49(20),
6107–6124 (2011)
Maxwell, W.L., Muckstadt, J.A.: Design of automated guided vehicle systems. IIE
Transactions 14(2), 114–124 (1982)
McCulloch, W., Pitts, W.: A logical calculus of the ideas immanent in nervous activity.
Bulletin of Mathematical Biophysics 5, 115–133 (1943)
Özcan, U.: Balancing stochastic two-sided assembly lines: A chance-constrained,
piecewise-linear, mixed integer program and a simulated annealing algorithm.
European Journal of Operational Research 205(1), 81–97 (2010)
Ozden: A simulation study of multiple-load carrying automated guided vehicles in a
flexible manufacturing system. International Journal of Production Research 26(8),
1353–1366 (1988)
Pamosoaji, A.K., Cat, P.T., Hong, K.S.: Sliding-mode and proportional-derivative-type
motion control with radial basis function neural network based estimators for wheeled
vehicles. International Journal of Systems Science (2013),
doi:10.1080/00207721.2013.772678
Pöyhönen, M., Hämäläinen, R.P.: On the convergence of multiattribute weighting methods.
Rosenblatt, F.: The perceptron: A probabilistic model for information storage and
organization in the brain. Psychological Review 65(6), 386–408 (1958)
Saaty, T.L.: The Analytic Hierarchy Process. McGraw-Hill, New York (1980)
Sarker, B.R., Gurav, S.S.: Route planning for automated guided vehicles in a manufacturing
facility. International Journal of Production Research 43(21), 4659–4683 (2005)
Sayarshad, H.R., Tavakkoli-Moghaddam, R.: Solving a multi periodic stochastic model of
the rail–car fleet sizing by two-stage optimization formulation. Applied Mathematical
Modelling 34(5), 1164–1174 (2010)
Shiode, S., Drezner, Z.: A competitive facility location problem on a tree network with
stochastic weights. European Journal of Operational Research 149(1), 47–52 (2003)
Smith, D.J.: Reliability, Maintainability and Risk. Butterworth Heinemann, Oxford (1993)
Stewart, W., Golden, B.: Stochastic vehicle routing: A comprehensive approach. European
Journal of Operational Research 14(4), 371–385 (1983)
Ultsch, A., Korus, D., Kleine, T.O.: Integration of Neural Networks and knowledge-based
systems in medicine. In: Wyatt, J.C., Stefanelli, M., Barahona, P. (eds.) AIME 1995.
LNCS, vol. 934, pp. 425–426. Springer, Heidelberg (1995)
Von Winterfeldt, D., Edwards, W.: Decision Analysis and Behavioral Research. Cambridge
University Press, Cambridge (1986)
Chapter 5
Uncertain Optimal Path for AGV System
5.1 Summary
Discrete event systems are characterized by changes in state over time, based on
current state and state transition rules, where each state is separated from its
neighbor by a step rather than a continuum of intermediate infinitesimal states.
Examples of such systems are information systems, operating systems, networking
protocols, banking systems, business processes and telecommunications systems,
and flexible manufacturing systems. Traditional manufacturing has relied on
dedicated mass-production systems to achieve high production volumes at low
costs. As living standards improve and the demands for new consumer goods rise,
manufacturing flexibility gains prominence as a strategic tool for rapidly changing
markets. Flexibility, however, cannot be properly incorporated in the decision-
making process if it is not well defined and measured in a quantitative manner.
Today, manufacturing flexibility remains an elusive notion because of its inherent
complexity and generality, in spite of a large body of published research work.
There exist more than 50 definitions of (Sethi and Sethi, 1990) and six different
approaches for obtaining a quantitative flexibility measure (Gupta and Goyal,
1989). Flexibility in its most rudimentary essence is the ability of a manufacturing
system to respond to changes and uncertainties associated with the production
process (Zelenovic, 1982; Buzacott, 1982; Gerwin, 1982). A comprehensive
classification of eight flexibility types was proposed in Browne et al. (1984).
Resource and system flexibilities were examined in Slack (1987), whereas global
measures for flexible manufacturing systems (FMSs) were defined in Gupta and
Buzacott (1989). Routing flexibility based on information theoretic concepts was
examined in Yao and Pei (1990) and Kumar (1987). Flexibility measures for one
machine, a group of machines, and whole industry were presented in Brill and
Mandelbaum (1989), involving appropriate weights and machine efficiencies in
carrying out sets of tasks.
Flexibility in its most rudimentary sense is the ability of a manufacturing
system to respond to changes and uncertainties associated with the production
process (Miettinen et al., 2010; Kumar and Sridharan, 2009; Das et al., 2009). In

58 5 Uncertain Optimal Path for AGV System
Barad and Sipper (1988), the period needed by a system to recover after a change
was used as the central flexibility measure, whereas a stochastic dynamic
programming model for its assessment was presented in Kulatilaka (1988).
Artificial intelligence (AI) methods seem appropriate in most practical situations
where numerical data are not readily available and linguistic variables are more
amenable to handling imprecise knowledge (Dooner, 1991). The flexibility of
competing systems can be ranked appropriately using an algorithmic approach
(Abdel-Malek and Wolf, 1991) or a decision support system (Suresh, 1991) based
on performance and economic criteria. Also, integer programming methods were
proposed in Chandra and Tombak (1992), and a graphical representation method
of production processes was presented in Kochikar and Narendran (1992).
Manufacturing flexibility is associated with uncertainty in all levels of a firm’s
operation, such as variation in the demand and characteristics of a product or
unanticipated interruptions of the production process because of machine failures.
In addition, human operators or managers use imprecise concepts and vague
notions when they attempt to define or measure flexibility. Fuzzy set theory
(Dubois and Prade, 1980; Zimmermann, 1991), and especially fuzzy logic,
constitute natural frameworks for the representation and manipulation of
uncertainty.Indeed, fuzzy set theory is an algebra of imprecise propositions and
gradual statements such as “machine A is more flexible than machine B because it
is more versatile.” In previous treatments, uncertainty was handled by probability
theory under the assumption that probabilities can be obtained precisely.
Mandelbaum and Buzacott (1990), examining the meaning and use of flexibility in
decision-making processes, admit that for real-world problems with increased
complexity, the existing modeling methods are inadequate to represent reality. For
context-dependent situations where conceptual imprecisions exist, however, as in
the description of machine flexibility itself, fuzzy sets and fuzzy logic appear to be
more appropriate for the definition and analysis of the problem.
Manufacturing costs for products are very crucial in decision making and
strategic planning. And with respect to cost estimation, research and development
departments in the past could only estimate the final product’s total cost.
Moreover, rules of thumb of the engineers are often applied as the cost estimation
benchmarks, making the results controversial in terms of accuracy (Mostafaee et
al., 2010; Eklin et al., 2009; Qian and Ben-Arieh, 2008; Wang, 2007; Verlinden et
al., 2008; Wang, 2007).
Jobshop is a flexible, scalable and intelligent production planning and control
system offering advanced functionality and value in key areas of manufacturing
and assembly. The flexible jobshop problem (FJP) is an extension of the classical
job shop problem allowing for an operation to be processed by any machine from
a given set.
Automated guided vehicles (AGVs) reduce costs of manufacturing and increase

efficiency in a manufacturing system. These trailers can be used to move raw
materials in line to get them ready for manufacturing (Aized, 2009; Hsueh, 2010).
To conceptualize an AGV, it is necessary to understand the fundamentals of
flexible manufacturing systems (FMSs). Rather than using humans to perform
repetitive tasks, a machine is used to perform the task. Each shop performs a
specific task to assist in the manufacturing of a product. Although FMS is fast and
efficient, but it is not cheap as it requires lots of expensive machines to operate. In
the next section, we give a comprehensive description of our proposed problem of
investigation.

Consider a jobshop layout which applies an AGV for material handling. The AGV
carries raw material, semi-produced and final products in batch sizes. Due to
mounting demands, advancing technology, and rising production capacity, the
need for increasingly more shops is mounting over time. The new shops are
expected to have more advanced machines. Therefore, more than one shop with
the same duty are evolved. The difference among shops having the same duty
shows up in the shop's specifications that affect the production cost. As a result,
the system would consider a flexible jobshop model where multi shops of the
same duty exist and each operation can be processed on any type of machine in
any shop. The sequences of jobs are specified and the jobs are assumed to be
independent.
The structure of such a problem would configure a network. In this network,
the nodes are the shops and the arcs are the flow paths of the AGV to each shop.
Shops in each stage are of the same type but have different specifications such as
different machine types and equipments, varied operator proficiencies, different
rates of defect, etc.. Each flow path for the AGV is associated with a time
parameter and also a cost parameter related to each shop. The aim is to find a path
for the AGV minimizing an aggregate time and cost objective. Considering the
variable status of the AGV flow among shops, the time of each flow path of the
AGV is a triangular fuzzy number. In each shop, different machines and operators
are working. Due to unpredictable events during working times a cost may incur.
This cost is inferenced from an expert system via fuzzy logic "IF …. Then……."
rules. Cost parameters of each shop are considered to be 3: (1) Equipment
sensitivity, (2) operator proficiency, and (3) product specifications, each being
specified by one of the three levels of low, moderate, and high. A configuration
for the proposed problem is presented in Figure 1.
Fig. 1 A configuration of the proposed problem
As stated before, time is considered to be a triangular fuzzy number and cost is

inferenced from an expert system. Our decision model is to consider both time and
cost parameters, and thus an integration of the two parameters would be required.
The integration is to have a weighted sum of cost and time as an arc length in the
proposed network. Time is a triangular fuzzy number and cost is considered to be
an indirect triangular fuzzy number. A description of fuzzy logic, numbers and
systems are given in the next section.
5.3 Mathematical Modeling
5.3.1 Fuzzy Logic and Membership Functions

In Zadeh's words, who first introduced the notion of fuzzy logic (Zadeh, 1965),
fuzzy logic is a tool for ‘‘Computing with Words’’. He stated that the main role of
fuzzy logic was to serve as a methodology for computing with words when no
other methodology could attain such purpose (Zadeh, 1996, 1999).
Fuzzy logic accommodates for the emulation of the human reasoning process
and making decisions based on vague or imprecise data. Linguistic terms can
better represent knowledge, experience, and subjective viewpoints of decision
makers in more intuitive ways and natural language formats. Each linguistic term
can be expressed by a fuzzy set. In fuzzy set theory, elements of a set are allowed
to have membership values between zero and one. If we depict this membership
value by α , then α can have any value in the interval [0,1]. Membership values
for a fuzzy set are usually determined by a membership function. Triangular
membership functions are common (Pedrycz, 1994).
5.3 Mathematical Modeling 61
Definition 1. A triangular fuzzy number a can be defined by a triplet

( a1 , a2 , a3 ) . Its conceptual schema and mathematical form are shown by:
⎧ 0, x ≤ a1
⎪
⎪ x − a1 , a1 < x ≤ a2
⎪⎪ a − a
μa ( x ) = ⎨ 2 1 (1)
⎪ a3 − x , a2 < x ≤ a3
⎪ a 3 −a 2
⎪
⎪⎩ 0, a3 < x .
A triangular fuzzy number a in the universe of discourse X that conforms to this

definition is shown in Figure 2.
μa ( x )
0 x
a1 a2 a3
Fig. 2 A triangular fuzzy number a
Definition 2. Assuming that both a = ( a1 , a2 , a3 ) and b = ( b1 , b2 , b3 ) are

triangular numbers, then the basic fuzzy operations are:
a × b = ( a1 × b1 , a2 × b2 , a3 × b3 ) for multiplication,
a + b = ( a1 + b1 , a2 + b 2 , a3 + b3 ) for addition. (2)
In Bastian (2000), different rule determination approaches have been described.

Expert knowledge is the most popular and accepted method, in practice, while
other common practices such as meta-rule techniques and fuzzy classifier systems
which utilize genetic algorithms or artificial neural networks are applicable when
expert knowledge is not available.
Roughly speaking, a fuzzy rule is the implication stated as an If-Then rule in
which the premise and the conclusion are fuzzy sets. The basic components are the
‘‘If’’ part, usually referred to as the antecedent, and the ‘‘Then’’ part, usually
referred to as the consequent (like If antecedent Then consequent). The antecedent
can be composed of a single condition or a set of conditions combined by
conjunction operators like ‘‘AND’’ and ‘‘OR’’.
Once the rules are determined, the next step is to determine the matching
degree of the inputs with respect to the fuzzy rules to perform the inference
process. If there are multiple inputs, a conjunction operator is used to combine the
matching degree of the fuzzy inputs utilizing ‘‘min’’, ‘‘max’’, or ‘‘product’’ (Klir
and Yuan, 1995; Zimmermann, 1996). The two most principal methods of fuzzy
inference are ‘‘clipping’’ and ‘‘scaling’’ methods (Klir and Yuan, 1995;
Zimmermann, 1996), both suppressing the membership functions for the
consequent, depending upon the degree of matching. Once the final outputs of the
fuzzy rules are obtained through fuzzy inference, the outputs are combined into
one single aggregated output. Since there might be more than one rule with a
matching degree greater than zero, then more than one rule may be configured
(Klir and Yuan, 1995; Zimmermann, 1996).
5.3.2 Fuzzy Systems

The fuzzy logic and rule-based reasoning have found various applications in the
control of industrial processes, modeling of complex systems and the development
of fuzzy inference systems (Lin and Chen, 2004; Machacha and Bhattacharya,
2000). Due to their varied usage, fuzzy inference systems are also known as fuzzy
expert systems, fuzzy rule-based systems, fuzzy associative memories or, in short,
fuzzy systems. In general, a fuzzy system is composed of five major blocks:
• Fuzzifier which transforms the crisp inputs into degrees of match with
linguistic values;
• dictionary which defines the membership functions of the fuzzy sets used
in the fuzzy rules;
• rule base which contains fuzzy if-then rules and along with the
dictionary, comprises the knowledge base of the fuzzy system;
• decision maker which performs the inference operations on the rules;
• defuzzifier which converts the fuzzy results of inference into crisp
outputs.
Based on the different trends in forming the major blocks of a fuzzy system and
various kinds of application to which the fuzzy systems are applied, different
types of fuzzy systems have been introduced. Mamdani fuzzy system and TSK
fuzzy system are two types being commonly used in the literature (Klir and Yuan,
1995; Zimmermann, 1996). Each one is being adopted in its own special domain
because of its particular accommodation for the different ways of knowledge

representation. However, other types of fuzzy systems such as Tsukamoto's
(Tsukamoto, 1979), ANFIS (Jang, 1993; Jang and Sun, 1995), etc., have their own
special applications. Next, we will give a brief discussion of the first two types.
• Mamdani Fuzzy System

Mamdani fuzzy system was proposed as a first attempt to control a steam engine
and boiler combination by a set of linguistic control rules obtained from
experienced human operators. Rules in Mamdani fuzzy systems are like these
(Klir and Yuan, 1995; Zimmermann, 1996):
If x1 is A1 AND/OR x2 is A2 Then y is B1,
whereA1, A2 and B1 are fuzzy sets. The fuzzy set acquired from aggregation of
rules’ results will be defuzzified using defuzzification methods like centroid
(center of gravity), max membership, mean-max, and weighted average. The
centroid method is very popular, in which the ‘‘center of mass’’ of the result
provides the crisp value. In this method, the defuzzified value of the fuzzy set A,
d(A), is calculated by
∫ x.μ A ( x) x
d ( A) = X
, (3)
∫ μ A ( x) x
X
where μ A (.) is the membership function of the fuzzy set A (Klir and Yuan, 1995;
Zimmermann, 1996).
For our problem in which various possible conditions of parameters are stated
in forms of fuzzy sets, we utilize the Mamdani fuzzy system, because the fuzzy
rules representing the expert knowledge in Mamdani fuzzy systems would
consider fuzzy sets in their consequences, while in TSK fuzzy systems, the
consequences are expressed as crisp functions. In general, designing a fuzzy
system is composed of the following major steps (Klir and Yuan, 1995):
Step 1. Identifying pertaining input and output variables. Besides, the

meaningful linguistic states along with appropriate fuzzy sets for the variable
ought to be selected.
Step 2. Introducing a fuzzification method for input variables that expresses the
associated measured uncertainty. The purpose of the fuzzification method is to
interpret measurements of input variables which are expressed by real numbers.
Step 3. Formulating pertaining knowledge in terms of fuzzy inference rules.

There are two principal ways in which relevant inference rules can be determined.
One is to elicit them from experienced humans and the other is to obtain them
from empirical data by suitable learning methods, usually with the help of neural
networks.
Step 4. Combining measurements of input variables with relevant fuzzy rules to
inference, regarding the output variables in which the purpose of inference engine
will be obtained.
Step 5. Ascertaining a suitable defuzzification method to convert the
aggregated fuzzy set of implications into a real number.
5.3.3 Modeling the Proposed Fuzzy Problem

As stated before, time is a triangular fuzzy number, ensued from the experts'
knowledge. Brainstorming and expert knowledge versus meta-rule techniques
(neural networks, genetic algorithms, etc.) are two common approaches for
defining fuzzy rules and membership functions. While available empirical data is
requisite for using the second approach, due to unavailability of historical data for
cost, we make use of the first approach to obtain the membership functions and
fuzzy rules.
The input to our Mamdani type fuzzy system is composed of equipment
sensitivity, operator proficiency, and product specifications. For any of the inputs,
three linguistic terms of "low", "moderate", and "high" are defined. The output of
the system is cost that is identified by any of the three linguistic terms, "low",
"moderate", or "high". The maximum membership grade of linguistic term
‘‘high’’ is 30.
As a result, a triangular fuzzy number as the time and a numerical value as the
cost are obtained. We intend to consider an integrated time-cost value as the value
of each proposed arc in the network. Time and cost having different scales, it
would not be possible to perform basic operations such as addition on their
original forms. Thus, to remove the scales, we normalize the time and cost values
throughout the network and then consider a weighing method to make them
appropriate for basic operations. On the other hand, time is a triangular fuzzy
number and cost is a crisp numerical value. But, as implied by Definition 2, we
need to have triangular numbers for both parameters to perform the basic
operations. For this, we consider cost as a trivial triangular fuzzy number and
show it by a triplet (0,C,0), where C is the numerical value inferenced from the
expert system for cost with the membership value equal to 1. The right and left
sides of C are zero since their membership values are zero. In the next section, we
present an approach for weighing the parameters.
5.3.4 Operations on the Parameters of the Problem:

Normalization of the Parameters
As stated, to perform basic operations on the time and cost parameters with their
different scales, we need to remove their scales. To do this, two normalization
processes are proposed for time and cost, separately. In (4), (5) and (6) below, i is
an index to show the node and j is an index to show the stage (shop type) in the
proposed network, rijis the value of either time or cost parameter in each node, and
nijis its corresponding normalized value. Considering time as a triangular fuzzy
number, the normal value with a positive view is given by
rij − r jmin
nij = , ∀i, j, (4)
r jmax − r jmin
while the normal value with a negative view is given by,

r jmax − rij
nij = , ∀i, j , (5)
r jmax − r jmin
max
where r j and rjmin are maximum and minimum values in each column of an
assumed matrix of time or cost, respectively. Since time in our proposed model is
a criterion implicating a negative aspect in decision making, then we choose (5)
for normalizing time in our approach.
Assuming cost as a crisp value, we normalize the cost values as follows:
rij
nij = , ∀i , j , (6)
∑r
k
2
kj
To compute the minimum or maximum value in (4) and (5), comparisons are
needed to be made. This means that it is necessary to have a method for ranking
and comparing fuzzy numbers. An operator ≺ for ordering fuzzy numbers can be
defined as follows (see Mahdavi et al., 2009):
A ≺ B ⇔ (a1 ≤ b1 ) ∧ (a2 ≤ b2 ) ∧ (a3 ≤ b3 ) ∧ (a4 ≤ b4 ). (7)
However, this relation is not a complete ordering, as fuzzy numbers A and B

satisfying
∃i, j ∈{1, 2,3, 4}) : (ai < bi ) ∧ (ai > bi ) , (8)
would not be comparable by ≺.

The ranking or ordering methods for fuzzy quantities have been proposed by
several authors. For summaries, see Bortolan and Degani (1985) and Delgado et
al. (1988). Admittedly, none of these methods is commonly accepted. Various
ranking functions may produce conflicting and controversial results in comparison
of fuzzy numbers (these issues have been discussed in Wang and Kerre, 2001,
through a specific example).
Here, we use a fuzzy ranking method recently adopted by Mahdavi et al.
(2009). Consider the fuzzy min and max operations defined in analogy with the
fuzzy addition as follows:
~
MINV = Min value (~
a, b) = (min(a1 , b1 ), min(a2 , b2 ), min(a3 , b3 ), min(a4 , b4 )). (9)
~
MAXV = Max value(~a, b) = (max(a1 , b1 ), max(a2 , b2 ), max(a3 , b3 ), max(a4 , b4 )).
~
~ and b , the fuzzy min
It is evident that, for non-comparable fuzzy numbers a
operation results in a fuzzy number different from both of them. For example, for
a = (5,10,13,19) and b = (6,9,15, 20) , we get from (9) a fuzzy
MINV = (5, 9, 13, 19 ) which differs from a and b . To alleviate this drawback,
a method based on the distance between fuzzy numbers is proposed. We use the
distance function introduced in Sadeghpour Gildeh and Gien (2001). The main
advantages of this distance function, to be defined next, are the generality of its
usage on various fuzzy numbers, and its reliability in distinguishing unequal fuzzy
numbers.
Definition 3. The D p ,q -distance, indexed by parameters 1 < p < ∞ and
0 < q < 1 , between two fuzzy numbers a and b is a nonnegative function

given by:
⎧ 1
⎡ 1
− − p
1
⎪ (1 − q ) ∫ aα − bα dα + q ∫ aα − bα dα p ,
+ + p ⎤ p < ∞,
~ ⎪ ⎢⎣ 0 0 ⎥⎦ (10)
D p , q (a~ , b ) = ⎨
⎪
( −
) −
⎪(1 − q) sup aα − bα + q 0inf
⎩ 0 ≤α ≤1 ≤α ≤
(
1
)
aα+ − bα+ , p = ∞,
+ −
where aα , aα and bα+ , bα− are the corresponding right and left α -cuts of a and
b, respectively. The analytical properties of D p ,q depend on the first parameter
p, while the second parameter q of D p ,q characterizes the subjective weight
attributed to the end points of the support; that is, aα+ and aα− . of the fuzzy
numbers. If there is no reason for distinguishing any side of the fuzzy numbers,
then D 1 is recommended. Having q close to 1 results in considering the right
p,
2
side of the support of the fuzzy numbers more favorably. Since the significance of
the end points of the support of the fuzzy numbers is assumed to be the same, then
1
we consider q = .
2
For triangular fuzzy numbers a = ( a1 , a2 , a3 ) and b = (b1 , b 2 , b3 ) , the
1
above distance with p = 2 and q = is then calculated as:
2
~, b~ ) = 1 ⎡ (b − a ) 2 + (b − a ) 2 + (b − a )(b − a ) ⎤
3 2
D
2,
1 ( a ∑ i i
6 ⎢⎣ i =1
2 2 ∑
i =1
i i i +1 i +1 ⎥
⎦
(11)
2
a , b and MINV using

We are now able to compute the distance between
~
~ and b as given before, we get
the proposed distance function (11); for a
~
D 1 (a~, MINV ) = 0.1667 and D 1 (b , MINV ) = 1.33 . The number with
2, 2,
2 2
lowest distance is the minimum. Now, we are ready to propose the following
algorithm for the comparison of fuzzy numbers.
Algorithm 1 (compare two triangular fuzzy numbers).

~ ~
Input: Two fuzzy numbers ( L1 and L2 ).
min
Output: Minimum or maximum between the two triangular fuzzy numbers ( L
~
or L max )
Step 1: Compute the minimum or maximum value ( MINV or MAXV ) by
Equation (9).
~ ~
Step 2: Find the distance of MINV or MAXV from ( L1 and L2 ) using
Equation (11).
~
Step 3: Determine Lmin or L max giving the smallest or largest distance.
5.3.5 Weighing the Parameters

To weigh the parameters, we take a multi criteria decision making approach.
Multi-criteria decision making (MCDM), dealing primarily with problems of
evaluation or selection (Keeney and Raiffa, 1976; Teng, 2002), is a rapidly
developing area in operatios research and management science. The analytical
hierarchy process (AHP), developed by Saaty (1980), is a technique of
considering data or information for a decision in a systematic manner
(Schniederjans and Garvin, 1997). AHP is mainly concerned with the way to solve
decision problems with uncertainties in multiple criteria characterization. It is
based on three principles: (1) constructing the hierarchy, (2) priority setting, and
(3) logical consistency. We apply AHP to weigh the parameters.
5.3.6 Construction of the Hierarchy

A complicated decision problem, composed of various attributes of an objective,
is structured and decomposed into sub-problems (sub-objectives, criteria,
alternatives, etc.), within a hierarchy.
5.3.7 Priority Setting

The relative “priority” given to each element in the hierarchy is determined by
pair-wise comparisons of the contributions of elements at a lower level in terms of
the criteria (or elements) with a causal relationship. In AHP, multiple paired
comparisons are based on a standardized comparison scale of nine levels (see
Table 1 due to Saaty, 1980).
Table 1 Scale of relative importance

Intensity of importance Definition of importance
1 Equal
2 Weak
3 Moderate
4 Moderate plus
5 Strong
6 Strong plus
7 Very strong or demonstrated
8 Very, very strong
9 Extreme
Let C = {c1 ,...,cn } be the set of criteria. The result of the pair-wise
comparisons on n criteria can be summarized in an n × n evaluation matrix A in
which every element aij is the quotient of weights of the criteria, as shown below:
A = (aij), i, j = 1, . . . , n. (12)
The relative priorities are given by the eigenvector (w) corresponding to the
largest eigenvalue
( λmax ) as:
Aw = λmaxw . (13)
When pair-wise comparisons are completely consistent, the matrix A has rank 1
and λmax = n . In that case, weights can be obtained by normalizing any of the
rows or columns of A.
The procedure described above is repeated for all subsystems in the hierarchy.
In order to synthesize the various priority vectors, these vectors are weighed with
the global priority of the parent criteria and synthesized. This process starts at the
top of the hierarchy. As a result, the overall relative priorities to be given to
the lowest level elements are obtained. These overall, relative priorities indicate
the degree to which the alternatives contribute to the objective. These priorities
represent a synthesis of the local priorities, and reflect an evaluation process that
permits integration of the perspectives of the various stakeholders involved.
5.3.8 Consistency Check

A measure of consistency of the given pair-wise comparison is needed. The
consistency is defined by the relation between the entries of A; that is, we say A is
consistent if aik= aij · ajk, for all i,j,k. The consistency index (CI) is:
(λmax − n)
CI = . (14)
(n − 1)
The final consistency ratio (CR), on the basis of which one can conclude whether
the evaluations are sufficiently consistent, is calculated to be the ratio of the CI
and the random consistency index (RI):
CI
CR = . (15)
RI
The value 0.1 is the accepted upper limit for CR. If the final consistency ratio
exceeds this value, the evaluation procedure needs to be repeated to improve
consistency. The measurement of consistency can be used to evaluate the
consistency of decision makers as well as the consistency of all the hierarchies.
We are now ready to give an algorithm for computing parameter weights using
the AHP. The following notations are used.
Notations and definitions:

n: Number of criteria.
i: Number of parameters.
p: Index for parameters, p=1or 2.
d: Index for criteria, 1≤ d ≤ D.
R pd : The weight of pth item with respect to dth criterion.
wd : The weight of dth criterion.
Algorithm 2: PWAHP (compute parameter weights using the AHP).

Step 1: Define the decision problem and the goal.
Step 2: Structure the hierarchy from the top through the intermediate to the lowest
level.
Step 3: Construct the parameter-criteria matrix using steps 4 to 8 using the AHP.
{Steps 4 to 6 are performed for all levels in the hierarchy.}
Step 4: Construct pair-wise comparison matrices for each of the lower levels for
each element in the level immediately above by using a relative scale
measurement. The decision maker has the option of expressing his or her intensity
of preference on a nine-point scale. If two criteria are of equal importance, a value
of 1 is set for the corresponding component in the comparison matrix, while a 9
indicates an absolute importance of one criterion over the other (Table 1 shows the
measurement scale defined by Saaty, 1980).
Step 5: Compute the largest eigenvalue by the relative weights of the criteria and
the sum taken over all weighted eigenvector entries corresponding to those in the
next lower level of the hierarchy.
Analyze pair-wise comparison data using the eigenvalue technique. Using these
pair-wise comparisons, estimate the parameters. The eigenvector of the largest
eigenvalue of matrix A constitutes the estimation of relative importance of the
attributes.
Step 6:Construct the consistency check and perform consequence weights

analysis as follows:
⎡ 1 w1
…
w1 ⎤
⎢ w2 wn⎥
⎢ ⎥
⎢w 2 1 …
w2 ⎥
A = ( aij ) = ⎢ w 1 wn⎥.
⎢ ⎥
⎢ ⎥
⎢w n wn
1 ⎥
⎣⎢ w 1 w2 ⎥⎦
Note that if the matrix A is consistent (that is, aik= aij · ajk, for all
i , j , k = 1, 2, ..., n ), then we have (the weights are already known),
wi
a ij = , i, j = 1,2,..., n.
wj
If the pair-wise comparisons do not include any inconsistencies, then λmax = n.
The more consistent the comparisons are, the closer the value of computed λmax
is to n. Set the consistency index (CI), which measures the inconsistencies of pair-
wise comparisons, to be:
CI =
( λmax − n )
,
( n − 1)
and let the consistency ratio (CR) be:
⎛ CI ⎞
CR = 100 ⎜ ⎟,
⎝ RI ⎠
wheren is the number of columns in A and RI is the random index, being the
average of the CI obtained from a large number of randomly generated matrices.
Note that RI depends on the order of the matrix, and a CR value of 10% or less
is considered acceptable (Saaty, 1980).
Step 7: Form the parameter-criteria matrix as specified in Table 2:
Table 2 The parameter-criteria matrix.
C1 C2 … Cd
parameter 1 R11 R12 … R1d
parameter 2 R21 R22 … R2d
Step 8: As a result, configure the pair-wise comparison for criteria-criteria matrix

as in Table 3:
Table 3 The criteria-criteria pair-wise comparison matrix.
C1 C2 … Cd wd
Criteria 1 1 a12 … a1d w1
Criteria 2 1/a12 1 … a2d w2
Criteria d 1/a1d 1/a2d … 1 wd
The wd are gained by a normalization process. The wd are the weights for
criteria.
Step 9: Compute the overall weights for the parameters, using tables 2 and 3, as
follows:
ψ = Total weight for parameter 1 = R11 × w1 + R12 × w2 + ... + R1d × wd ,

(16)
ψ ' = Total weight for parameter 2 = R21 × w1 + R22 × w2 + ... + R2 d × wd ,
where ψ ' = 1 − ψ . Note that due to the significance and sensitivity of different
shops of the same type in any stage the weighing process is performed for each
arc. Here, we obtain the weights for the parameters of each arc in the proposed
n
network. If we consider C n and T as normalized cost and time, respectively,
then the total weighted normalized value of each arc is determined as follows:
P= (ψ × T n ) + (ψ '×C n ) . (17)
In the next section, we propose an approach to identify the shortest path in the
network using the total weighted normalized value for each arc. Before that,
consider a ranking method for triangular fuzzy numbers.
5.3.9 Ranking of Triangular Fuzzy Numbers

Several fuzzy ranking methods have been proposed (Bortalan and Degani, 1985;
Luis and Antonio, 1989; Kim and Park, 1990). Since the graded mean integration
representation method (Chen and Hseih, 2000) not only alleviates some
drawbacks of the existing methods, but also possesses the advantages of being
easily implementable and quite effective in problem solving (see Lee et al., 2007),
we will use it to transform the total weighted normalized value of each arc in our
proposed network. If a = ( a1 , a2 , a3 ) is a triangular fuzzy number, then the

~ ) ) is defined to be:
graded mean integration ( R (a
a + 4a 2 + a 3
R ( a~ ) = 1 . (18)
6
We apply (18) to transform the fuzzy numbers to crisp values and use them to
find the optimal path.
5.3.10 Shortest Path in a Network

Let G = (V,A) be a graph, where V = {1,..., N } is the set of nodes, and A ⊂ V × V is
the set of arcs. We write (i , j ) ∈ A , if there exists an arc from node i ∈V to node
j ∈ V . Furthermore, let t ij ≥ 0 denote the distance (or travel time, or any other
measure of cost) from i to j. If (i , j ) ∉ A ,then set t ij = +∞ . Note that the travel
time from node i to node j is assumed to be stationary; i.e., independent of the
actual arrival time at node i. Let fijdenote the length of the shortest-path from i to j
in the graph.
It is a well-known principle that every additive deterministic dynamic
programming formulation can be equivalently viewed as the problem of finding
the shortest path in a directed network, where the states, decisions, and decision
costs of the former correspond to the nodes, arcs, and arc lengths of the latter
(Dreyfus and Law, 1977). It is perhaps for this reason that the task of efficiently
computing shortest paths is found prominence in the mathematical programming
literature. Next, we describe a dynamic programming approach for computing the
optimal path.
Dynamic programming (DP) was introduced by Bellman (1957). Toth (1980)
presented the early DP-based approaches and reported numerical experiments with
a limited success. Hybrid methods, combining dynamic programming and implicit
enumeration, were developed later. The first approach was developed by Plateau
and Elkihel (1985). A recent approach, the so-called combo algorithm, is able to
solve very large instances of up to 10000 variables within less than one second,
with basically no difference in the required solution times for ‘‘easy’’ and ‘‘hard’’
instances (Martello et al., 1999). Marsten and Morin (1978) proposed the first
hybrid method, which combined heuristic algorithms, dynamic programming and
branch-and-bound approaches. More sophisticated methods can be found in
Ibaraki (1987).
Dynamic programming is a technique to tackle multistage decision processes.
A given problem is subdivided into smaller subproblems, which are sequentially
solved until the initial problem is solved by the aggregation of the subproblem
solutions. In each stage, a set of states is defined. The states would describe all
possible conditions of the process in the current decision stage, which corresponds
to every feasible partial solution. The set of all possible states is known as the state
space. The states of a stage u can be transformed to states of a stage u+1, using a
transition. A transition indicates the decisions adopted in a stage, and a sequence
of transitions taken to reach a state starting from another state is known as a
policy. Dynamic programming approaches can be seen as transformations of the
original problem to one associated with the exploration of a multistage graph
G(S,T), where the vertices in S correspond to the state space and the arcs in T
correspond to the set of transitions, leading to an optimal policy.
The basis of dynamic programming can be traced to the optimality principle of
Bellman (2003). The optimality principle states that an optimal policy should be
constituted by optimal policies from every state of the decision chain to the final
state.
Here, we make use of a dynamic programming approach for our proposed
network to identify the optimal manufacturing path. This model helps the
manufacturing system to determine the more profitable shops. The advantages of
such a model are simplicity, the ability to determine the exact optimal value, and
implementability on sophisticated networks. The backward dynamic model would
be defined as:
Indices:
s Number of stages; s= 0,1,2,…, n.
i' Start node number; i'=1,2, …, I; i'=0 (for the start node).
i End node number; i=1,2, …, I.
Notations:
ϕs (i' ) The minimum value of moving from node i' in stage s to an end
node i in stage s-1.
Pi'i Numerical value of an arc between node i' to node i.
Objective function:
ϕ s (i ' ) = Min {ϕ s +1 (i ) + Pi 'i }, ∀i ' in stage s, ( s = 0,1,2,..., n).

i in layer s +1
(19)
ϕ s (i ) = (0),
ϕ * = ϕ 0 (0).
Note that ϕ * identifies the optimal path.

5.4 Conclusions 75
5.4 Conclusions
We proposed an approach for finding an optimal path in a flexible jobshop

manufacturing system considering two criteria of time and cost. The proposed
flexible jobshop system has more than one shop with the same duty. The
difference among shops with the same duty is in their machines with various
specifications. The shops configure a network in which they are considered as
nodes and the paths among them are considered as network arcs. An Automated
Guided Vehicle (AGV) functions as a material handling device through the
manufacturing network. Time is considered to be a triangular fuzzy number and
cost is inferred from an expert system considering three parameters of equipment
sensitivity, operator proficiency, and product specification via linguistic variables.
As unique contributions of the work, the objective was to find a path which
minimizes both the time and cost criteria, aggregately. Since time and cost had
different scales, a normalization process was used to remove the scales and
because the model was bi-objective, the AHP weighing method was applied to
gain a single objective. A dynamic programming approach was used to compute a
shortest path in the proposed network. Finally, the proposed approach was
illustrated by a numerical example. Various applicability and advantages of the
proposed model were discussed. The limitations of the research were the work
load of the AGVs and the defect rate of the guide path. As future research, other
multi-objective optimization approaches could be tested and compared with the
proposed method.
References
Abdel-Malek, L., Wolf, C.: Evaluating flexibility of alternative FMS designs-A
comparative measure. Int. J. Prod. Econ. 23(1), 3–10 (1991)
Industrial Engineering 57(3), 822–831 (2009)
Barad, M., Sipper, D.: Flexibility in manufacturing systems: Definitions and Petri net
modeling. Int. J. Prod. Res. 26(2), 237–248 (1988)
Bastian, A.: Identifying fuzzy models utilizing genetic programming. Fuzzy Sets and
Systems 113(3), 333–350 (2000)
Bellman, R.: Dynamic Programming. Princeton University Press, Princeton (1957)
Bellman, R.: Dynamic Programming. Dover Publications (2003)
Bortolan, G., Degani, R.: A review of some method for ranking fuzzy subsets. Fuzzy Sets
and Systems 15, 1–19 (1985)
Brill, P.H., Mandelbaum, M.: On measures of flexibility in manufacturing systems. Int. J.
Prod. Res. 27(5), 747–756 (1989)
Browne, J., Dubois, D., Rathmill, K., Sethi, S.P., Stecke, K.E.: Classification of flexible
manufacturing systems. FMS Mag. 2(2), 114–117 (1984)
Buzacott, J.A.: The fundamental principles of flexibility in manufacturing system. In: Proc.
1st. Int. Conf. on FMS, pp. 23–30 (1982)
Chandra, P., Tombak, M.M.: Models for the evaluation of routing and machine flexibility.
Eur. J. Oper. Res. 60, 156–165 (1992)
Chen, S.H., Hsieh, C.H.: Representation, ranking, distance, and similarity of L-R type
fuzzy number and application. Aust. J. Intell. Process. Syst. 6(4), 217–229 (2000)
Das, K., Baki, M.F., Li, X.: Optimization of operation and changeover time for production
planning and scheduling in a flexible manufacturing system. Computers & Industrial
Engineering 56(1), 283–293 (2009)
Delgado, M., Verdegay, J.L., Vila, M.A.: A procedure for ranking fuzzy numbers using
fuzzy relations. Fuzzy Sets and Systems 26, 49–62 (1988)
Dooner, M.: Conceptual modeling of manufacturing flexibility. Int. J. Comput. Integr.
Manuf. 4(3), 135–144 (1991)
Dreyfus, S.E., Law, A.M.: The Art and Theory of Dynamic Programming. Academic Press,
New York (1977)
Dubois, D., Prade, H.: Fuzzy Sets and Systems: Theory and Applications. Academic Press,
New York (1980)
Eklin, M., Arzi, Y., Shtub, A.: Model for cost estimation in a finite-capacity stochastic
environment based on shop floor optimization combined with simulation. European
Engineering 56(3), 779–808 (2009)
Gerwin, D.: Do’s and don’t’s of computerized manufacturing. Harvard Bus. Rev. 60(2),
107–116 (1982)
Gupta, D., Buzacott, J.A.: A framework for understanding flexibility of manufacturing
systems. J. Manuf. Syst. 8(2), 89–97 (1989)
Gupta, Y., Goyal, S.: Flexibility of manufacturing systems: Concepts and measurements.
Eur. J. Oper. Res. 43, 119–135 (1989)
Hsueh, C.F.: A simulation study of a bi-directional load-exchangeable automated guided
vehicle system. Computers & Industrial Engineering 58(4), 594–601 (2010)
Ibaraki, T.: Enumerative approaches to combinatorial optimization–Part II. Annals of
Operations Research 11, 397–439 (1987)
Jang, J.R.: ANFIS: Adaptive-network-based fuzzy inference system. IEEE Transactions on
Systems, Man and Cybernetics 23, 665–685 (1993)
Jang, J.R., Sun, C.: Neuro-fuzzy modeling and control. Proceedings of IEEE 83, 378–405
(1995)
Keeney, R., Raiffa, H.: Decision with multiple objectives: Preference and value tradeoffs.
Wiley and Sons, New York (1976)
Kim, K., Park, K.S.: Ranking fuzzy numbers with index of optimism. Fuzzy Sets and
Systems 35, 143–150 (1990)
Klir, G.J., Yuan, B.: Fuzzy Sets and Fuzzy Logic: Theory and Applications. Prentice-Hall
(1995)
Kochikar, V.P., Narendran, T.T.: A framework for assessing the flexibility of
manufacturing systems. Int. J. Prod. Res. 30(12), 2873–2895 (1992)
Kulatilaka, N.: Valuing the flexibility of flexible manufacturing systems. IEEE Trans. Eng.
Manag. 78, 250–257 (1988)
Kumar, N.S., Sridharan, R.: Simulation modelling and analysis of part and tool flow control
decisions in a flexible manufacturing system. Robotics and Computer-Integrated
Manufacturing 25(4-5), 829–838 (2009)
References 77
Kumar, V.: Entropic measures of manufacturing flexibility. Int. J. Prod. Res. 25(7), 957–
966 (1987)
Lee, K.L., Huang, W.C., Teng, J.Y.: Locating the competitive relation of global logistics
hub using quantitative SWOT analytical method. Quality & Quantity 43(1), 87–107
(2007)
Lin, C., Chen, C.: New product Go/No-Go evaluation at the front end: A Fuzzy linguistic
approach. IEEE Transactions on Engineering Management 51(2), 197–207 (2004)
Luis, M., Antonio, G.M.: A subjective approach for ranking fuzzy numbers. Fuzzy Sets and
Systems 29, 145–153 (1989)
Machacha, L., Bhattacharya, P.: A Fuzzy-logic-based approach to project selection. IEEE
Transactions on Engineering Management 47(1), 65–73 (2000)
Mahdavi, I., Nourifar, R., Heidarzade, A., Mahdavi-Amiri, N.: A dynamic programming
approach for finding shortest chains in a fuzzy network. Applied Soft Computing 9(2),
503–511 (2009)
Mandelbaum, M., Buzacott, J.: Flexibility and decision making. Eur. J. Oper. Res. 44, 17–
27 (1990)
Marsten, R.E., Morin, T.L.: A hybrid approach to discrete mathematical programming.
Mathematical Programming 14, 21–40 (1978)
Martello, S., Pisinger, D., Toth, P.: New trends in exact algorithms for the 0–1 knapsack
problem. European Journal of Operational Research 123, 325–336 (1999)
Miettinen, K., Eskelinen, P., Ruiz, F., Luque, M.: NAUTILUS method: An interactive
technique in multiobjective optimization based on the nadir point. European Journal of
Operational Research 206 (2) 2010, 426–434 (2010)
Mostafaee, A., Saljooghi, F.H.: Cost efficiency measures in data envelopment analysis with
data uncertainty. European Journal of Operational Research 202(2), 595–603 (2010)
Niazi, A., Dai, J.S., Balabani, S., Seneviratne, L.: Product cost estimation: Technique
classification and methodology review. Journal of Manufacturing Science and
Engineering 128, 563–575 (2006)
Pedrycz, W.: Why triangular membership functions? Fuzzy Sets and Systems 64, 21–30
(1994)
Plateau, G., Elkihel, M.: A hybrid method for the 0–1 knapsack problem. Methods of
Sadeghpour Gildeh, B., Gien, D.: La distance-Dp,q et le cofficient de corrélation entre deux
variables aléatoires floues. In: Actes de LFA 2001, Monse, Belgium, pp. 97–102 (2001)
Schniederjans, M.J., Garvin, T.: Using the analytic hierarchy process and multi-objective
programming for the selection of cost drivers in activity based costing. European
Journal of Operational Research 100, 72–80 (1997)
Chapter 6
Cross Entropy Model for AGV Routing Time
6.1 Summary
Today, in order to survive in the rapid challenging environment of the modern

manufacturing era, manufacturers are forced to adopt new technologies, especially
for products that are made in small batch production. Automated Guided Vehicles
(AGV) has been applied for the automated manufacturing system. In industrial
application, manufacturing factory is brought the mobile vehicle to incorporate
working with other machine in order to being the automated manufacturing
system (Arai et al., 2002).Advanced automated manufacturing systems are widely
used in industrial companies where productivity objectives have to be met. These
systems often being costly, they must be designed to be as efficient as possible.
Here, an automated manufacturing system in a job shop layout considering AGV
as a material handling resource is focused. The key issue in manufacturing
operations is how to produce high quality products at low costs in such a way that
the diversified demand is met. Hence, modern manufacturing companies should
become as responsive as possible in order to satisfy customer demands. Material
handling accounts for 30–75% of the total cost of a product, and efficient material
handling can result in reducing the manufacturing system operations cost by 15–
30% (Sule, 1994). These points underscore the importance of material handling
costs reduction as a key element in improving the cost structure of a product. The
determination of a material handling system involves both the selection of suitable
material handling equipment and the assignment of material handling operations
to each individual piece of equipment. Hence, material handling system selection
can be defined as the selection of material handling equipment to perform material
handling operations within a working area considering all aspects of the products
to be handled.The material handling system plays a crucial role in automated
manufacturing systems. When inadequately designed, the material handling
system indeed can adversely affect the overall performance of the system and lead
to substantial losses in productivity and competitiveness, and to unacceptably long
lead times. Thus, to avoid such pitfalls, material handling system design must be
integrated into the overall design of the manufacturing system centering on the
selection of machines and the allocation of operations to the machines (Butdee and

80 6 Cross Entropy Model for AGV Routing Time
Suebsomran, 2006; Shirazi et al., 2010).Automated guided vehicle is an intelligent

machine that has ‘intelligence’ to determine its motion status according to the
environmental conditions systems. AGVs are advanced material handling devices
extensively used in automated manufacturing systems (AMS) to transport
materials among workstations (Vis, 2006).
Automated manufacturing systems (AMSs), equipped with several CNC
machines and an AGV-based material handling system, are designed and
implemented to secure automation and gain efficiency of production. To achieve
these benefits, the planning in the AMS decision making process is critical
because the planning decision has an influence on the subsequent decision
processes such as scheduling, dispatching, etc. Subtasks are classified into two
types for material processing and delivery. To pursue a global optimization of
material processing and transportation cost, a manufacturing plan should consider
the two costs simultaneously. Moreover, to obtain a better solution for subsequent
decision making processes, not only total workloads are evenly distributed among
the workstations, but workloads are well balanced among AGV’s network
segments. Many factories were adopted it into assembly line or production line
such as automobile, food processing, wood working, port container terminal, and
other factories.
The cross-entropy (CE) method is a general Monte Carlo approach to
combinatorial and continuous multi-extremal optimization(De Boer et al., 2005).
The method originated from the field of rare event simulation (Rubinstein, 1997),
where very small probabilities need to be accurately estimated, for example in
network reliability analysis, queuing models, or performance analysis of
telecommunication systems. The CE method can be applied to static and noisy
combinatorial optimization problems such as the traveling salesman problem, the
quadratic assignment problem, DNA sequence alignment, the max-cut problem
and the buffer allocation problem, as well as continuous global optimization
problems with many local extrema (Rubinstein, 1999).
In a nutshell the CE method consists of two phases:

 Generate a random data sample (trajectories, vectors, etc.) according to a
specified mechanism.
 Update the parameters of the random mechanism based on the data to
produce a "better" sample in the next iteration.
Here, we aim to obtain the total material handling time for an AGV while the
AGV activity times are stochastic. Due to importance of material handling in
manufacturing systems specially AMS, estimating the material handling time is
substantial for planning purposes to cover demands considering the capability and
availability of the machines and equipment.
Automated Guided Vehicle has firstly developed and conducted as a research
by (Butdee and Suebsomran, 2006 and 2007; Butdee et al., 2006) in the attempt to
be used at Jumbo Truck Manufacturing in Thailand. An AGV need to sense its
environment to be able to plan its operations and then act based on this plan.
Thus, an AGV system possesses more flexibility and capacity then other
6.1 Summary 81
conventional material-handling systems and plays an important role in AMS

(Reveliotis, 2000).
An automated guided vehicle system enables flexible material routing and
dispatching, and is especially suited for flexible manufacturing environments in
which product mix and priorities may continuously vary (Dubois, 1983). An AGV
based material handling system is designed and implemented to gain production
the flexibility and efficiency (Co and Wysk, 1986; Suri and Hildebrant, 1984;
Dallery and David, 1984). There are four basic AGV material flow configurations:
conventional network, unidirectional loop network, tandem configuration and
segmented loop topology (Asef-Vaziri and Laporte, 2005). For tandem layout we
can point to a recent work by Shirazi et al., (2010) where they illustrated a non-
linear multi-objective problem for minimizing the material flow intra and inter-
loops and minimization of maximum amount of inter cell flow, considering the
limitation of tandem automated guided vehicle (TAGV) work-loading.
The use of AGV has increased considerably since their introduction in the
1950s, along with the number of AGV application areas and types (Vis, 2006). As
the complexity and size of AGVS applications increase, evaluating performance
becomes more difficult. The last decade has also seen great progress in control
methodologies of autonomous multi-robot systems (Buzacott, 1999). The
objective of optimization model is the minimization of the total distance traveled
by vehicles to transport the material handling system. The route planning of AGVs
in FMS was proposed by (Co and Wysk, 1986; Suri and Hildebrant, 1984; Dallery
and David, 1984). Authors in (Co and Wysk, 1986) presented the new approach
for dynamic route planning and scheduling problem of AGVs. They applied the
search algorithm and some heuristic rules to solve the route assignment in
dynamic situations. Suri and Hildebrant (1984) also proposed the path planning
strategy of AGV to navigation, collision avoidance and docking to the target. The
path planning was implemented on-board computer in order to avoid the wire-
guided path. Not only the AGV was moved along the path with collision
avoidance, but also it should be navigated with no deadlock condition as done by
Srinivasan et al. (1994).
For an AMS, it is desirable to be able to increase or decrease the product with the
rise and fall of demand. Therefore, Fazlollahtabar et al. (2010) proposed a flexible
jobshop automated manufacturing system to optimize the material flow. The flexibility
was on the multishops of the same type and also multiple products that can be
produced. An AGV was applied for material handling. The objective was to optimize
the material flow regarding the demand fluctuations and machine specifications.
Application of cross entropy (CE) method to rare event simulation has been
studied to obtain an accurate estimation of very small probabilities. CE mainly uses
various importance sampling schemes to estimate rare event probabilities in
Markovian systems such as Markovian reliability models and Jackson
networks.Kaynar and Ridder (2010), present a general state-dependent importance
sampling method which partitions the state space and applies the cross entropy
method to each partition. They investigated two versions of their proposed algorithm
and employed them in several examples of reliability and queuing models.
We consider a jobshop manufacturing system make use of AGV for material

handling purpose as shown in Figure 1. Since uncertainty is elusive in automated
manufacturing systems, the activities' times are stochastic and follow statistical
distributions. The statistical distribution of each AGV activity is considered for
obtaining the total material handling time. The activities that the AGVs perform in
material handling process are such as direct movement, waiting, pickup/delivery,
breakdown, etc. These activities are accompanied with time span. These time
spans follow statistical distributions due to uncertainty and stochastic nature of
automated equipment such as AGV. Therefore, the collected data (activity times)
are considered by goodness of fit technique to obtain the probability distribution
function.
Fig. 1 A configuration of AGV-based jobshop manufacturing system
The material handling begins with taking the raw materials from the depot and
moving to the shops considering the production plan and job sequences to produce
the final product. Products are in batches. Since, material handling activities for
each shop are effective on the total material handling time, all the shops' activities
should be considered, i.e., the total AGV material handling time is an aggregation
of all shops' activities.
The aim here is to obtain the AGV total stochastic material handling time.
While the times are considered to follow statistical distributions, we need to
propose special techniques to obtain the total AGV material handling time. Our
proposed approach is called heuristic statistical method (HSM). The flowchart of
the proposed methodology to consider AGV stochastic activity times to obtain
total material handling time is shown in Figure 2. To analyze the system, the data
are collected for each activity. Then, the collected data are tested via goodness of
fit technique to clarify the probability distribution function (pdf). If the goodness
of fit technique does not provide a known pdf, then cross entropy (to be discussed
later) approach is used. If the pdfs are known and same, i.e., all activity times
follow exponential distribution functions, the HSM is employed to find total
material handling time. Again, if the pdfs are not same, then cross entropy
approach is used. Both HSM and cross entropy approach are verified using
simulation study.

To model the problem, we consider a simple case in which an AGV is assumed to
do the material handling with two activities of direct movement and waiting
showing by X and Y, respectively. The probability distribution for each activity is
exponential as follows:
X ≈ EXP (λ ) , (1)
and the corresponding probability density function is,

f ( X ) = λe −λx , (2)
Also,
Y ≈ EXP ( λ ) , (3)
f (Y ) = λe −λy . (4)
To obtain the total handling time, we need to investigate the joint distribution
function of X and Y. Since they are times, then they are independent. Thus, the
joint density function of independent variables is equal to the products of their
density functions, as follows:
f ( X , Y ) = f ( X ) × f (Y ) , (5)
and thus we want to obtain,
f ( X < z , Y < w) = ? (6)
This problem is easily solved using the distribution function integral of each
variable due to independent variables. But we are looking for a density function to
illustrate the total handling time of the AGV. Therefore, we apply cumulative
distribution function technique.
Data collection
Goodness of fit
No
Known probability distribution functions?
Yes
No
Are the probability distribution functions same?
Yes
Use CE approach to obtain total Use HSM approach to obtain total

AGV material handling time AGV material handling time
Validation by simulation study
Termination
Fig. 2 Flowchart of the proposed methodology

Here, we want to know what the distribution function of W = X + Y is, i.e.,

what the cumulative distribution function of X and Y is. The space of this
cumulative function is shown in Figure 3.
Fig. 3 The space representation of the cumulative function
We have,
f ( X , Y ) = f ( X ) × f (Y ) = λe −λx × λe −λy , (7)
FW (a) = P(W < a) = P( X + Y < a) , (8)
a a−x a a a
∫ ∫ λe
− λx
( )
× λe −λy dydx = ∫ λe −λx 1 − e −λ ( a − x ) dx = ∫ λe −λx − ∫ λe −λa dx =1 − e −λa − λae −λa
.
(9)
0 0 0 0 0
Now, we set partial derivative to obtain the summation density function,
λe −λa (λa )1
f W (a ) =
dFW (a)
da
( )
= λe −λa − λe −λa − λ2 ae −λa = λ2 ae −λa =
1!
, a >0. (10)
Therefore, the obtained density function for W follows Erlang probability

distribution with λ and α = 2 .
Note that the Erlang probability density function is,
λα x α −1e −λx
f ( x;α , λ ) = , x, λ ≥ 0 . (11)
(α − 1)
As a result, if X1, X2,…, Xn are independent exponential random variables with λ,

n
then Y = ∑ X i follows Erlang distribution with λ and α = n .
i =1
To minimize the total AGV handling time, we have to investigate the minimum
cumulative distribution function. Therefore, in an easy case, we consider,
X ≈ EXP(λ1 ) , Y ≈ EXP(λ2 ) and we are looking for W = min( X , Y ) . We
apply the cumulative distribution function, as follows:
FW (a) = 1 − P(W > a) = 1 − P(min(X , Y ) > a ) = 1 − P( X > a, Y > a ) = 1 − (P( X > a) × P(Y > a)) =
( )
1 − e −aλ1 × e − aλ2 = 1 − e −(λ1 +λ2 ) a . (12)
Here, we set the derivative of the cumulative distribution function to obtain the
density function, as below:
dFW ( a )
fW (a) = = (λ1 + λ 2 )e − ( λ1 + λ2 ) a , a ≥ 0. (13)
da
As a result, if X1, X2,…, Xn are independent exponential random variables with

λ1 , λ2 ,..., λn , then W = min( X 1 , X 2 ,..., X n ) follows exponential distribution
with λ = λ1 + λ2 + ... + λn .
This result is reliable when all the activities' probability distributions are
exponential, but if the distributions are different, then we have to explore another
approach.
To break through the model when the pdfs are different or even when the pdfs are
not known, a cross entropy algorithm is proposed and developed in next section.
6.3.1 Cross Entropy (CE)

Due to the drawback of our heuristic proposed approach to aggregate stochastic
AGV activity times when the probability distribution functions are different, we
employ cross entropy to break through the problem.
Let X = {X 1 ,..., X n } be a random vector taking values in some space χ . Let

( f (.;v) ) be a family of probability density functions (pdfs) on χ , with respect to
some base measure μ , where v is a real-valued parameter (vector). Thus,
Ε H ( X ) = ∫ H ( X ) f ( x; v ) μ ( dx) , (14)
χ
for any function H. Note that E stands for expected value. For simplicity μ(dx) is
considered to be dx . Let S be some real function on χ
.
Suppose we are interested in the probability that S(X) is greater than or equal to
some real number
ξ - we consider it as AGV activity time- under f(.;u). This
probability can be expressed as
l = Ρu (S ( X ) ≥ ξ ) = E u I {S ( X )≥ξ } , (15)
Where I is an indicator random variable. A straightforward way to estimate l is to

use Monte Carlo simulation. Draw a random sample X 1 ,..., X n from f(.;u); then
1 N
∑ I{S ( X i ≥ξ } ,
N i =1 (16)
is an unbiased estimator of l. This way is a time consuming simulation effort. An

alternative way is based on importance sampling (Smith et al., 1997; Srinivasan,
2002). Take a random sample X 1 ,..., X n from an importance sampling density g
on χ , and estimate l using the likelihood ratio estimator.
1 N f ( X i ; u)
lˆ = ∑ I{S ( X i )≥ξ } . (17)
N i =1 g( X i )
The best way to estimate l is to use change of measure with density

I {S ( x )≥ξ } f ( x; u )
g ∗ ( x) = . (18)
l
Therefore, we obtain from (17)

f ( X i ; u)
I {S ( X i )≥ξ } =l, for all i. (19)
g∗(Xi )
l is a constant value with zero variance. Here, the problem is that g* depends on
the unknown parameter l. it is convenient to choose a g in the family of densities
{ }
f (.;v) . Now, we need to choose the reference parameter v such that the
distance between the density g* in (18) and f(.;v) is minimal. A usual measure of
distance between two densities, say g and h, is the Kullback-Leibler distance
(Kullback and Leibler, 1951; Kullback, 1959) or the cross entropy between g and
h. The Kullback-Leibler distance is defined as,
g(X )
h( X ) ∫
D = ( g , h ) = Ε g ln = g ( x ) ln g ( x ) dx − ∫ g ( x ) ln h ( x ) dx . (20)
Minimizing the Kullback-Leibler distance between g* in (18) and f(.;v) is
∫g
∗
equivalent to choosing v such that − ( x ) ln f ( x, v ) dx is minimized, which
is solving the following maximization problem,
max ∫ g ∗ ( x) ln f ( x, v)dx , (21)

v
Substituting g* from (18) into (21) we obtain the maximization program

I {S ( x )≥ξ } f ( x; u )
max
v ∫ l
ln f ( x , v ) dx , (22)
which is,
max D ( x ) = max Ε u I {S ( x )≥ξ } ln f ( x, v ) . (23)
v v
Again, using importance sampling with a change of measuref(.;w) we can rewrite

(23) as
max D (v ) = max Ε w I {S ( X )≥ξ }W ( X ; u , w) ln f ( x , v ) , (24)
v v
for any reference parameter w, where,

f ( x, u )
W ( x; u , w) = , (25)
f ( x, w)
Is the likelihood ratio at x, between f(.;u) and f(.;w). The optimal solution of (24)
can be written as:
v ∗ = arg max Ε w I {S ( X )≥ξ }W ( X ; u , w) ln f ( X ; v) , (26)
v
We can solve v* by solving the following stochastic program
N
1
max D̂(v) = max
v v N
∑ I{
i =1
W ( X i ; u, w) ln f ( X i ; v) .
S ( X i )≥ξ } (27)
where X 1 ,..., X n is a random sample from f(.;w).

this way we obtain an estimated probability for stochastic AGV activity time.
While these probabilities are real values, we can obtain the aggregated value by
simple adding all activities' times. Here, all stochastic times from various
probability distribution functions can be aggregated by CE algorithm to compute
the total AGV material handling time. A substantial advantage of the proposed CE
algorithm is to be capable for data collections without any known probability
distribution function. The following algorithm is composed to facilitate the CE
computations:
Algorithm 1. Cross entropy method for aggregating various probability

distribution functions
Step 1. Define vˆ0 = u . Set t=1 (iteration)

Step 2. Generate a sample X 1 ,..., X n from density f(.;vt-1) and compute the
sample (1-q)-quantile ξˆt of the performances according to ξˆt = S (⎡(1−q ) N ⎤) ,

provided ξˆt is less than ξ . Otherwise set ξˆt = ξ .
Note that,
Ρvt −1 (S ( X ) ≥ ξ t ) ≥ q ,
Ρvt −1 (S ( X ) ≤ ξ t ) ≥ 1 − q .
Step 3. Use the same sample X 1 ,..., X n to solve the stochastic program
N
1
max D̂(v) = max
v v N
∑ I{
i =1
S ( X i )≥ξˆt }W ( X i ; u, vˆt −1 ) ln f ( X i ; v) .
Denote the solution by v̂t .
Step 4.If ξˆt < ξ , set t=t+1 and reiterate from Step 2. Else proceed with Step 5.
Step 5. Estimate probability l using likelihood ratio estimate
N
1
lˆ =
N
∑ I{
i =1
S ( X i ) ≥ξ } W ( X i ; u , vˆT ) .
where T denotes the final number of iterations.
Next section presents verification and validation of both HSM and CE approach
for an AGV-based jobshop manufacturing system. The most significant advantage
of our proposed heuristic method and CE is to be helpful when the data are few.
6.4 Conclusions
Here, a heuristic statistical technique to compute total stochastic material handling

time in an automated guided vehicle (AGV) equipped jobshop manufacturing
system was proposed. With respect to stochastic times of AGVs material handling
process, each activity was associated with a probability distribution. Using the
probability distributions, we modeled the AGV material handling problem using a
heuristic statistical method (HSM) when the activities' probability distribution
functions are same. To model the problem when the activities' probability
distribution functions are different, a cross entropy algorithm is developed. The
problem was investigated and implemented in two cases with same probability
distributions and different ones. The effectiveness of the proposed model was
validated in an example. An experimental test problem was worked out for each of
the cases. For the second case, to verify the obtained probability distribution of the
total AGV handling time, we applied chi-square and Kolmogorov-Smirnov
goodness-of-fit tests. The simulation was also implemented by several softwares
to analyze their behaviors for the proposed problem ensuing to better performance
of ARENA. The computations validated by simulation illustrated the efficiency of
the HSM approach in the case while all distributions were same and the CE
method for the case having different probability distributions of activities.
References
Arai, T., Pagello, E., Parker, L.E.: Guest editorial, Special issue on advancement in multi-
robot systems. IEEE Trans. RA 18(5), 655–662 (2002)
Asef-Vaziri, A., Laporte, G.: Loop-Based Facilities Planning and Material Handling.
Butdee, S., Suebsomran, A.: Learning and recognition algorithm of intelligent AGV
system. In: Proceedings of the Global Congress on Manufacturing and Management,
Santos, Brazil, pp. 13–72 (2006)
Butdee, S., Suebsomran, A.: Localization Based on Matching Location of AGV. In:
Proceeding of the 24th International Manufacturing Conference, IMC24, pp. 1121–
1128. Waterford Institute of Technology, Ireland (2007)
Butdee, S., Vignat, F., Suebsomran, A.: Self-alignment control of an automated unguided
vehicle. In: Proceedings of the 6th International Conference on Integrated Design and
Manufacturing in Mechanical Engineering, Grenoble, France, pp. 48–64 (2006)
Buzacott, J.A.: The structure of manufacturing systems: insights on the impact of
variability. International Journal of Flexible Manufacturing System 11, 127–146 (1999)
Co, H.C., Wysk, R.A.: Robustness of CAN-Q in modeling automated manufacturing
systems. International Journal of Production Research 24(6), 1485–1503 (1986)
Dallery, Y., David, R.: Some new results on operational analysis. In: Galenbe, E. (ed.)
Performance, vol. 84. Elsevier Science Publishers, B.V., Amsterdam (1984)
De Boer, P.T., Kroese, D.P., Mannor, S., Rubinstein, R.Y.: A Tutorial on the Cross-
Entropy Method. Annals of Operations Research (2005)
References 91
Dubois, D.: A mathematical model of a flexible manufacturing system with limited in-
process inventory. International Journal of Production Research 14(3), 66–78 (1983)
Fazlollahtabar, H., Rezaie, B., Kalantari, H.: Mathematical programming approach to
optimize material flow in an AGV-based flexible jobshop manufacturing system with
performance analysis. International Journal of Advanced Manufacturing Technology 51,
1149–1158 (2010)
Kaynar, B., Ridder, A.: The cross-entropy method with patching for rare-event simulation
of large Markov chains. European Journal of Operational Research 207, 1380–1397
(2010)
Kullback, S.: Information theory and statistics. John Wiley and Sons, NY (1959)
Kullback, S., Leibler, R.A.: On Information and Sufficiency. Annals of Mathematical
Statistics 22(1), 79–86 (1951)
Papadopoulos, H.T., Heavy, C.: Invited review: queueing theory in manufacturing systems
analysis and design. A classification of models for production and transfer lines.
European Journal of Operational Research 92, 1–27 (1996)
Reveliotis, S.A.: Conflict resolution in AGV systems. IIE Trans. 32, 647–659 (2000)
Rubinstein, R.Y.: Optimization of Computer simulation Models with Rare Events.
European Journal of Operations Research (1997)
Rubinstein, R.Y.: The simulated entropy method for combinatorial and continuous
optimization. Methodology and Computing in Applied Probability (1999)
Shirazi, B., Fazlollahtabar, H., Mahdavi, I.: A six sigma based multi-objective optimization
for machine grouping control in flexible cellular manufacturing systems with guide-path
flexibility. Advances in Engineering Software 41, 865–873 (2010)
Smith, P.J., Shafi, M., Gao, H.: Quick simulation: A review of importance sampling
techniques in communication systems. IEEE J. Select. Areas Commun. 15, 597–613
(1997)
Srinivasan, M.M., Bozer, Y.A., Chao, M.: Trip-based material-handling systems.
Throughput capacity analysis. IIE Trans. 26(1), 70–89 (1994)
Srinivasan, R.: Importance sampling - Applications in communications and detection.
Springer, Berlin (2002)
Sule, D.R.: Manufacturing Facilities: Location, Planning, and Design. PWS, Boston (1994)
Suri, R., Hildebrant, R.R.: Modeling flexible manufacturing systems using mean value
analysis. Journal of Manufacturing Systems 3(1), 21–38 (1984)
Vis, I.F.A.: Survey of research in design and control of automated guided vehicle systems.
European Journal of Operational Research 170, 677–709 (2006)
Chapter 7
Neuro-Fuzzy-Regression Expert System for
AGV Optimal Path
7.1 Summary
According to Turban and Aronson (1998), a decision support system (DSS) is a

computer-based information system combining models and data in an attempt to
solve non-structured problems with extensive user involvement. They advocated
that an expert system (ES) was a computer system that applied reasoning
methodologies on knowledge to render advice or recommendations much like a
human expert. When expert system technology was first applied to decision-
making problems, it fell short in several respects. Early expert systems were rule-
based and thus were not capable of handling the classical DSS functions being
more computational than logical. Recently, artificial intelligence researchers have
noted the necessity for using statistical techniques to build intelligent decision
support systems (Nolan, 1998; Weiss and Kulikowski, 1991). Examples of such
statistical techniques include fuzzy logic, neural networks, rule induction and
various Bayesian techniques. Turban and Aronson (1998) believe that although
uncertainty is widespread in the real world, but practical treatment in artificial
intelligence is very limited.
Cost estimation plays a very important role in design and production stages as
well as a fairly important role in company business decision-making. Effective
cost estimation can contribute to an optimal design leading to customers’
satisfaction due to cost reduction, improved quality and timely delivery (Wu and
Zhou, 2005).
Manufacturing costs in products are very crucial in decision making and
strategic planning. With respect to cost estimation, research and development
departments were mainly concerned with the estimate of the total cost for the
finished product (Vis, 2006). Moreover, engineers often applied rules of thumb as
the cost estimation benchmarks, making the results controversial in terms of
accuracy (Khayat et al., 2006). Although calculation by cost model has the
advantage of timeliness, only representative values exclusive of indirect tasks cost
and raw materials cost are calculated resulting in inadequate estimation accuracy.
Niazi et al. (2006) pointed out that back propagation network (BPN) could be

94 7 Neuro-Fuzzy-Regression Expert System for AGV Optimal Path
applied for training to deduce unprecedented problems by accumulated knowledge

and information. Specially, it can find out solutions in uncertain circumstances
and has satisfactory results in dealing with nonlinear problems. Therefore, BPN is
one of the most popular neural network models being applied and it properly fits
the nature of product cost estimation. McKim (1993) proposed a discussion on
applying BPN in cost estimation projects. Finally, by integrating a user interface,
the demand of a fast response cost estimation model was met at the initial product
development stage.
Cavalieria et al. (2004) on the spare parts cost estimation in auto industry
pointed out that, at the final stage of the product life cycle, most product costs
continuously rise because they are determined at the product concept design stage.
Yang and Lin (1997), Rehman and Guenov (1998) and Gayretli and Abdalla
(1999) also pointed out time and again that the product design initial stage
determined 70–80% of the total product cost.
Cost affects sales price, sale volume and profit most directly. In addition,
Vrbsky (1997) pointed out that the research and development design cost often
accounted for almost 50% of the product development cost and was an important
factor determining whether the product total cost would be reasonable or not. For
the same reason, accurately estimated design cost cannot only help reduce the
capital-losing risks in the product research and development process but also can
shorten the product development time period. Such methods can really reduce the
expense for product development (Ghasemzadeh et al., 2009).
The automotive industry operates in a very competitive market which requires
controlling product costs, improving the product quality and shortening the
development lead time. Cost estimation requires quality data and information. Roy
et al. (2011) presented the various data and information requirements for detailed
cost estimating in automotive industry. The research project has identified the
common cost estimation process model within the identified industry sector. The
study identified the types of data and information requirements for cost estimating.
It has also constructed the relevant data infrastructure as the basis for a Web
Portal, which was the physical presentation of the information infrastructure.
Multiple sources of data collection techniques were employed to identify the types
of data needed for detailed cost estimates within a manufacturing company.
Quintana and Ciurana (2011) proposed a machine tool selection problem
consisting of selecting the most suitable machine to satisfy manufacturing
company requirements. The main goal of the work was to develop a cost
estimation support tool for vertical high speed machining centers based on final
part and productivity requirements of the company linked with machine tool
characteristics available in the catalogues in order to apply the cost model and to
calculate machine tool cost estimations. The cost model presented was based on
multiple regression analyses and provided reasonably accurate market cost
predictions. Applying the proposed cost model will help the user to determine the
approximate market cost of the machine and can be especially interesting for
decision makers in the preliminary stages of a selection process because it avoids
long and costly studies.
7.1 Summary 95
Flexible material handling systems (MHS) have been widely used to enhance
productivity involved with product proliferation, and thus far, only fixed-track
MHSs such as Eton systems in the apparel industry are commonly used. Dai and
Lee (2012) explored the economic feasibility of a flexible MHS using free-ranging
automated guided vehicles (AGV) with a local positioning system (LPS) for the
apparel industry. A component-based and modified activity-based costing
methodology was proposed to estimate the additional cost of adopting flexible
MHSs, and then the internal rate of return (IIR) and payback periods were applied
to evaluate the project economic performance.
Accurate cost estimation plays a significant role in industrial product
development and production. Deng and Yeh (2011) applied least squares support
vector machines (LS-SVM) method solving the problem of estimating the
manufacturing cost for airframe structural projects. They evaluated the estimation
performance using back-propagation neural networks and statistical regression
analysis.
Product cost estimation varies widely ranging from standard spare parts
manufacturing cost estimation to the cost analysis of the optimized technology and
marketing fees of highly customized assembled products, with appropriate product
estimation models available from the product concept design stage to the product
design cycle’s final stages. Zhang et al. (1996) categorized cost estimation
techniques into traditional detailed breakdown, simplified-breakdown, group-
technology-based, regression-based and activity-based cost approaches. Ben-Arieh
and Qian (2003) divided cost estimation models into intuitive, analogical,
parametric and analytical approaches. Shehab and Abdalla (2001) proposed
intuitive, parametric, variant-based and generative cost estimating approaches.
Cavalieria et al. (2004) provided three cost analyses of analogy-based, parametric
and engineering approaches. Niazi et al. (2006), on the basis of the integrated cost
estimation approaches, categorized the qualitative and quantitative cost estimation
techniques along with key advantages and limitations of each cost estimation
technique.
Here, we employ soft computing techniques for configuring an expert system in
the presence of uncertainty. The purpose of our study is to design a cost estimation
model for an AGV based automated manufacturing system. We choose three main
factors accompanied by their sub-factors and then propose a neural network based
approach to analyze the effects of the sub-factors on the main factors. Due to
uncertainty of cost elements in manufacturing environment, we apply fuzzy logic
to determine an appropriate range for each of the linguistic variables. Then, using
the linguistic variables, we extract the possible fuzzy rules and using a multiple
linear regression analysis we investigate the significant rules and eliminate
inappropriate ones. A sensitivity analysis is considered along with the regression
method. We then use a dynamic programming approach to find an optimal path
for the proposed manufacturing system.
carries raw material, semi-produced and finished products in batch sizes. Due to
shows up in the shop's specificities that affect the production cost. As a result, the
system would consider a flexible jobshop model where multi shops of the same
duty exist and each operation can be processed on any type of machine in any
shop. The sequences of jobs are specified and the jobs are assumed to be
independent.
rates of defect, etc.. Each flow path for the AGV is associated with a cost
parameter related to each shop. The aim is to find a path for the AGV minimizing
the cost objective. In each shop, different machines and operators are working.
Due to unpredictable events during working times a cost may incur. This cost is
inferred from an expert system via fuzzy logic "IF …. Then……." rules. Cost
parameters of each shop are considered to be 3 main factors: (1) Equipment
sensitivity, (2) operator proficiency, and (3) product specifications, each being
specified by one of the three levels of low, moderate, and high. A configuration
for the proposed problem is presented in Figure 1.
Fig. 1 A configuration of the proposed problem.

As stated before, cost is inferred from an expert system. A flowchart for the
proposed expert system is presented in Figure 2. In our expert system, a back
propagation neural network is considered to estimate the cost factors (outputs)
using their corresponding sub-factors (inputs). The cost factors and their related
sub-factors are shown in Figure 3. Using the existing data, we train the network
and then by the resulting pattern we can apply the model to obtain the output with
respect to the proposed sub-factors. The numerical results would show a lower
bound and an upper bound for each cost factor. We utilize these bounds as our
cost factors ranges for the fuzzy rule base. Considering the uncertainty in the cost
factors due to dynamic changes in neural network inputs, we specify the cost
factors in three levels of low, moderate, and high. The problem is to specify the
range of these levels, i.e., the numerical range of low, moderate or high for each of
the cost factors. This process is repeated for each AGV movement between any
two shops for determining the arcs cost values.
Fig. 2 A proposed expert system.

Fig. 3 The cost factors and their related sub-factors.
We apply an inductive reasoning technique to obtain the appropriate range for the
levels. This method is based on an ideal scheme describing the input and output
relationships for a well-established data base. This method is called entropy. A
key goal of entropy minimization analysis is to determine the quantity of
information in a given data set. The entropy of a probability distribution is a
measure of the uncertainty of the distribution. This information measure estimates
the uncertain range of data using a predetermined inappropriate range to start the
process. The higher the prior estimate of the probability for an outcome to occur,
the lower will be the information gained by observing it to occur. The entropy on a
set of possible outcomes of a trial where exactly one outcome is possible is
defined by the sum of probabilities. In other words, the entropy is the expected
value of information. For a simple one-dimensional (one uncertain variable) case,
let us assume that the probability of the ith sample wito be true is p(wi ) . If we
actually observe the sample wi in the future and discover that it is true, then we
gain the following information, I(wi):
I (wi ) = −k ln p(wi ) , (1)

wherek is a normalizing parameter. If we discover that it is false, we still gain this

information:
I (wi ) = −k ln[1 − p(wi )] . (2)
Thus, the entropy of the inner product of all the samples (N) is:
N
S = −k ∑ [ pi ln pi + (1 − pi ) ln(1 − pi )], (3)
i =1
wherepiis the probability of the ith sample to be true. Note that S ≥ 0 , because
ln x ≤ 0 , for 0 ≤ x ≤ 1 .
The entropy of a rule should be minimized. Minimum entropy (S) is associated
with all the pi being as close to ones or zeros as possible, which in turn implies
that they have a very high probability of either happening or not happening,
respectively. Note in equation (3) that if pi=1 then S=0.
This result makes sense, since pi is the probability measure of whether a value
belongs to a partition or not. The precedence for partitioning the data is as follows.
First, we assume that we are seeking a threshold value for a sample in the range
between [x1, x] and [x, x2]. We denote the first region p and the second region q.
by moving an imaginary threshold value x between x1 and x2, we calculate entropy
for each value of x.
An entropy with each value of x in the region x1 and x2 is:
S ( x) = p ( x) S p ( x ) + q ( x) S q ( x) , (4)
where,
S p ( x) = −[ p1 ( x) ln p1 ( x) + p 2 ( x) ln p 2 ( x)] , (5)
S q ( x) = −[q1 ( x) ln q1 ( x) + q 2 ( x) ln q 2 ( x)] , (6)
with pk(x) and qk(x) as conditional probabilities that the class k sample is in the
region [x1, x1+x] and [x1+x, x2], respectively, p(x) and q(x) are probabilities that all
samples are in the region [x1, x1+x] and [x1+x, x2], respectively. Also,
p ( x) + q ( x) = 1 . (7)
A value of x that gives the minimum entropy is the optimum threshold value. We
calculate entropy estimates pk(x), qk(x), p(x), and q(x) as follows:
n k ( x) + 1
p k ( x) = , (8)
n( x ) + 1
N k ( x) + 1
q k ( x) = , (9)
N ( x) + 1
n( x )
p ( x) = , (10)
n
q ( x) = 1 − p( x) , (11)
where nk(x) is the number of class k samples located in [x1, x1+x], n(x) is the total
number of samples located in [x1, x1+x], Nk(x) is the number of class k samples
located in [x1+x, x2], N(x) is the total number of samples located in [x1+x, x2], and
n is the total number of samples in [x1, x2]. While moving x in the region [x1, x2],
we calculate the values of entropy for each position of x.
This procedure finds the region for the levels of the cost factors. After each
update in input data the ranges are specified using an entropy technique. After
determining the regions of the cost factors, we then compose the fuzzy rules which
clarify the total cost for each shop. To find the effective rules, we configure a
multiple linear regression model using the previous data and identify the
regression coefficients. Since we have three cost factors each of which contains
three levels of low, moderate and high, then twenty seven (3*3*3) possible rules
exist. Thus, using the hypothesis testing, the effective rules are identified. By these
rules we estimate the cost of each shop (arc length in the network). Consequently,
a dynamic program is applied to find the optimal path in the proposed jobshop
automated manufacturing network.
7.3.1 Artificial Neural Network

Neural networks are being widely used in many fields of study. This could be
attributed to the fact that these networks attempt to model the capabilities of
human brains. Since the last decade, neural networks have been used as a
theoretically sound alternative to traditional statistical models. Although neural
networks (NNs) originated in mathematical neurobiology, the rather simplified
practical models currently in use have moved steadily towards the field of
statistics. A number of researchers have illustrated the connection of neural
networks to traditional statistical models. For example, Gallinari et al. (1991)
presented analytical results establishing a link between discriminant analysis and
multilayer perceptrons (MLP) used for classification problems. Cheng and

Titterington (1994) made a detailed analysis and comparison of various neural
network models with traditional statistical models. They showed strong
associations of the feed-forward neural networks with discriminant analysis and
regression models, and unsupervised networks such as self-organizing neural
networks with clustering. Neural networks are being used in areas of prediction
and classification, areas where regression models and related statistical techniques
have traditionally been used. Ripley (1994) discusses the statistical aspects of
neural networks and classifies neural networks as one of a class of flexible
nonlinear regression models. Sarle (1994) translates neural network terminologies
into statistical ones and shows the relationship between neural networks and
statistical models such as generalized linear models, projection pursuit and cluster
analysis. He explains that neural networks and statistical approaches are not
competing methodologies for data analysis and there is a considerable overlap
between the two. Warner and Misra (1996) present a comparison between
regression analysis and neural network computation in terms of notation and
implementation. They also discuss when it would be advantageous to use a neural
network model in place of a parametric regression model, as well as some of the
difficulties in implementation. Schumacher et al. (1996) and Vach et al. (1996)
present a comparison between feed-forward neural networks and the logistic
regression. The conceptual similarities and discrepancies between the two
methods are also analyzed.
Some of the commonly used traditional statistical techniques applied for
prediction and classification are multiple regression, discriminant analysis, and
logistic regression. These methods are being widely used in various applications
due to their established methodologies. However, neural networks are being used
as alternatives to these traditional techniques and gaining much popularity in
recent years. This has led to a number of studies comparing the traditional
statistical techniques with neural networks in a variety of applications.
Artificial neural networks have been applied successfully to many
manufacturing and engineering areas. Zhengrong et al. (1996) used quadratic
regression to assess the results of neural network for improving the efficiency of
fermentation process development. The results show that different sizes of neural
nets within a certain range give an equally good prediction by using the ‘‘stopping
training” technique, while quadratic regressions are sensitive to the size of the data
sets. Smith and Mason (1997) mentioned that regression and neural network
modeling methods have become two competing empirical model-building
methods. They compared the predictive capabilities of NNs and regression
methods in manufacturing cost estimation problems. This study examines the
performance, stability and ease of cost estimation modeling using regression
versus neural networks to develop cost estimating relationships (CERs). Results
show that neural networks have advantages when dealing with data not adhering
to the generally chosen low order polynomial forms, or data for which there is
little a priori knowledge of the appropriate CER to select for regression modeling.
However, in cases where an appropriate CER can be identified, regression models
have significant advantages in terms of accuracy, variability, model creation and

model examination. Both simulated and actual data sets were used for comparison
purposes. Researchers have also discussed some important issues other than model
accuracy to be considered when using regression versus neural network to
estimate cost functions. Finnie et al. (1997) used neural network models to
estimate software development efforts. Improving the estimation techniques
available to project managers would facilitate more effective control of time and
budgets in software development. They compared three estimation techniques
using function points as an estimate of system size. The models considered were
based on regression analysis, artificial neural networks and case-based reasoning.
Although regression models performed poorly on the data set of 299 software
projects, both artificial neural networks and case-based reasoning appeared to be
useful for software development effort estimation models. Also, Heiat (2002)
compared artificial neural network and regression to estimate the amount of effort
required for developing an information system, an important project management
concern. The results of the study indicated that for one experiment (when only
third generation language data set was considered), regression analysis and neural
networks performance were nearly identical and for a second experiment (when a
combined third generation and fourth generation language data set as used), the
neural network produced improved performance over conventional regression
analysis in terms of mean absolute percentage error (MAPE). They also stated that
autocorrelation could have been a possible reason for the large MAPE values and
not its non-normality as in both cases the data sets were not normally distributed.
Lee and Um (2000) used multiple regression analysis and artificial neural network
methods for predicting the relationship between welding parameters and geometry
of the welding area in arc welding. It was found that the error rate predicted by the
artificial neural network was smaller than that predicted by the multiple regression
analysis. Shuhui et al. (2001) compared regression and neural networks to predict
the power produced by wind farms and found that neural networks performed
better than regression models. Applicability of NN for cost estimation in building
construction was tested by Setyawati et al. (2002). The authors concluded that
neural networks outperformed linear regression models given the same training
data and the same variables. Krishnaswamy and Krishnan (2002) used regression
analysis and neural network in nozzle wear rate prediction and found a statistical
significant difference in the prediction of validation data. Feng and Wang (2002)
compared nonlinear regression and neural network models in developing
empirical models for estimating the digitizing accuracy of a given coordinate-
measuring machine. It was shown that both prediction models appear to provide a
satisfactory prediction with the regression model providing a slightly better
performance in both model construction and model verification. Feng and Wang
(2002) in a different article applied two competing data mining techniques,
nonlinear regression analysis and computational neural networks in developing the
empirical models for surface roughness prediction. It was demonstrated by
hypothesis testing of relative errors and absolute errors for the process of
modeling construction, model validation, and combined data modeling that no
statistically significant differences were found between the performances of the

two methods. Kim et al. (2004) applied three techniques, namely multiple
regression analysis, neural network and case based reasoning for estimating
construction costs of Korean residential buildings. Authors concluded that
although the best NN model performed better than the other models for accuracy
of estimation, the training procedure to train the best NN model was slow because
of the trial and error process. In addition to the accuracy of estimation results, the
time and accuracy tradeoffs and also the clarity of explanation should be
considered in a cost estimation model. Yesilnacar and Topal (2005) used a neural
network and compared it with logistic regression for landslide susceptibility
mapping in a study for Hendek region in Turkey. The landslide susceptibility map
produced using the neural network method predicts higher percentages of
landslides, specially in high and very high zones than the logistic regression
method. Few studies in the project management literature concentrated on the
critical success factors that affect project success or failure. A comparison
between neural network and linear regression analysis was used for identifying
critical managerial factors affecting the success of high-tech defense projects by
Dvir et al. (2006). The study showed that neural networks have significantly better
explanatory and prediction powers, and they enable the exploration of
relationships among the data that are difficult to arrive at by traditional statistical
methods. Pendharkar (2006) investigated the factors influencing the object-
oriented (OO) component size and source code documentation.
Using empirical data, researcher compared the performance of nonlinear
artificial neural network (ANN) forecasting models and linear regression models.
Experimental results confirmed the superior performance of ANN over the linear
multiple regression model when variable returns to scale economies exist between
multiple inputs and multiple outputs.
7.3.2 The Backpropagation Neural Network

The backpropagation algorithm trains a given feed-forward multilayer neural
network for a given set of input patterns with known classifications. When each
entry of the sample set is presented to the network, the network examines its
output response to the sample input pattern. The output response is then compared
to the known and desired output and the error value is calculated. Based on the
error, the connection weights are adjusted. The backpropagation algorithm is
based on Widrow-Hoff delta learning rule in which the weight adjustment is done
through mean square error of the output response to the sample input (Abdi et al.,
1996).
Algorithm 1. General steps of backpropagation
1. Propagate inputs forward in the usual way, i.e., all outputs are computed using
sigmoid thresholding of the inner product of the corresponding weight and input
vectors. All outputs at stage n are connected to all the inputs at stage n+1
2. Propagate the errors backwards by apportioning them to each unit according to

the amount of the error the unit is responsible for.
We now discuss how to develop the stochastic backpropagation algorithm for the
general case. The derivation is simple, but unfortunately the book-keeping is a
little messy. The following notations and definitions are needed:
x j : input vector for unit j (xji = ith input to the jth unit)
w j : weight vector for unit j (wji = weight on xji)
z j = w j .x j : the weighted sum of inputs for unit j
oj : output of unit j ( o j = σ (z j ) )
tj : target for unit j
Downstream(j) : set of units whose immediate inputs include the output of j
Output : Set of output units in the final layer.
Since we update after each training example, we can simplify the notation
somewhat by assuming that the training set consists of exactly one example and so
the error can simply be denoted by E.
∂E
We want to calculate corresponding to each input weight wji of each
∂w ji
output unit j. Note first that since zj is a function of wji regardless of where in the
network unit j is located,
∂E ∂E ∂z j ∂E
= . = .x ji , (12)
∂w ji ∂z j ∂w ji ∂z j
Furthermore, ∂E is the same regardless of which input weight of unit j we are

∂z j
trying to update. So, we denote this quantity by δj.
Consider the case when j is an output unit. We know that

1
E= ∑ (t k − σ ( z k )) 2 .
2 k∈Outputs
(13)
Since the outputs of all units k ≠ j are independent of wji, we can then drop the
summation and consider just the contribution to E by j and we call it δj:
∂E ∂ 1 ∂o j ∂
δj = = (t j − o j ) 2 = −(t j − o j ) = − (t j − o j ) σ (z j )
∂z j ∂z j 2 ∂z j ∂z j (14)
= −(t j − o j )(1 − σ ( z j ))σ ( z j ) = −(t j − o j )(1 − o j )o j .
Thus,
∂E
Δw ji = −η = ηδ j x ji . (15)
∂w ji
Now, consider the case when j is a hidden unit. Like before, we make the
following two important observations:
1. For each unit k downstream from j, zk is a function of zj.
2. The contribution to error by all units l ≠ j , in the same layer as j, is
independent of wji.
∂E
We want to calculate for each input weight wji for each hidden unit j.
∂w ji
Note that wji influences just zj which influences oj which influences zk,
∀k ∈ Downstream( j ), each of which influences E. So, we can write,
∂E ∂E ∂z k ∂o j ∂z j ∂E ∂z k ∂o j
= ∑ . . . = ∑ . .
∂w ji k∈Downstream ( j ) ∂z k ∂o j ∂z j ∂w ji k∈Downstream ( j ) ∂z k ∂o j ∂z j
.x ji . (16)
Again, note that all the terms except xji in (16) are the same regardless of which
input weight of unit j we are trying to update. Like before, we denote this common
∂E ∂z ∂o
quantity by δ j . Also, note that δ k , k wkj and j = o j (1 − o j ) .
∂z k ∂o j ∂z j
Substituting them in (14),
∂E ∂z k ∂o j
δj = ∑ . .
k∈Downstream( j ) ∂z k ∂o j ∂z j
= ∑ δ k .wkj .o j (1 − o j ) ,
k∈Downstream( j )
(17)
we obtain:
δ k = o j (1 − o j ) ∑δ k .w .
k∈Downstream( j )
kj (18)
We are now in a position to state the backpropagation algorithm formally.

Algorithm 2. Formal statement of stochastic backpropagation

(Training examples, η , ni, nh, no)
Each training example is of the form x, t , where x is the input vector and t is
the target vector, η is the learning rate (e.g., 0.05), ni, nh and no are the number of
input, hidden and output nodes, respectively. Input from unit i to unit j is denoted
by xji and its weight is denoted by wji. Create a feed-forward network with ni
inputs, nh hidden units, and no output units.
Initialize all the weights to small random values (e.g., between -0.05 and 0.05).
While termination condition is not met Do
For each training example x, t ,
1. Input the instance x and compute the output ou of every unit.
2. For each output unit k, calculate
δ k = ok (1 − ok )(t k − ok ) . (19)
3. For each hidden unit h, calculate
δ h = oh (1 − oh ) ∑δ k .w .
k∈Downstream( h )
kh (20)
4. Update each network weight wji as follows:
w ji ← w ji + Δw ji . (21)
where,
Δw ji = ηδ j x ji . (22)
7.3.3 Fuzzy Logic

Fuzzy Logic (FL) is a problem-solving control system methodology that lends
itself to implementation of systems ranging from simple, small, embedded micro-
controllers to large, networked, multi-channel PC or workstation-based data
acquisition and control. It can be implemented in hardware, software, or a
combination of both. FL provides a simple way to arrive at a definite conclusion
based upon vague, ambiguous, imprecise, noisy, or missing input information. The
FL approach to control problems mimics how a person would make decisions,
only much faster.
FL incorporates a simple, rule-based IF X AND Y THEN Z approach to solving

a control problem rather than attempting to model a system mathematically. The
FL model is empirically-based, relying on the operators’ experience rather than
their technical understanding of the system.
FL requires some numerical parameters in order to operate such as what is
considered to be significant error and significant rate-of-change-of-error, but exact
values of these numbers are usually not critical unless very responsive
performance is required in which case empirical tuning would determine them.
7.3.4 Mamdani Fuzzy System

Mamdani fuzzy system was proposed as a first attempt to control a steam engine
and boiler combination by a set of linguistic control rules obtained from
experienced human operators. Rules in Mamdani fuzzy systems are like these
(Klir and Yuan, 1995; Zimmermann, 1996):
If x1 is A1 AND/OR x2 is A2 Then y is B1,
whereA1, A2 and B1 are fuzzy sets. The fuzzy set acquired from aggregation of
rules’ results are defuzzified using defuzzification methods like centroid (center of
gravity), max membership, mean-max, and weighted average. The centroid
method is very popular, in which the ‘‘center of mass’’ of the result provides the
crisp value. In this method, the defuzzified value of the fuzzy set A, d(A), is
calculated by:
∫ x.μ A ( x) x
d ( A) = X
, (23)
∫ μ A ( x) x
X
where μ A (.) is the membership function of the fuzzy set A (Klir and Yuan, 1995;
Zimmermann, 1996).
For our problem in which various possible conditions of parameters are stated
in forms of fuzzy sets, we utilize the Mamdani fuzzy system, because in this
system the fuzzy rules representing the expert knowledge would consider fuzzy
sets as their consequences. In general, designing a fuzzy system is composed of
the following major steps (Klir and Yuan, 1995).
Algorithm 3: General steps for the design of a fuzzy logic system
Step 1. Identify pertaining input and output variables. Besides, select the
meaningful linguistic states along with appropriate fuzzy sets for the variable.
Step 2. Employ a fuzzification method for input variables that expresses the
associated measured uncertainty. The purpose of the fuzzification method is to
interpret measurements of input variables which are expressed by real numbers.
Step 3. Formulate pertaining knowledge in terms of fuzzy inference rules. There

are two principal ways in which relevant inference rules can be determined. One is
to elicit them from experienced humans and the other is to obtain them from
empirical data by suitable learning methods, usually with the help of neural
networks.
Step 4. Combine measurements of input variables with relevant fuzzy rules to

inference, considering the output variables in a way to admit the purpose of the
inference engine.
Step 5. Ascertain a suitable defuzzification method to convert the aggregated

fuzzy set of implications into a real number.
7.3.5 Multiple Linear Regression

As we stated before, some rules affect the cost of the system. We should
investigate whether a rule is significant on cost or not. One way to survey the
effect of independent variables on dependent variable is to use a multiple linear
regression model. Therefore, we consider the following equation,
ui = β 0 + β1r1 + ... + β j r j + ε j , i = 1,..., n, j = 1,..., m, (24)
where β0 is the intercept, the βj are the coefficients of the rjparameters and the
ε j are the error terms. The aim is to identify the β j not being important on total
cost of the proposed system. Here, we apply two-sided hypothesis testing as
follows:
H 0 : β j = 0,
(25)
H 1 : β j ≠ 0.
The null hypothesis indicates whether a rule coefficient is zero or not. If the
hypothesis is accepted (a certain rule coefficient is zero) then we can omit the
corresponding rule from cost estimation process. The test of the null hypothesis H0
against the two sided alternative proceeds in three steps. The first is to compute
the standard error of β j , SE ( β j ) . The standard error of β j is an estimator of
σβ j
, the standard deviation of the sampling distribution of βj:
m
1
× ∑ (r j − r ) 2 ε 2j
1 m − 2 j =1
σ β2 = × 2
, (26)
m ⎡1 m 2⎤
j
⎢ ∑ (r j − r ) ⎥
⎣ m j =1 ⎦
SE ( β j ) = σ β2 j . (27)

is complicated, but in practice the standard error is
computed by a regression software. The second step is to compute the t-statistic:
βj −0
t= . (28)
SE ( β j )
act
⎡ βj −0 β jact − 0 ⎤
[ ]
p − value = PrH 0 β j − 0 > β jact − 0 ⇒ PrH 0 ⎢
β
>
β
⎥ ⇒ PrH 0 ( t > t act )
,
(29)
⎣⎢ SE ( j ) SE ( )
j ⎦⎥
where PrH 0 denotes the probability computed under the null hypothesis, the second
equality follows by dividing into SE ( β j ) , and t act is the value of the t-statistic
actually computed. Because β j is approximately normally distributed in large
samples, under the null hypothesis the t-statistic is approximately distributed as a
standard normal random variable, and so for large samples,
p − value = Pr( Z > t act ) = 2Φ(− t act ) . (30)

A small value of the p-value, say less than 5%, provides evidence against the null
hypothesis in the sense that the chance of obtaining a value of βj by pure random
variation from one sample to the next is less than 5% if, in fact, the null hypothesis
is correct. If so, the null hypothesis is rejected at the 5% significance level. Simply,
we can reject the null hypothesis at the 5% significance level if t act > 1.96 .
Therefore, if the hypothesis is accepted, then we ensue that the corresponding rule is
not important and is not influential on the cost. The estimation of the coefficients is
performed using a regression software (EViews 3.0).
7.3.6 Shortest Path in a Network

Let G = (V,A) be a graph, where V = {1,..., N } is the set of nodes, and A ⊆ V ×V is
in the graph.
It is a well-known principle that every additive deterministic dynamic
programming formulation can be equivalently viewed as the problem of finding
the shortest path in a directed network, where the states, decisions, and decision
costs of the former correspond to the nodes, arcs, and arc lengths of the latter
(Dreyfus and Law, 1977). It is perhaps for this reason that the task of efficiently
computing shortest paths is found prominence in the mathematical programming
literature. Next, we describe a dynamic programming approach for computing the
optimal path.
7.3.7 Dynamic Programming

Dynamic programming (DP) was introduced by Bellman (1957). Toth (1980)
presented the early DP-based approaches and reported numerical experiments with a
limited success. Hybrid methods, combining dynamic programming and implicit
enumeration, were developed later. The first approach was developed by Plateau and
Elkihel (1985). A recent approach, the so-called combo algorithm, is able to solve very
large instances of up to 10000 variables within less than one second, with basically no
difference in the required solution times for ‘‘easy’’ and ‘‘hard’’ instances (Martello et
al., 1999). Marsten and Morin (1978) proposed the first hybrid method, which
combined heuristic algorithms, dynamic programming and branch-and-bound
approaches. More sophisticated methods can be found in Ibaraki (1987).
Dynamic programming is a technique to tackle multistage decision processes.

A given problem is subdivided into smaller subproblems, which are sequentially
solved until the initial problem is solved by the aggregation of the subproblem
solutions. In each stage, a set of states is defined. The states would describe all
possible conditions of the process in the current decision stage, which corresponds
to every feasible partial solution. The set of all possible states is known as the state
space. The states of a stage u can be transformed to states of a stage u+1, using a
transition. A transition indicates the decisions adopted in a stage, and a sequence
of transitions taken to reach a state starting from another state is known as a
policy. Dynamic programming approaches can be seen as transformations of the
original problem to one associated with the exploration of a multistage graph
G(S,T), where the vertices in S correspond to the state space and the arcs in T
correspond to the set of transitions, leading to an optimal policy.
The basis of dynamic programming can be traced to the optimality principle of
Bellman (2003). The optimality principle states that an optimal policy should be
constituted by optimal policies from every state of the decision chain to the final state.
Here, we make use of a dynamic programming approach for our proposed
network to identify the optimal path in the manufacturing system. The advantages
of such a model are simplicity, the ability to determine the exact optimal value,
and implementability on sophisticated networks. The backward dynamic model
would be defined as:
Indices:
s Number of stages; s= 0,1,2,…, n.
i' Start node number; i'=1,2, …, I; i'=0 (for the start node).
i End node number; i=1,2, …, I.
Notations:
ϕs (i' ) The minimum value of moving from node i' in stage s to an end
node i in stage s-1.
Pi'i Numerical value of an arc from node i' to node i.
Optimal policy:
ϕ s (i' ) = Min {ϕ s +1 (i ) + Pi 'i }, ∀i ' in stage s ( s = 0,1,2,..., n).
i in layer s +1
(31)
ϕ s (i ) = (0),
ϕ * = ϕ 0 (0).
Note that ϕ * identifies the optimal path.

7.4 Conclusions
We proposed an approach for finding an optimal path in a flexible jobshop

manufacturing system. The proposed flexible jobshop system has more than one
shop with the same duty. The difference among shops with the same duty is in
their machines with various specifications. The shops configure a network in
which they are considered as nodes and the paths among them are considered as
network arcs. An Automated Guided Vehicle (AGV) functions as a material
handling device through the manufacturing network. Cost of traveling from one
node to another is uncertain and needs to be approximated. Cost is inferenced
from an expert system considering three parameters of equipment sensitivity,
operator proficiency, and product specification via linguistic variables. This cost is
associated with some main and sub factors. To approximate the cost, a back
propagation neural network was applied to estimate the sub-factors with the
corresponding main factors and gain a bound for the main factors. These bounds
were processed by the entropy technique to obtain regions for the linguistic
variables. After determining the regions for the linguistic variables, we composed
the rules and verified their significance using regression analysis. The purified
rules were used to infer the cost of the system. The objective was to find a path
minimizing the cost in the manufacturing network. A dynamic programming
approach was used to compute an optimal path in the proposed network. Finally,
the proposed approach was illustrated by a numerical example.
References
Abdi, H., Valentin, D., Edelman, B., O’Toole, A.J.: A Widrow-Hoff learning rule for a
generalization of the linear auto-associator. Journal of Mathematical Psychology 40,
175–182 (1996)
Bellman, R.: Dynamic Programming. Princeton University Press, Princeton (1957)
Bellman, R.: Dynamic Programming. Dover Publications (2003)
Ben-Arieh, D., Qian, L.: Activity-based cost management for design and development
stage. International Journal of Production Economics 83, 169–183 (2003)
Cavalieria, S., Maccarroneb, P., Pinto, R.: Parametric vs. neural network models for the
estimation of production costs: A case study in the automotive industry. International
Journal of Production Economics 91, 165–177 (2004)
Cheng, B., Titterington, D.M.: Neural networks: A review from a statistical perspective.
Statistical Science 9(1), 2–30 (1994)
Dai, J.B., Lee, N.K.S.: Economic feasibility analysis of flexible material handling systems:
A case study in the apparel industry. International Journal of Production
Economics 136(1), 28–36 (2012)
Deng, S., Yeh, T.-H.: Using least squares support vector machines for the airframe
structures manufacturing cost estimation. International Journal of Production
Economics 131(2), 701–708 (2011)
Dreyfus, S.E., Law, A.M.: The Art and Theory of Dynamic Programming. Academic Press,
New York (1977)
References 113
Dvir, D., Ben-Davidb, A., Sadehb, A., Shenhar, A.J.: Critical managerial factors affecting
defense projects success: A comparison between neural network and regression
analysis. Engineering Applications of Artificial Intelligence 19, 535–543 (2006)
Feng, C.-X., Wang, V.: Surface roughness predictive modeling: Neural networks versus
regression. IIE Transactions on Design and Manufacturing 40(3), 683–697 (2002)
Feng, C.-X., Wang, X.: Digitizing uncertainty modeling for reverse engineering
applications: Regression versus neural networks. Journal of Intelligent
Manufacturing 13(3), 189–199 (2002)
Finnie, G.R., Wittig, G.E., Desharnais, J.-M.: A comparison of software effort estimation
techniques: Using function points with neural networks, case-based reasoning and
regression models. Journal of Systems and Software 39(3), 281–289 (1997)
Gallinari, P., Thiria, S., Badran, F., Fogelman-Soulie, F.: On the relations between
discriminant analysis and multilayer perceptrons. Neural Networks 4(3), 349–360
(1991)
Gayretli, A., Abdalla, H.S.: A featured based prototype system for the evaluation and
optimization of manufacturing processes. Computers & Industrial Engineering 37, 481–
484 (1999)
Ghasemzadeh, S., Behrangi, E., Azgomi, M.A.: Conflict-free scheduling and routing of
automated guided vehicles in mesh topologies. Robotics and Autonomous Systems 30,
738–748 (2009)
Heiat, A.: Comparison of artificial neural network and regression models for estimating
software development effort. Information and Software Technology 44(15), 911–922
(2002)
Ibaraki, T.: Enumerative approaches to combinatorial optimization–Part II. Annals of
Khayat, G.E., Lagevin, A., Riopel, D.: Integrated production and material handling
scheduling using mathematical programming and constraint programming. European
Kim, G.-H., An, S.-H., Kang, K.-I.: Comparison of construction cost estimating models
based on regression analysis, neural networks, and case-based reasoning. Building and
Environment 39(10), 1235–1242 (2004)
(1995)
Krishnaswamy, M., Krishnan, P.: Nozzle wear rate prediction using regression and neural
network. Biosystems Engineering 82(1), 53–64 (2002)
Lee, J., Um, K.: A comparison in a back-bead prediction of gas metal arc welding using
multiple regression analysis and artificial neural network. Optics and Lasers in
Engineering 34(3), 149–158 (2000)
Marsten, R.E., Morin, T.L.: A hybrid approach to discrete mathematical programming.
Mathematical Programming 14, 21–40 (1978)
Martello, S., Pisinger, D., Toth, P.: New trends in exact algorithms for the 0–1 knapsack
problem. European Journal of Operational Research 123, 325–336 (1999)
McKim, R.A.: Neural network applications to cost engineering. Cost Engineering 35, 31–
35 (1993)
Engineering 128, 563–575 (2006)
Nolan, A.: An intelligent system for case review and risk assessment in social services. AI
Mag. 19(1), 39–46 (1998)
Pendharkar, P.C.: Scale economies and production function estimation for object-oriented
software component and source code documentation size. European Journal of
Operational Research 172, 1040–1050 (2006)
Plateau, G., Elkihel, M.: A hybrid method for the 0–1 knapsack problem. Methods of
Quintana, G., Ciurana, J.: Cost estimation support tool for vertical high speed machines
based on product characteristics and productivity requirements. International Journal of
Production Economics 134(1), 188–195 (2011)
Rehman, S., Guenov, M.D.: A methodology for modeling manufacturing costs at
conceptual design. Computers & Industrial Engineering 35, 623–626 (1998)
Ripley, B.D.: Neural networks and related methods for classification. Journal of the Royal
Statistical Society, Series B (Methodological) 56(3), 409–456 (1994)
Roy, R., Souchoroukov, P., Shehab, E.: Detailed cost estimating in the automotive industry:
Data and information requirements. International Journal of Production
Economics 133(2), 694–707 (2011)
Sarle, W.S.: Neural networks and statistical models. In: Proceedings of the 19th Annual
SAS Users Group International Conference, pp. 215–221 (1994)
Schumacher, M., Robner, R., Vach, W.: Neural networks and logistic regression: Part I.
Computational Statistics and Data Analysis 21(6), 661–682 (1996)
Setyawati, B.R., Sahirman, S., Creese, R.C.: Neural networks for cost estimation. AACE
International Transactions, EST13, 13.1–13.8 (2002)
Shehab, E.M., Abdalla, H.S.: Manufacturing cost modeling for concurrent product
development. Robotics and Computer- Integrated Manufacturing 17, 341–353 (2001)
Shuhui, L., Wunsch, D.C., O’Hair, E., Giesselmann, M.G.: Comparative analysis of
regression and artificial neural network models for wind turbine power curve
estimation. Journal of Solar Energy Engineering 123, 327–332 (2001)
Smith, A.E., Mason, A.K.: Cost estimation predictive modeling: regression versus neural
network. The Engineering Economist 42(2), 137–161 (1997)
Toth, P.: Dynamic programming algorithms for the zero–one knapsack problem.
Computing 25, 29–45 (1980)
Turban, E., Aronson, J.: Decision Support Systems and Intelligent Systems. Prentice-Hall,
London (1998)
Vach, W., Robner, R., Schumacher, M.: Neural networks and logistic regression: Part II.
Vis, I.F.A.: Survey of research in the design and control of automated guided vehicle
systems. European Journal of Operational Research 170, 677–709 (2006)
Vrbsky, S.V.: A data model for approximate query processing of real time database. Data &
Knowledge Engineering 21, 79–102 (1997)
Warner, B., Misra, M.: Understanding neural networks as statistical tools. The American
Statistician 50(4), 284–293 (1996)
Weiss, N., Kulikowski, I.: Computer Systems That Learn. Morgan-Kaufmann, California
(1991)
Wu, N., Zhou, M.: Modeling and deadlock avoidance of automated manufacturing systems
with multiple automated guided vehicles. IEEE Transactions on Systems Man and
Cybernetics B 35, 1193–1201 (2005)
Yang, C.O., Lin, T.S.: Developing an integrated framework for feature-based early
manufacturing cost estimation. International Journal of Advanced Manufacturing
Technology 13, 618–629 (1997)
References 115
Yesilnacar, E., Topal, T.: Landslide susceptibility mapping: A comparison of logistic

regression and neural networks methods in a medium scale study, Hendek region
(Turkey). Engineering Geology 79(3-4), 251–266 (2005)
Zhang, Y.F., Fuh, J.Y.H.: A neural network approach for early cost estimation of packaging
products. Computers & Industrial Engineering 34, 433–450 (1998)
Gu, Z., Lam, L.H., Dhurjati, P.S.: Feature correlation method for enhancing fermentation
development: A comparison of quadratic regression with artificial neural networks.
Computers and Chemical Engineering 20(suppl. 1), S407–S412 (1996)
Zimmermann, H.J.: Fuzzy Set Theory and its Applications, 3rd edn. Kluwer Academic
Publishers (1996)
Chapter 8
Optimal Path for AGV System with Intelligent
Agents
8.1 Summary
An AGV is a material handling equipment that travels on a network of guide

paths. The guide path is composed of aisle segments on which the vehicles are
assumed to travel at a constant speed. The vehicles can travel forward or
backward. As many vehicles travel on the guide path simultaneously, collisions
must be avoided. AGV systems are implemented in various industrial contexts:
container terminals, part transportation in heavy industry, flexible manufacturing
systems.
The most elementary version of the vehicle routing problem is the capacitated
vehicle routing problem (CVRP). The CVRP is described as follows: n customers
must be served from a unique depot. Each customer asks for a quantity qi of goods
(i = 1,..., n) and a vehicle of capacity Q is available to deliver goods. Since the
vehicle capacity is limited, the vehicle has to periodically return to the depot for
reloading. In the CVRP, it is not possible to split customer delivery. Therefore, a
CVRP solution is a collection of tours where each customer is visited only once
and the total tour demand is at most Q.
From a graph theoretical point of view the CVRP may be stated as follows: Let
G = (C,L) be a complete graph with node set C = (co, c1, c2,..., cn) and arc set L =
(ci, cj): ci , cj ∈ C, i < j. In this graph model, co is the depot and the other nodes are
the customers to be served. Each node is associated with a fixed quantity qi of
goods to be delivered (a quantity qo = 0 is associated to the depot co). To each arc
(ci, cj) is associated a value tij representing the travel time between ciand cj. The
goal is to find a set of tours of minimum total travel time. Each tour starts from
and terminates at the depot co, each node ci(i = 1,..., n) must be visited exactly
once, and the quantity of goods to be delivered on a route should never exceed the
vehicle capacity Q.
An important extension of the CVRP is the vehicle routing problem with time
windows (VRPTW). In addition to the mentioned CVRP features, this problem
includes, for the depot and for each customer ci (i = 0,..., n) a time window [bi, ei]

118 8 Optimal Path for AGV System with Intelligent Agents
during which the customer has to be served (with b0 the earliest start time and e0
the latest return time of each vehicle to the depot). The tours are performed by a
fleet of l identical vehicles. The additional constraints are that the service
beginning time at each node ci (i = 1,..., n) must be greater than or equal to bi, the
beginning of the time window, and the arrival time at each node ci must be lower
than or equal to ei, the end of the time window. In case the arrival time is less than
bi, the vehicle has to wait till the beginning of the time window before starting
servicing the customer. In the literature the fleet size l is often a variable and a
very common objective is to minimize l. Usually, two different solutions with the
same number of vehicles are ranked by alternative objectives such as the total
traveling time or total delivery time (including waiting and service times).
For a general review on AGV problems, the reader is referred to (Co and
Tanchoco, 1991; King and Wilson, 1991; Ganesharajah et al., 1998). For a recent
review on AGVs scheduling and routing problems and issues, the reader is
referred to the survey of Qiu et al. (2002). These authors identified three types of
algorithms for AGVs problems:
(1) for general path topology,
(2) for path optimization and
(3) for specific path topologies.
Methods of the first type can be divided in three categories:
(1a) static methods, where an entire path remains occupied until a vehicle
completes its route; (1b) time-window based methods, where a path segment may
be used by different vehicles during different time-windows;
and (1c) dynamic methods, where the utilization of any segment of path is
dynamically determined during routing rather than before as with categories (1a)
and (1b). The method presented in this article belongs to the third category (1c)
and addresses the conflict free routing problem with an optimization approach.
A number of authors have addressed the conflict free routing problem with a
static transportation requests set, i.e., with all requests known a priori. Lee et al.
(1998) present a two-staged traffic control scheme to solve a conflict free routing
problem. Their heuristic method consists of generating off-line k-shortest paths in
the first stage before the on-line traffic controller picks a conflict free shortest path
whenever a dispatch command for an AGV is issued (second stage). Rajotia et al.
(1998) propose a semi-dynamic time window constrained routing strategy. They
use the notions of reserved and free time windows to manage the motion of
vehicles. Krishnamurthy et al. (1993) propose an optimization approach. Their
objective is to minimize the make span. They assume that the assignment of tasks
to AGVs is given and they solve the routing problem by column generation (Ho
and Wu, 2002). Their method generates very good solutions in spite of the fact
that it is not optimal (column generation is performed at the root node of the
8.1 Summary 119
search tree only). Oboth et al. (1999) present a heuristic method to solve the
dispatching and routing problems but not simultaneously. Scheduling is performed
first and a sequential path generation heuristic (SPG) is used to generate conflict
free routes. The SPG is inspired from Krishnamurthy et al. (1993) static version of
the AGV routing problem and applied to a dynamic environment while relaxing
some of the limiting assumptions like equal and constant speeds of AGVs. When
conflict is encountered, no feedback is sent to the scheduling module. The AGV
being routed has to be delayed if an alternate route cannot be generated (Lin et al.,
2002).
The authors use rules for positioning idle AGVs instead of letting the system
manage them. Langevin et al. (1996) propose a dynamic programming based
method to solve exactly instances with two vehicles. They solve the combined
problem of dispatching and conflict free routing. Desaulniers et al. (2003) propose
an exact method that enables to solve instances with up to four vehicles. Their
approach combines a greedy search heuristic (to find a feasible solution and set
bound on delays), column generation and a branch and cut procedure. Their
method presents however some limits since its efficiency depends highly on the
performance of the starting heuristic. If no feasible solution is found by the search
heuristic, then no optimal solution can be found. The search heuristic performs
poorly when the level of congestion increases.
The problem considered in this chapter is a generalization of the Vehicle
Routing Problem (VRP) and the Pickup and Delivery Problem (PDP) see Cordeau
et al. (2004), Mitrović-Minić (1998) and secondary literature given there. The
most widely studied vehicle routing problems are the capacitated VRP and the
Vehicle Routing Problem with Time Windows (VRPTW) which are surveyed by
Laporte and Semet (2002) and Cordeau et al. (2002). Efficient methods for
handling complex side constraints in insertion methods are presented in Campbell
and Savelsbergh (2004). Comprehensive surveys on construction methods,
neighborhood search methods, and meta-heuristics for the VRPTW are given by
Bräysy and Gendreau (2005a) and Bräysy and Gendreau (2005b). Variable
Neighborhood Search (VNS) is a meta-heuristic based on the idea of
systematically changing the neighborhood structure during the search, see
Mladenovi ć and Hansen (1997) and Hansen and Mladenović (2003). VNS
systematically exploits the following observations:
a) a local optimum with respect to one neighborhood structure is not necessary
so for another;
b) a global optimum is a local optimum with respect to all possible
neighborhood structures;
c) for many problems local optima with respect to one or several
neighborhoods are relatively close to each other.
A recent example of a VNS algorithm for vehicle routing problems is the

algorithm for the multi-depot VRPTW presented by Polacek et al. (2004). Large
Neighborhood Search (LNS) has been introduced for the VRPTW by Shaw (1997)
and can be interpreted as a special case of VNS. Kilby et al. (2000) have shown
that LNS is well suited for rich VRP.
In many cases it is assumed that transportation requests are accepted before
planning begins and tours are generated assuming that all accepted transportation
requests must be served (Lin et al., 2004). Work regarding load acceptance issues
for the Traveling Salesman Problem (TSP) has been surveyed by Feillet et al.
(2005), but only few attempts have been made to tackle extensions of this
problem, for example, by Schönberger et al. (2002). VRP with multiple pickup
and delivery locations have been studied by Savelsbergh and Sol (1995),
Savelsbergh and Sol (1998) and Hasle (2003). A comprehensive discussion of
dynamic vehicle routing can be found in Psaraftis (1988) and Psaraftis (1995).
Dynamic real-life problems often require rich models, in most of the literature on
dynamic routing problems; however, some simplifying assumptions are made. For
example, in the dynamic full-truckload PDP, which recently has received
increasing attention, see Fleischmann et al. (2004), Yang et al. (2004), and Powell
et al. (2000), each vehicle can only carry one transportation request at a time and
cannot load further shipments until all currently loaded shipments are unloaded.
The only work known to the authors regarding rich VRP in a dynamic context is
presented by Savelsbergh and Sol (1998). A column generation approach is used
to solve the General Pickup and Delivery Problem (GPDP) presented by
Savelsbergh and Sol (1995).

In a material requirement planning (MRP) for a product some shops are existed to
provide the needed materials. According to the bill of material, an AGV is applied
to provide the materials for the final assembly in the master shop. Considering the
hierarchical structure of the BOM, shops are in levels. For each level some shops
are existed being differentiated by the cost, time, and AGV capability. In this way
from each shop of a level to the next shop from another level some paths are
considered. As it is stated, the difference between paths is in cost, time, and AGV
capability. Hence, the AGV should decide to select the path to minimize time,
cost, and AGV capability, concurrently. The materials are assembled from the
lowest level of the BOM and lead to the final product. Each shop in each level can
satisfy a corresponding level of the proposed BOM. The proposed BOM is
illustrated in Figure 1.
Fig. 1 The Proposed BOM
Based on the hierarchical structure of the BOM, the shops are in a level
structure each of which provides the materials in the certain level of the BOM.
The AGV would begin to assemble the materials from shops in varied levels to
reach the final product in the last stage in the main shop. The model is looking for
a path that fulfills the BOM and also is optimal considering the decision
parameters. A configuration of the shops in the Material Requirement Planning is
presented in Figure 2.
Fig. 2 A Configuration of the Shops in the Material Requirement Planning

AGV needs to decide which route to select to satisfy the objectives of the
problem. In the stated condition, AGV should consider all time, cost, and AGV
capability to select a route i.e. the lowest cost, shortest time, and the lowest
occupied AGV capability must be regarded. To achieve this goal, mathematical
programming approach is applied for optimization. To find out the cost, time, and
capability parameters intelligent agents are proposed. Intelligent agents help AGV
to evaluate all paths.
In artificial intelligence, an intelligent agent (IA) is an autonomous entity which
observes and acts upon an environment (i.e. it is an agent) and directs its activity
towards achieving goals (i.e. it is rational). Intelligent agents may also learn or use
knowledge to achieve their goals. They may be very simple or very complex: a
reflex machine such as a thermostat is an intelligent agent, as is a human being, as
is a community of human beings working together towards a goal.
Intelligent agents are often described schematically as an abstract functional
system similar to a computer program. For this reason, intelligent agents are
sometimes called abstract intelligent agents (AIA) to distinguish them from their
real world implementations as computer systems, biological systems, or
organizations. Some definitions of intelligent agents emphasize their autonomy,
and so prefer the term autonomous intelligent agents. Still others considered goal-
directed behavior as the essence of intelligent and so prefer a term borrowed from
economics, "rational agent".
Intelligent agents in artificial intelligence are closely related to agents in
economics, and versions of the intelligent agent paradigm are studied in cognitive
science, ethics, the philosophy of practical reason, as well as in many
interdisciplinary socio-cognitive modeling and computer social simulations.
Fig. 3 A configuration of the intelligent agent and information transfer process

Intelligent agents are also closely related to software agents (an autonomous
software program that carries out tasks on behalf of users). In computer science,
the term intelligent agent may be used to refer to a software agent that has some
intelligence, regardless if it is not a rational agent.
In this chapter, we propose three intelligent agents for time, cost and capability,
respectively. The core of these agents is optimization. The results of all agents are
transferred to a computer data base which conducts the AGV through guide path
in the shop floor. A configuration of the intelligent agent and information transfer
process is shown in Figure 3.
The AGV collects the information and process them in the separate intelligent
agents. The results are sent to the data base and view by the controller. The
controller issue required orders to the AGV.
To estimate the time of traveling from one shop (node) to another one in our
proposed network, a mathematical optimization is worked out by an intelligent
agent. While the velocity (v) of AGV is known, the time is computed as follows:
d = v.t , (1)
Where d is the movement distance, v is AGV velocity and t is the time of

movement. Due to a Cartesian coordinates structure as shown in Figure 4.
Fig. 4 Cartesian coordinates for AGV movement

to compute the distance we use,
d = ( x − x0 ) 2 + ( y − y0 ) 2 , (2)
wherex is the horizontal movement and y is the vertical movement. As a result, we

have,
d
t= . (3)
v
Therefore, time is computed and transferred to the central computer.
Cost is integrated with the time computed from last section. Our proposed
intelligent agent use the following formulae to estimate the cost of moving from
one shop to another one,
cost = ( variable cost × d ) + fix cost , (4)
Variable cost depends on the distance that the AGV move (such as, defect,
amortization, etc.). But the fix cost is independent of the distance and include set
up cost, controller salary, etc.
Here, we want to estimate the AGV capability during the assembly process.
Some parameters affect the capability of an AGV. We should investigate whether
a parameter is significant on capability or not. One way to survey the effect of
Therefore, we consider the following equation which is computed in our proposed
intelligent agent for capability,
ui = β 0 + β1 r1 + ... + β j r j + ε j , i = 1,..., n; j = 1,..., m. (5)
where β0 is the intercept, β j s are the coefficients for rjparameters and ε is the
error term. The aim is identifying the β j s which are not important on capability
(ui). Here, we apply two-sided hypothesis testing as follows;
H 0 : β j = 0;
(6)
H 1 : β j ≠ 0.
The test of the null hypothesis H0 against the two sided alternative proceeds as in
the three steps. The first is to compute the standard error of β j , SE ( β j ) . The
standard error of βj is an estimator of σβ j
, the standard deviation of the
sampling distribution of βj.

m
1
× ∑ (r j − r ) 2 ε 2j
1 m − 2 j =1
σ β2 = × 2
, (7)
m ⎡1 2⎤
j
m
⎢ ∑ (r j − r ) ⎥
⎣ m j =1 ⎦
SE ( β j ) = σ β2 j . (8)

is complicated, in applications the standard error is
computed by regression software. The second step is to compute the t-statistic,
βj −0
t= (9)
SE ( β j )
act
⎡ βj −0 β act −0⎤
[ ]
p − value = PrH 0 β j − 0 > β jact − 0 ⇒ PrH 0 ⎢
β
>
j
β
⎥ ⇒ PrH 0 ( t > t act ) (10)
⎣⎢ SE ( j ) SE ( ⎥
j) ⎦
Where PrH 0 denotes the probability computed under the null hypothesis, the
second equality follows by dividing by SE ( β j ) , and t act is the value of the
t-statistic actually computed. Because β j is approximately normally distributed in
large samples, under the null hypothesis the t-statistic is approximately distributed
as a standard normal random variable, so in large samples,
p − value = Pr( Z > t act ) = 2Φ(− t act ) (11)
A small value of the p-value, say less than 5%, provides evidence against the null
hypothesis in the sense that the chance of obtaining a value of βj by pure random
variation from one sample to the next is less than 5% if, in fact, the null hypothesis
is correct. If so, the null hypothesis is rejected at the 5% significance level. Simply
we can reject the null hypothesis at the 5% significance level if, t act > 1.96 .
Therefore, if the hypothesis is accepted then we ensue that the corresponded
parameter is not important and has no effect on the capability.
While we determined the effective parameters on capability, we can find the

regression value for capability omitting the noneffective parameters.

Considering the network nature of the problem and the objective which is
identifying shortest path or route, mathematical programming approach is applied.
The network of the proposed problem is represented in Figure 5.
i’ )
Pi
i ',
,Ti
ii'
(C
Fig. 5 The Network of the Problem
All nodes indicate the shops except node 0 which shows the AGV. Arcs are
identified with three criteria, cost, time, and AGV capability which are
consequenced from the intelligent agents. Number of nodes are N and depends on
the number of the shops that provide materials. Nodes are in levels as stated
before. Mathematical programming approach in details is as follows.
Indices:
i Start node i=0,1,2,….,N
i' End node i'=1,2,…,N
j Level number j=0,1,2,….J
Notations:
Cii’ Cost of moving from node i to node i'

Tii' Time duration of moving from node i to node i'
ϕ j (i) The minimum value of moving from the start node i to the end nod i'
in layer j
N Number of nodes
δ Maximum capability of an AGV

τ Maximum time duration in hand to satisfy the demands
B Maximum budget in hand
Pii’ Capability of an AGV for moving from node i to node i’
⎧1 if path ii' is selected

ξ ii ′ = ⎨
⎩0 otherwise
Objective Function:
ϕ j (i ′) = Min {ϕ j −1 (i) + Cii′ } and (12)

i ≠ i′
Min {ϕ j −1 (i) + Tii′ }and (13)

i ≠ i′
Min {ϕ j −1 (i) + Pii′ } (14)

i ≠ i′
Subject to:
ϕ 0(0) = 0 (15)
N N
∑∑ ξ
i′ i
ii ′ × Cii′ ≤ B (16)
N N
∑∑ ξ
i′ i
ii ′ × Tii′ ≤ τ (17)
N N
∑∑ ξ
i′ i
ii ′ × Pii ' ≤ δ (18)
Tii′ , Cii′ , Pii ' ≥ 0 (19)
Equations (12), (13), and (14) are the objective functions which aim to find the
shortest path based on cost, time, and AGV capability, respectively. Equation (15)
is the primary condition for starting the travel on nodes. Equation (16) guarantees
that our travels won’t exceed our accessible budget. Equation (17) represents that
the time needed to satisfy the demands is confined. Equation (18) guarantees the
capability of AGV during the travels. Equation (19) is the sign relation.
As it is cleared, the objective functions cause the decision maker not to be able
to decide at once about time and cost and capability simultaneously. Hence, we do
some changes on objective functions to mix them in to a single function that take
into account all time and cost and capability as follows.
8.3.1 Normalization Process

To transform the above three objective functions (12, 13 and 14) the normalization
is required for making it possible to obtain time, cost and AGV capability,
simultaneously. The normalization process is as follows:
Tii′
Tiin′ = (20)
N N
∑∑ T
i '=1 i =0
2
ii '
n
where T is the normalized value of time duration for each arc.
ii′
The same procedure is existed for costs:
C ii ′
C ini′ = (21)
N N
∑∑ C
i '=1 i = 0
2
ii '
n
where C ii ′ is the normalized value of cost for each arc.
The same procedure is existed for AGV capability:
Pii′
Piin′ = (22)
N N
∑∑ P
i '=1 i =0
2
ii '
By the means of normalization the objective functions could be transformed to a

single one as follows:
ϕ j (i ′) = Min {ϕ j −1 (i) + Tiin′ + Cini′ + Piin′ } (23)
Equation (23) does time-cost-AGV capability optimization at the same time and
its result is the optimum combination of time, cost, and AGV capability which
ensue to the optimal path or route. Also, note that their importance weights are
considered to be equal. But, if a decision maker tends to consider different
importance for time, cost and capability, it is easily performed using common
weighing approaches in the literature.
Algorithm for Solving

The following heuristic algorithm could be applied for solving the proposed
problem in some stages.
Step0- Identify number of nodes (factories)
Step1- Identify Tii' ,Cii' and, Pii'

Step2- Normalizing process
Step3- Selecting the best combination of time, cost, and capability
Step4- Satisfying the constraints
Step5- Finding the optimal path (route)
Step6- Termination
The flow chart of the heuristic algorithm is illustrated in Figure 6.
Number of N
Identify Tii' ,Cii'andPii'
Normalizing
Process
Selecting best
time and cost
No
If constraints
are satisfied
Yes
Finding the
Optimal Path
Termination
Fig. 6 The Flowchart of the Solving Algorithm

Regarding to the flowchart, in termination stage a path with all optimal time,
cost, and AGV capability is identified at the same time.
8.4 Conclusions
The obtained results extracted from the proposed methodology in a material
handling system are helpful in providing better managerial insights for decision
making in AGV planning and guide path maintenance. While a path contains
several arcs in the proposed material handling network determining the arcs
especially the most used ones helps fortifying them and prohibit halt in the system
due to AGV or path break downs. Another managerial aspect is to make real time
decisions underlying the proposed agents as decision supports. While demand and
material supplies are dynamic therefore AGV planning should also be dynamic.
Therefore the proposed agents help to make decisions at any time with respect to
any demand fulfilled and material provision. We have considered an extension of
the classical vehicle routing problem in which three dimensional packing
constraints were introduced i.e. time, cost, and AGV capability. This problem
featured three classical combinatorial optimization problems. Intelligent agents
were designed for the triple criteria which used optimization tools to estimate the
values and report them to the computer data base. We have applied a mathematical
programming approach for optimizing the problem. By a normalization process all
time, cost, and AGV capability were included as the arc’s value of the proposed
network. The effectiveness of the proposed algorithm was validated using two
illustrative examples.
References
Bräysy, O., Gendreau, M.: Vehicle routing problem with time windows, Part I: Route
construction and local search algorithms. Transportation Science 39(1), 104–118
(2005a)
Bräysy, O., Gendreau, M.: Vehicle routing problem with time windows, Part II:
Metaheuristics. Transportation Science 39(1), 119–139 (2005b)
Campbell, A., Savelsbergh, M.: Efficient insertion heuristics for vehicle routing and
scheduling problems. Transportation Science 38(3), 369–378 (2004)
Co, C.G., Tanchoco, J.M.A.: A review of research on AGVs vehicle management.
Engineering Costs and Production Economics 21, 35–42 (1991)
Cordeau, J.-F., Desaulniers, G., Desrosiers, J., Solomon, M., Soumis, F.: VRP with time
windows. In: Toth, P., Vigo, D. (eds.) The Vehicle Routing Problem, pp. 157–193.
SIAM Monographs on Discrete Mathematics and Applications, Philadelphia (2002)
Cordeau, J.-F., Gendreau, M., Hertz, A., Laporte, G., Sormany, J.-S.: New heuristics for the
vehicle routing problem. Les cahiers du GERAD G-2004-33, Université de Montréal
HEC, Montréal, Canada (2004)
References 131
Desaulniers, G., Langevin, A., Riopel, D., Villeneuve, B.: Dispatching and conflict-free
routing of automated guided vehicles: an exact approach. The International Journal of
Flexible Manufacturing Systems 15, 309–331 (2003)
Feillet, D., Dejax, P., Gendreau, M.: Traveling Salesman Problems with Profits.
Transportation Science 39(2), 188–205 (2005)
Fleischmann, B., Gnutzmann, S., Sandvoß, E.: Dynamic vehicle routing based on on-line
traffic information. Transportation Science 38(4), 420–433 (2004)
Ganesharajah, T., Hall, N.G., Sriskandarajah, C.: Design and operational issues in AGV-
served manufacturing systems. Annals of Operations Research 76, 109–154 (1998)
Hansen, P., Mladenović, N.: A tutorial on Variable Neighborhood Search. Les cahiers du
GERAD G-2003- 46, Université de Montréal HEC, Montréal, Canada (2003)
Hasle, G.: Heuristics for rich VRP models. Presented at the Seminar at GERAD, Montréal,
Canada (2003)
Ho, Y.C., Wu, F.C.: A bidding-based control strategy for multiple-load automated guided
vehicles. Journal of the Chinese Institute of Industrial Engineers 19(4), 82–94 (2002)
Kilby, P., Prosser, P., Shaw, P.: A comparison of traditional and constraint based heuristic
methods on vehicle routing problems with side constraints. Constraints 5, 389–414
(2000)
King, R.E., Wilson, C.: A review of automated guided vehicle system design and
scheduling. Production Planning and Control 2, 44–51 (1991)
Krishnamurthy, N.N., Batta, R., Karwan, M.H.: Developing conflict-free routes for
automated guided vehicles in a flexible manufacturing system. Operations Research 41,
1077–1090 (1993)
Langevin, A., Lauzon, D., Riopel, D.: Dispatching, routing and scheduling of two
automated guided vehicles in a flexible manufacturing system. International Journal of
Laporte, G., Semet, F.: Classical heuristics for the capacitated VRP. In: Toth, P., Vigo, D.
(eds.) The Vehicle Routing Problem, pp. 109–128. SIAM Monographs on Discrete
Mathematics and Applications, Philadelphia (2002)
Lee, J.H., Lee, B.H., Choi, M.H.: A real-time traffic control scheme of multiple AGV
systems for collision free minimum time motion: a routing table approach. IEEE
Transactions on Systems, Man and Cybernetics—Part A: Systems and Humans 28, 347–
358 (1998)
Lin, J.T., Wang, F.K., Young, J.R.: Virtual vehicle in the connecting transport automated
material-handling system (AMHS). International Journal of Production Research 42(13),
2599–2610 (2004)
Lin, J.T., Wu, C.K., Yang, C.J.: Vehicle management of AMHS in 300 MM wafer FAB.
Journal of the Chinese Institute of Industrial Engineers 19(4), 1–10 (2002)
Mitrović-Minić, S.: Pickup and delivery problem with time windows: A survey. Technical
report TR, School of Computing Science, Simon Fraser University, Burnaby, BC,
Canada (1998)
Mladenović, N., Hansen, P.: Variable Neighborhood Search. Computers and Operations
Research 24, 1097–1100 (1997)
Oboth, C., Batta, R., Karwan, M.: Dynamic conflict-free routing of automated guided
vehicles. International Journal of Production Research 37, 2003–2030 (1999)
Polacek, M., Hartl, R., Doerner, K., Reimann, M.: A variable neighborhood search for the
multi depot vehicle routing problem with time windows. Journal of Heuristics 10, 613–
627 (2004)
Powell, W., Snow, W., Cheung, R.: Adaptive labeling algorithms for the dynamic
assignment problem. Transportation Science 34(1), 50–66 (2000)
Psaraftis, H.: Dynamic vehicle routing problems. In: Golden, B., Assad, A. (eds.) Vehicle
Routing: Methods and Studies, pp. 233–248. North-Holland, Amsterdam (1988)
Psaraftis, H.: Dynamic vehicle routing: Status and prospects. Annals of Operations
Research 61, 143–164 (1995)
Qiu, L., Hsu, W.-J., Wang, H.: Scheduling and routing algorithms for AGVs: a survey.
International Journal of Production Research 40, 745–760 (2002)
Rajotia, S., Shanker, K., Batra, J.L.: A Semi-dynamic window constrained routing strategy
in an AGV system. International Journal of Production Research 36, 35–50 (1998)
Savelsbergh, M., Sol, M.: The general pickup and delivery problem. Transportation
Science 29(1), 17–30 (1995)
Savelsbergh, M., Sol, M.: DRIVE: dynamic routing of independent vehicles. Operations
Research 46, 474–490 (1998)
Shaw, P.: A new local search algorithm providing high quality solutions to vehicle routing
problems. Technical report, APES group, Department of Computer Sciences, University
of Strathclyde, Glasgow, Scottland (1997)
Yang, J., Jaillet, P., Mahmassani, H.: Real-time multi-vehicle truckload pickup-and-
delivery problems. Transportation Science 38(2), 135–148 (2004)
Chapter 9
Earliness/Tardiness for a Multiple AGV System
9.1 Summary
Flexible manufacturing systems (FMSs), container terminals, warehousing systems,

and service industries including hospital transportations are employing automated
guided vehicle systems (AGVs) for the material handling to maintain flexibility and
efficiency of production and distribution. For the efficient operation, it is requested
to realize the synchronized operations for the simultaneous scheduling of production
systems and transportation systems. The main issue treated in this chapter is the
simultaneous optimization problems for penalized earliness and tardiness for the
AGVs in the manufacturing system. The production scheduling problems asks an
optimal production sequence and starting time of operations for jobs at machines for
multi-stages with respect to a specified technical precedence relation. The vehicle
management problems are classified into:
(1) dispatching, which is to assign tasks to vehicles;
(2) routing, which is to select specific paths taken by vehicles;
(3) scheduling, which is to determine the arrival and departure times.
Unlike the classical vehicle routing problem (VRP) formulation, conflict-free
constraints should be considered for the routing of AGVs for semiconductor
fabrication. The interaction between production and transportation control is
discussed by Mantel and Landerweerd (1995). In the flowshop production
systems, the production and transportation schedules are usually controlled by a
pull type of policy in case of fork lifts or conveyor systems. However, for FMSs
environment with AGV systems, the optimal machine schedules highly depend on
the selection of dispatching and routing because it is extremely difficult to predict
the transportation time when the conflicts and interferences between vehicles
cannot be neglected.
Automated guided vehicle (AGV) is a material handling equipment traveling
on a network of guide paths. The FMS is a configuration of various shops, also
called working stations, each with a specific function such as milling, washing, or
assembly. Each shop is connected to the guide path network by a pick-up/delivery
(P/D) station where pallets are transferred from/to the AGVs. Pallets of products

134 9 Earliness/Tardiness for a Multiple AGV System
are moved between the shops by the AGVs. The guide path is composed of aisle
segments on which the vehicles are assumed to travel at a constant speed. The
vehicles can travel forward or backward. As many vehicles travel on the guide
path simultaneously, collisions must be avoided. AGV systems are implemented
in various industrial contexts: container terminals, part transportation in heavy
industry, flexible manufacturing systems. For a general review on AGV problems,
the reader is referred to (Co and Tanchoco, 1991; King and Wilson, 1991;
Ganesharajah et al., 1998). For a recent review on AGVs scheduling and routing
problems and issues, the reader is referred to the survey of Qiu et al. (2002). These
authors identified three types of algorithms for AGVs problems:
(1) for general path topology,
(2) for path optimization,
(3) for specific path topologies.
Methods of the first type can be divided in three categories:
(1a) static methods, where an entire path remains occupied until a vehicle
completes its route;
(1b) time-window based methods, where a path segment may be used by different
vehicles during different time-windows;
(1c) dynamic methods, where the utilization of any segment of path is
dynamically determined during routing rather than before as with categories (1a)
and (1b).
This chapter addresses a penalized earliness and tardiness scheduling problem
for AGVs in a manufacturing system.
Scheduling problems arise in areas as diverse as production planning, personnel
planning, product configuration, and transportation. An overview of the wide range
of constraints in scheduling, together with the most powerful propagation algorithms
for these constraints are given (Baptiste et al., 2001; Baptiste et al., 1995).
Production scheduling, dispatching, routing and scheduling decisions for AGVs
can be made simultaneously or separately. Most of the literature treats one or two
of the problems at the same time. An extensive review has been addressed by Vis
(2006) for operational control of AGVs. A widely used technique for dispatching
is the simulation. The heuristic rules are used in on-line control systems. For
routing and scheduling of AGVs, several techniques have been used to maximize
the total system performance taking in to account for deadlock or conflicts for
AGVs. Kim and Tanchoco (1991) studied the problem of finding conflict-free
routes in a bi-directional network. The algorithm is based on the shortest path
methods through the concept of time-window graph. Petri net is used to analyze
deadlock and conflict-free conditions (Dotoli and Fanti, 2004; Wu and Zhou,
2005). Singh and Tiwari (2002) presented an intelligent agent framework to find a
conflict-free shortest-time path. Nishi et al. (2005) provided a mathematical model
for routing problem. Lagrangian decomposition technique was used solve the
problem. Ghasemzadeh et al. (2009) presented a conflict-free scheduling and
9.1 Summary 135
routing in mesh topologies. It can generate the shortest path for scheduling
predicting conflicts and select another path in the case of failure.
The literature discussed above on scheduling of AGVs hardly considers the
capacity constraints of the machines where transportation jobs become available
and sequencing of operations at the machines. The simultaneous production
scheduling and transportation routing problem is one of the difficult joint
problems. The problems for AGVs have been studied mostly in operations
research and/or FMS literature. A common approach for FMS scheduling is based
on the discrete event simulation with dispatching rules (Tunali, 1998).
Lacomme et al. (2005) introduced a branch and bound algorithm coupled with
discrete event simulation. Blazewicz et al. (1994) addressed the two steps
algorithm for integrating machine scheduling and the conflict-free routing
problems. In their approach, the production scheduling and routing problems are
solved separately. Bilge and Ulusoy (1995) developed a time-window approach to
solve the simultaneous scheduling of machines and material handling in FMSs.
They formulated the problem as a mixed integer programming problem. Ulusoy et
al. (1997) and Jerald et al. (2006) dealt with the application of the genetic
algorithm on the problem. Khayat et al. (2006) studied an integrated method with
mixed integer linear programming (MILP) and constrained programming. Never
the less in their model, vehicles can always select a shortest path from a machine
station to another machine station without the consideration of conflict and
collision on the detailed routing for vehicles. Correá et al. (2007) proposed an
integrated scheduling of dispatching and vehicle routing with the consideration of
conflict-free path selection, but it does not take into account the scheduling of
machines and vehicles simultaneously.
In the above literature, it is extremely difficult to consider production
scheduling and conflict-free routing because the number of decision variables is
significantly increased. Therefore, the conventional decomposition algorithm is
not sufficient to solve the problem efficiently. The integration of cut generation
with various decomposition methods is widely studied recently (2006). The logic-
based Benders decomposition method was introduced by Hooker (2003). The
advantage of the logic-based Benders is that it permits to combine MILP and the
constraint programming approach. Similar idea was applied to solve the
simultaneous planning and scheduling problems (2006).
A number of authors have addressed the conflict free routing problem with a
static transportation requests set, i.e., with all requests known a priori. Lee et al.
(1998) present a two-staged traffic control scheme to solve a conflict free routing
problem. Their heuristic method consists of generating off-line k-shortest paths in
the first stage before the on-line traffic controller picks a conflict free shortest path
whenever a dispatch command for an AGV is issued (second stage). Rajotia et al.
(1998) propose a semi-dynamic time window constrained routing strategy. They
use the notions of reserved and free time windows to manage the motion of
vehicles. Krishnamurthy et al. (1993) propose an optimization approach. Their
objective is to minimize the makespan. They assume that the assignment of tasks
to AGVs is given and they solve the routing problem by column generation. Their
method generates very good solutions in spite of the fact that it is not optimal
(column generation is performed at the root node of the search tree only).
Oboth et al. (1999) present a heuristic method to solve the dispatching and
routing problems but not simultaneously. Scheduling is performed first and a
sequential path generation heuristic (SPG) is used to generate conflict free routes.
The SPG is inspired from Krishnamurthy et al. (1993) static version of the AGV
routing problem and applied to a dynamic environment while relaxing some of the
limiting assumptions like equal and constant speeds of AGVs. When conflict is
encountered, no feedback is sent to the scheduling module. The AGV being routed
has to be delayed if an alternate route cannot be generated. The authors use rules
for positioning idle AGVs instead of letting the system manage them.
Langevin et al. (1996) propose a dynamic programming based method to solve
exactly instances with two vehicles. They solve the combined problem of
dispatching and conflict free routing. Desaulniers et al. (2003) propose an exact
method that enables to solve instances with up to four vehicles. Their approach
combines a greedy search heuristic (to find a feasible solution and set bound on
delays), column generation and a branch and cut procedure. Their method presents
however some limits since its efficiency depends highly on the performance of the
starting heuristic. If no feasible solution is found by the search heuristic, then no
optimal solution can be found. The search heuristic performs poorly when the
level of congestion increases.
Consider a jobshop manufacturing system with multiple AGVs performing

material handling. There is a number of AGVs pre-specified for material handling.
The AGVs guide paths may be occupied in the time that an AGV is sent to do the
material handling. Therefore, finding the free path to fulfill the function is
important. The manufacturing process plan of all jobs processing time is cleared.
If an AGV arrives early, it should waits until the part processing is finished. The
waiting time is related to the distance the AGV moves and the due date of jobs in
shops. The overall problem is to determine the manufacturing schedule and
routing for AGVs to minimize the total penalized earliness/tardiness and AGVs’
waiting times at the shops in jobshop configuration. The following assumptions
are considered for modeling the proposed problem.
• The number of jobs, processing time for each job and the number of AGVs
are given.
• The started job cannot be interrupted once the processing of that is started at
a shop.
• Each AGV can transport a single load at a time. A transportation task is the
set of a starting node and a destination node.
The overall decision variables for the problem consist of the allocation of
AGVs to the shops for material handling and the conflict-free routing for vehicles.
To estimate the time of traveling from one shop (node) to another in our
proposed network, a mathematical relation is employed. While the velocity (v) of
AGV is known, the time is computed as follows:
d = v.t , (1)
where d is the movement distance, v is AGV velocity and t is the time of

movement, due to a Cartesian coordinates structure as shown in Figure 1.
en
t
em
ov
M
Fig. 1 Cartesian coordinates for AGV movement
to compute the distance, we use rectilinear formulae,
d = x − x0 + y − y0 , (2)
where x is the horizontal movement and y is the vertical movement. As a result,

we have,
d
t= . (3)
v
Therefore, AGV movement time is computed.

Here, we formulate the proposed problem, mathematically. The indices, parameters
and decision variable are introduced below.
Indices:
i Index for jobs; i = 1, 2,..., m
j Index for AGVs; j = 1,2,..., n
k ,k ' Index for shops; k = 1,2,..., o, k ' = 2,3,..., o
Parameters:
α weight for total earliness; α ≥ 0
β weight for total tardiness; β ≥ 0
p ik processing time of job i being processed in shop k
duik due date of job i being processed in shop k
dkk' distance between any two shops k and k'
V velocity of AGV (same for all AGVs)
tjkk' Movement time for AGV j between nodes (shops) k and k'
Eik earliness of job i being processed in shop k
TAik tardiness of job i being processed in shop k
Cik completion time for job i being processed in shop k
Qik allocation of job i to shop k; 1, if job i is allocated to shop k, 0,
otherwise
Decision variables:
⎧1 if AGV j is scheduled on job i in shop k
X jik = ⎨
⎩0 otherwise,
⎧1 if a path between shops k and k ' is busy by AGV j
Y jkk ' = ⎨
⎩0 otherwise,
for i = 1 , ..., m , j = 1 ,..., n , k = 1,... , o.
Objective function:
o m 0 m
Min z = ∑∑ α .Eik + ∑∑ β .TAik (4)
k =1 i =1 k =1 i =1
where,
o m
α = ∑∑ pik − du ik ,
2
k =1 i =1
o m
i = 1,....,m, k = 1,.....,o . (5)
β = ∑∑ pik − C ik ,
2
k =1 i =1
The weights for total earliness and tardiness are loss functions of processing
times, due dates and completion times. Note that, α and β are computed based
on quadratic loss function with a positive squared subtraction of processing times
from due dates and completion times for earliness and tardiness, respectively.
Constraints:
n
∑X
j =1
jik = 1, i = 1,...., m, k = 1,....., o, (6)
m o
∑∑ X
i =1 k =1
jik = 1, j = 1,...., n, (7)
X jik .Y jkk ' = 1, i = 1,...., m, j = 1,..., n, k = 1,.....o, (8)
∑Y
j =1
jkk ' = 1, k = 1,..., o, k ' = k + 1, (9)
∑Y
j =1
jkk ' .Qik = 1, i = 1,..., m, k = 1,..., o, k ' = k + 1, (10)
∑X
j =1
jik .Qik = 1, i = 1,..., m, k = 1,..., o, (11)
n
pik = ∑ X jik .t jkk ' , i = 1,...., m, k = 1,....., o, (12)
j =1
Cik ≥ Ci ,k −1 + pik , i = 1,...., m, k = 1,....., o, (13)
TAik = max {t jkk ' , Cik − duik }, i = 1,...., m, k = 1,....., o, (14)
Eik = max{t jkk ' , duik − Cik }, i = 1,....,m, k = 1,.....,o , (15)
X jik , Y jkk ' ∈ {0,1}, i = 1,...., m, k = 1,....., o, j = 1,...., n. (16)

Constraints (6) specify that exactly one AGV is scheduled at the kth shop.
Constraints (7) define each AGV to be scheduled only once. Constraints (8)
emphasize that when an AGV is scheduled for a job in a shop then the path from
that shop to the next immediate shop is busy. Constraints (9) imply that only one
AGV is allocated to a shop. Constraints (10) show that for a job-shop allocation
only one AGV occupies the path between any two nodes, too. Constraints (11)
indicate that for a job-shop allocation only one AGV is scheduled. Relations (12)
compute the processing times of jobs in shops. Relations (13) show an inequality
between the completion times of job in a previous shop and the processing time of
the job in next shop. Relations (14) present the tardiness. Relations (15) present
the earliness. Relations (16) indicate that all variables are binary.
Since formulae (14) and (15) are nonlinear, we linearize them as follows:
Eik ≥ t jkk ' , i = 1,....,m, k = 1,.....,o , (17)
Eik ≥ duik − Cik , i = 1,....,m, k = 1,.....,o , (18)
TAik ≥ t jkk ' , i = 1,....,m, k = 1,.....,o , (19)
TAik ≥ Cik − duik , i = 1,....,m, k = 1,.....,o . (20)
Clearly, the equation (8) is nonlinear due to product of two binary decision
variables leading to inefficiency of the outputs and spending longer time for
convergence purposes. So, we linearize it by substituting the binary variable ε ijkl
as follows,
ε ijkk ' = X jik .Y jkk ' , ∀i, j , k , k ' . (21)
Meanwhile, the following linear inequalities should be added to the

mathematical model,
ε ijkk ' − X jik − Y jkk ' + 1.5 ≥ 0, ∀i, j, k , k ' , (22)
1.5ε ijkk ' − X jik − Y jkk ' ≤ 0, ∀i, j, k , k ' . (23)
Therefore, the final linearized mathematical model is given below:

Eq. (4)
Eq. (5)
s.t.
Eq. (6) and (7)
Eq. (21)-(23)
Eq. (9)-(13)
Eq. (17)-(20)
Eq. (16).
Next, we test the applicability of the proposed mathematical model in the

numerical results.
Here, we work out on the computational aspects of the proposed mathematical
model. The number of variables and constraints in the linearized model are
presented parametrically in tables 1 and 2, respectively, based on the variable
indices.
Table 1 The number of variables in the linearized model
Variable Count
Xjik n×m×o
Yjkk' n×o×(o-1)
Sum= (n×m×o) + (n×m×(o-1))
Table 2 The number of constraints in the linearized model
Con. Count Con. Count

(6) m×o (11) i×o
(7) n (12) i×o
(21-23) 3(n×m×o×(o-1)) (13) i×o
(9) o×(o-1) (17-20) 4(i×o)
(10) m×o×(o-1) (16) m×o×n
Sum= (m×o) +( n)+3(n×m×o×(o-1))+(o×(o-1))+(m×o×(o-1))+7(i×o)+( m×o×n)
According to tables 1 and 2, when the number of jobs, shops and AGVs
increase, the search space increases drastically. The solution should provide the
optimal paths for each AGV and prohibit the conflict among AGVs minimizing
the total earliness and tardiness.
AGVs begin their movement from a corresponding shop according to the
process plan and go on through other shops to complete the pre-defined job
sequence and deliver the finished product to the depot. These movements of
AGVs configure a network as shown in Figure 2. The AGVs movements are based
on the due dates of the jobs and distances between any two shops. Thus, to prevent
the AGVs conflicts the solution algorithm should consider an allocation between
jobs and AGVs so that no conflict occurs.
At each iteration, the arcs (paths between any two shops) in the network are
divided into two sets; the arcs occupied by an AGV (A); the free arcs (B). A
network structure (A, B) is optimal if an AGV serves as a handling device for a job
(k, k ' ) ∈ A is one and at the same time the AGV waiting time to move to next
job is minimized. With those conditions, the current solution is optimal.
Otherwise, there are arcs in the network that violate the optimal conditions. An arc
is a violated arc if it belongs to B and has lower waiting time than an ark in A. This
Algorithm maintains a feasible arc at each iteration and successfully goes toward
the optimality conditions until it becomes optimal.
Fig. 2 A network configuration of the problem
Heuristic Search Algorithm

To create the initial or Basic Feasible Solution (BFS) in Step 0, an artificial node 0
and artificial arcs are appended to the network. The node ‘0’ will be the source of
the network and the artificial arcs connect the nodes to the source. The set B
consists of the main arcs in the network, and the set A is empty.
Selection of an earliness/tardiness scheme is an important decision in Step 1.
During this step the waiting time of the AGVs on non-basic arcs are recalculated.
If there is at least one that violates its optimality condition it is a candidate to enter
the basis.
In Step 1, appending the entering arc (k, k'), an arc with violation, to the
network forms a unique cycle, W, with the arcs of the basis.
In order to eliminate this cycle (Step 2), one of its arcs must leave the basis. The
cycle is eliminated when we have augmented flow by a minimum amount of
earliness or tardiness to force the flow in one or more arcs of the cycle to the next
shop. By augmenting the flow in an empty arc, the objective value of the solution is
improved. The first task in determining the leaving arc is the identification of all arcs
of the cycle. The flow (AGV path) change is determined by the equation
θ = min{Y jkk ' for all (k , k ' ) ∈W }. The leaving arc is selected based on cycle W.
The substitution of entering for the leaving arc and the reconstruction of new
network is called a pivot (Step 3). After pivoting to change the basis, the
earliness/tardiness for each arc (k , k ' ) ∉ A is calculated. If the
earliness/tardiness for all (k , k ' ) ∈ {A, B} satisfies the optimality condition then
the current basic feasible solution is optimal. Otherwise, an arc (k, k') where there
is a violation should be chosen and operations of the algorithm should be repeated.
The steps of the proposed search algorithm are specified in Figure 3.
Fig. 3 The proposed search algorithm
Architecture of Solution Approach

Here, the architecture and operations of the proposed solution approach is
described, briefly. At the start of process, the Job Generator generates a few jobs
for the AGVs. These jobs will be appended to the remaining jobs, which is empty
at the beginning. The remaining jobs are used by the heuristic search algorithm
and the output of this method is an individual job for every AGV. When the time
is running, the travelled and waited times of every AGV should be updated. At the
same time, if the AGV picks up the job, the assigned job for the AGV will be
deleted and removed from the list of remaining jobs. If the job should be delivered
to the shop, it could not be removed until meeting time between the preceding and
proceeding shops. The Job Generator has to generate a few new jobs, when it finds
out any idle AGV. The main architecture of the proposed solution approach is
shown in Figure 4.
Fig. 4 Block diagram for the proposed solution approach
9.4 Conclusions
In this chapter, the automated guided vehicle scheduling problem was formulated
as a special case of the earliness/tardiness minimization. In the proposed problem
the number of jobs, processing time for each job and the number of AGVs were
assumed to be given. Also, the started job couldn't be interrupted once the
processing of that was started at a shop, and each AGV could transport a single
load at a time. Besides scheduling, the model was able to fulfill the conflict free
routs for the AGVs. Due to nonlinear equations, a linearization process performed
to make the mathematical model smooth for the solution approach. The number of
parameters and variables and therefore the search space increased when the
number of jobs, shops and AGVs increase. The solution should provide the
optimal paths for each AGV and prohibit the conflict among AGVs minimizing
the total earliness and tardiness. Then, a heuristic search algorithm and a solution
methodology based on network concepts were developed for tackling the problem.
Our computational results implied that the algorithms were efficient to find the
optimal solution for large scale problems in negligible time. The heuristic search
algorithm accompanied with the proposed solution architecture was a complete
algorithm and it was efficient for dealing with the problems without any certain
limits in size.
References
Baptiste, P., Le Pape, C., Nuijten, W.: Incorporating efficient operations research
algorithms in constraint-based scheduling. In: First International Workshop on
Artificial Intelligence and Operations Research, Timberline Lodge, Oregon (1995)
Baptiste, P., Le Pape, C., Nuijten, W.: Constraint-Based Scheduling, Applying Constraint
Programming to Scheduling Problems. Kluwer’s International Series (2001)
Bilge, U., Ulusoy, G.: A time window approach to simultaneous scheduling of machines
and material handling system in an FMS. Operations Research 43, 1058–1070 (1995)
References 145
Blazewicz, J., Burkard, R.E., Finke, G., Woeginger, G.J.: Vehicle scheduling in two-cycle
flexible manufacturing systems. Mathematical and Computer Modelling 20, 19–31
(1994)
Co, C.G., Tanchoco, J.M.A.: A review of research on AGVs vehicle management.
Engineering Costs and Production Economics 21, 35–42 (1991)
Corréa, A.I., Lagevin, A., Rousseau, L.-M.: Scheduling and routing of automated guided
vehicles: A hybrid approach. Computers & Operations Research 34, 1688–1707 (2007)
Desaulniers, G., Langevin, A., Riopel, D., Villeneuve, B.: Dispatching and conflict-free
routing of automated guided vehicles: an exact approach. The International Journal of
Dotoli, M., Fanti, M.P.: Coloured timed Petri net model for real-time control of automated
guided vehicle systems. International Journal of Production Research 9, 1787–1814
(2004)
Erdirik-Dogan, M.E., Grossmann, I.E.: A decomposition method for the simultaneous
planning and scheduling of single stage continuous multi product plants. Industrial
Engineering & Chemistry Research 45, 299–315 (2006)
Ganesharajah, T., Hall, N.G., Sriskandarajah, C.: Design and operational issues in AGV-
served manufacturing systems. Annals of Operations Research 76, 109–154 (1998)
Ghasemzadeh, S., Behrangi, E., Azgomi, M.A.: Conflict-free scheduling and routing of
automated guided vehicles in mesh topologies. Robotics and Autonomous Systems 30,
738–748 (2009)
Hooker, J.N.: Logic-based benders decomposition. Mathematical Programming 96, 33–60
(2003)
Jerald, J., Asokan, P., Saravanan, R., Rani, A.D.C.: Simultaneous scheduling of parts and
automated guided vehicles in an FMS environment using adaptive genetic algorithm.
The International Journal of Advanced Manufacturing Technology 29, 584–589 (2006)
Khayat, G.E., Lagevin, A., Riopel, D.: Integrated production and material handling
scheduling using mathematical programming and constraint programming. European
Kim, C.W., Tanchoco, J.M.A.: Conflict-free shortest time bi-directional AGV routing.
King, R.E., Wilson, C.A.: A review of automated guided vehicle system design and
scheduling. Production Planning and Control 2, 44–51 (1991)
Krishnamurthy, N.N., Batta, R., Karwan, M.H.: Developing conflict-free routes for
automated guided vehicles in a flexible manufacturing system. Operations Research 41,
1077–1090 (1993)
Lacomme, P., Moukrim, A., Tchernev, N.: Simultaneous job input sequencing and vehicle
dispatching in a single-vehicle automated guided vehicle system: A heuristic branch-
and-bound approach coupled with a discrete events simulation model. International
Journal of Production Research 43, 1911–1943 (2005)
Langevin, A., Lauzon, D., Riopel, D.: Dispatching, routing and scheduling of two
automated guided vehicles in a flexible manufacturing system. International Journal of
Lee, J.H., Lee, B.H., Choi, M.H.: A real-time traffic control scheme of multiple AGV
systems for collision free minimum time motion: a routing table approach. IEEE
Transactions on Systems, Man and Cybernetics—Part A: Systems and Humans 28,
347–358 (1998)
Mantel, R.J., Landeweerd, H.R.A.: Design and operational control of an AGV system.
International Journal of Production Economics 41, 257–266 (1995)
Nishi, T., Ando, M., Konishi, M.: Distributed route planning for multiple mobile robots
using an augmented Lagrangian decomposition technique. IEEE Transactions on
Robotics 22, 1191–1200 (2005)
Oboth, C., Batta, R., Karwan, M.: Dynamic conflict-free routing of automated guided
vehicles. International Journal of Production Research 37, 2003–2030 (1999)
Qiu, L., Hsu, W.-J., Wang, H.: Scheduling and routing algorithms for AGVs: a survey.
Rajotia, S., Shanker, K., Batra, J.L.: A Semi-dynamic window constrained routing strategy
in an AGV system. International Journal of Production Research 36, 35–50 (1998)
Ralphs, T.K., Galati, M.V.: Decomposition and dynamic cut generation in integer linear
programming. Mathematical Programming 106, 261–285 (2006)
Singh, S.P., Tiwari, M.K.: Intelligent agent framework to determine the optimal conflict-
free path for an automated guided vehicle system. International Journal of Production
Research 40, 4195–4223 (2002)
Tunali, S.: Evaluation of alternate routing policies in scheduling a jobshop type FMS.
Computers & Industrial Engineering 32, 243–250 (1998)
Ulusoy, G., Sivrikaya-Srifoglu, F., Bilge, U.: A genetic algorithm approach to the
simultaneous scheduling of machines and automated guided vehicles. Computers &
Vis, I.F.A.: Survey of research in the design and control of automated guided vehicle
systems. European Journal of Operational Research 170, 677–709 (2006)
Wu, N., Zhou, M.: Modeling and deadlock avoidance of automated manufacturing systems
with multiple automated guided vehicles. IEEE Transactions on Systems Man and
Cybernetics B 35, 1193–1201 (2005)
Chapter 10
Markovian Model for Multiple AGV System
10.1 Summary
Traditional manufacturing has relied on dedicated mass-production systems to

achieve high production volumes at low costs. As living standards improve and
the demands for new consumer goods rise, manufacturing flexibility gains
prominence as a strategic tool for rapidly changing markets. Flexibility, however,
cannot be properly incorporated in the decision-making process if it is not well
defined and measured in a quantitative manner. Flexibility in its most rudimentary
sense is the ability of a manufacturing system to respond to changes and
uncertainties associated with the production process (Miettinenet al., 2010; Kumar
and Sridharan, 2009; Das et al., 2009). A comprehensive classification of eight
flexibility types was proposed in Browne et al. (1984).
Flexible manufacturing systems (FMS) are crucial for modern manufacturing to
enhance productivity involved with high product proliferation (Paraschidiset al.,
1994). As one of the critical components of the FMS, the flexible material
handling system (MHS) plays a strategic role in the implementation of the FMS
(Beamon, 1998). According to Tompkins et al. (2002), about 20–50% of the total
production cost is spent on material handling. This makes the subject of material
handling increasingly important. In addition, all the complexity of manufacturing
is passed on to the MHS. Therefore, the flexible MHS has been vital for
improving the FMS to fulfill the requirements of high product proliferation.
Automated manufacturing systems (AMS), which are equipped with several
CNC machines and AGV-based material handling system are designed and
implemented to gain the automation and efficiency of production. To make use of
all features of AMS, the planning in the AMS decision making process is critical
because the planning decision has influence on the subsequent decision processes
such as scheduling, dispatching, etc. The planning in automated manufacturing
systems can be characterized as being online and short-term nature to respond to
frequently changing production order. Given a production order, manufacturing
planning function is responsible to establish a plan by decomposing the production
task into a set of subtasks. An analysis of AMS dealing with changing demand can
be found in (Terkaj et al., 2009). An extensive review of the loading problem for

148 10 Markovian Model for Multiple AGV System
an FMS can be found in (Grieco et al., 2001). An early stochastic programming

approach to address the short-term production planning for an FMS can be found
in (Terkaj and Tolio, 2006).
Automated Guided Vehicle System (AGVS) becomes popular in many
industrial fields because of its flexibility, reliability, safety, and contribution to the
increase of productivity and to the improvement of housekeeping. But, the
performance of the material handling system is significantly influenced by several
operating policies. One of the important operating policies is the positioning
strategy of idle vehicles on the guide path (Egbehi, 1993; Kim, 1995).
In most manufacturing systems, decision making is worked out at several
stages of design, planning and operation. The role of performance modeling is
significant in advanced manufacturing systems from economic viewpoints.
However, events such as machine breakdown, changes in part type and volume,
tool replacement, raw material and other short interruptions are effective on the
desired performance of a manufacturing system. This problem is critical due to its
impacts on the capacity of the system (Stoop and Wiers, 1996). Researches on the
automated manufacturing systems imply that the machine failure is the major
problem in analyzing system performance in comparison with other factors like
raw material, equipment, software and workers (Sanchez, 1994). Therefore,
reliability considerations should be taken into account for manufacturing system
analysis. Researchers who studied this problem include (Hilderbrant, 1980;
Kimemia, 1982; Liberopulos, 1993; Viswanadham and Narahari, 1992; Vinod,
1983; Vinod and Solberg, 1984; Choiand Lee, 1998).
Since the manufacturing systems experience different failure states, therefore
considering these states in modeling a reliability problem is of importance. The
best way for considering system states in modeling is to employ Markovian
property. Reibman (1990) stated the problem in estimating the probability of
failure in different state is vital for reliability computations.
The increasing demand for the reliability assessment in manufacturing systems
under several random parameters has been investigated by several approaches
facilitating the computations of probability estimations. According to the following
brief literature review, studies to compare Markovian and neural networks are
few. Especially, modeling the reliability of an advanced manufacturing system
considering AGVs is also rare.
The improvement of safety in the process industries is related to assessment
and reduction of risk in a cost-effective manner. Kančev and Čepin(2011)
addressed the trade-off between risk and cost related to standby safety systems.
An age-dependent unavailability model that integrated the effects of the test and
maintenance (T&M) activities as well as component ageing was developed and
represented the basis for calculating risk. The repair “same-as-new” process was
considered regarding the T&M activities. Costs were expressed as a function of
the selected risk measure. The time-averaged function of the selected risk measure
was obtained from probabilistic safety assessment, i.e. the fault tree analysis. This
function was further extended with inclusion of additional parameters related to
T&M activities as well as ageing parameters related to component ageing. In that
10.1 Summary 149
sense, a new model of system unavailability, incorporating component ageing and

T&M costs, was presented. The testing strategy was also addressed. Sequential
and staggered testings were compared. The developed approach was applied on a
standard safety system in nuclear power plant although the method was applicable
to standby safety systems that were tested and maintained in other industries as
well.
The problem of selecting a suitable maintenance policy for repairable systems
and for a finite time period was presented by Marquez and Heguedas (2002).
Since the late seventies, examples of models assessing corrective and preventive
maintenance policies over an equipment life cycle exist in the literature. However,
there are not too many contributions regarding real implementation of these
models in the industry, considering realistic timeframes and for repairable
systems. Modeling this problem requires normally the representation of different
corrective and/or preventive actions that could take place at different moments,
driving the equipment to different states with different hazard rates. An approach
to pattern the system under finite periods of time has been the utilization of semi-
Markovian probabilistic models, allowing later a maintenance policy optimization
using dynamic programming. These models are very flexible to represent a given
system, but they are also complex and therefore very difficult to handle when the
number of the system possible states increases. Marquez and Heguedas (2002)
explored the trade-off between flexibility and complexity of these models, and
presented a comparison in terms of model data requirements versus potential
benefits obtained with the model.
In generalized renewal process (GRP) reliability analysis for repairable
systems, Monte Carlo (MC) simulation method instead of numerical method is
often used to estimate model parameters because of the complexity and the
difficulty of developing a mathematically tractable probabilistic model. Wang and
Yang (2012) proposed a nonlinear programming approach to estimate restoration
factor for the Kijima type GRP model I, as well as the model II based on the
conditional Weibull distribution for repairable systems, using negative log-
likelihood as an objective function and adding inequality constraints to model
parameters. The method minimized the negative log-likelihood directly, and
avoided solving the complex system of equations. Three real and different types of
field failure data sets with time truncation for NC machine tools were analyzed by
the proposed numerical method. The sampling formulas of failure times for the
GRP models I and II were derived and the effectiveness of the proposed method
was validated with MC simulation method.
Ke et al. (2013) considered a multi-repairmen problem comprising of M
operating machines with W warm standbys (spares). Both operating and warm
standby machines were subject to failures. With a coverage probability c, a failed
unit was immediately detected and attended by one of R repairmen if available. If
the failed unit was not detected with probability 1−c, the system entered an unsafe
state and must be cleared by a reboot action. The repairmen were also subject to
failures which result in service (repair) interruptions. The failed repairman
resumed service after a random period of time. In addition, the repair rate
depended on number of failed machines. The entire system was modeled as a

finite-state Markov chain and its steady state distribution was obtained by a
recursive matrix approach. The major performance measures were evaluated based
on this distribution. Under a cost structure, the authors proposed to use the Quasi-
Newton method and probabilistic global search Lausanne method to search for the
global optimal system parameters.
Nowadays VoIP has become an evolutionary technology in telecommunications.
Hence it is very important to study and enhance its dependability attributes. An
analytical dependability model for VoIP was proposed by Gupta and Dharmaraja
(2011). The study was focused on analyzing the combined effects of resource
degradation and security breaches on the Quality of Service (QoS) of VoIP, to
enhance its overall dependability. As a preventive maintenance policy to prevent or
postpone software failures which cause resource degradation, software rejuvenation
was adopted. The dependability model was analyzed using semi-Markov process
which captures the effects of non-Markovian nature of the time spent at various
states of the system. The steady-state as well as the time-dependent analysis of the
dependability model was presented.
Zhou et al. (2012) presented a maintenance optimization method for a multi-
state series-parallel system considering economic dependence and state-dependent
inspection intervals. The objective function considered in the chapter was the
average revenue per unit time calculated based on the semi-regenerative theory
and the universal generating function (UGF). A new algorithm using the stochastic
ordering was also developed in the chapter to reduce the search space of
maintenance strategies and to enhance the efficiency of optimization algorithms.
A numerical simulation was presented in the study to evaluate the efficiency of the
proposed maintenance strategy and optimization algorithms.
Reliability assessment for hard disk drives (HDDs) is important yet difficult for
manufacturers. Motivated by the fact that the particle accumulation in the HDDs,
which accounts for most HDD catastrophic failures, is contributed from the
internal and external sources, a counting process with two arrival sources was
proposed by Ye et al. (2013) to model the particle cumulative process in HDDs.
The model successfully explained the collapse of traditional ALT approaches for
accelerated life test data. Parameter estimation and hypothesis tests for the model
were developed and illustrated with real data from a HDD test. A simulation study
was conducted to examine the accuracy of large sample normal approximations
that were used to test existence of the internal and external sources.
An R out of N repairable system consisting of N independent components is
operating if at least R components are functioning. The system fails whenever the
number of good components decreases from R to R − 1. A failed component is
sent to a repair facility having several repairmen. Life times of working
components are i.i.d random variables having an exponential distribution. Repair
times are i.i.d random variables having a phase type distribution. Both cold and
warm stand-by systems are considered. Barron et al. (2006) presented an
algorithm deriving recursively in the number of repairmen the generator of the
Markov process that governs the process. Then they derived formulas for the point
availability, the limiting availability, the distribution of the down time and the up
time. Numerical examples were given for various repair time distributions. The
numerical examples showed that the availability is not very sensitive to the repair
time distribution while the mean up time and the mean down time might be very
sensitive to the repair time distributions.
According to the brief reviewed literature, studies to compare Markovian and
neural networks are few. Especially, modeling the reliability of an advanced
manufacturing system considering AGVs is also rare.
Here, a jobshop manufacturing system having multiple AGVs for material

handling purpose is considered. In each shop several machines perform the part
processing according to a process plan. To transfer the parts among different
shops AGVs are employed. The reliability of the whole manufacturing system is
concerned with the reliability of the machines in shops and the reliability of the
AGVs. The failure of the machines and AGVs could be considered in different
states. The failure causes for machines are:
• Amateur operator
• Equipment deficiency
• Inappropriate part specifications
Also, the failures of AGVs are due to:

• Carrier overload
• Guide path fracture
Using Markovian property, we can configure the transition diagram and the
corresponding matrix.
The result of the Markovian process is the failure probability for machines and
AGVs. These probabilities are applied in reliability computations. For reliability,
first we conceptualize different scenarios exist in the proposed manufacturing
system. The shops are in parallel since the parts are disseminated through the
system according to the process plan. The sequence of machines in a shop may be
important or not, i.e., the part processing in a shop should be performed
sequentially on the machines or the sequence is not important and parallel
machining is possible. Therefore, two separate cases of series and parallel should
be modeled. AGVs are in series since if one AGV break down then the whole
system should wait until the AGV is repaired or taken out of the system.
The aim of the decision maker is to maximize the performance of the whole
system. To achieve the aim, two objectives namely maximizing the total reliability
of machines in shops in the whole jobshop system and the maximizing the total
reliability of the AGVs, should be investigated. Also, for the economic viewpoint
of the system performance the third objective is to minimize the total repair cost in
the system. As a unit (machine or AGV) in the system is broken down, the repair
should be performed on it for preparing it to function.
The aims of conducting this study are:
 Developing a reliability assessment methodology for AGV based
manufacturing systems
 Analyzing and including fault sources in machine-AGV state modeling in
manufacturing systems
 Markovian modeling for reliability assessment of a machine-AGV
manufacturing system
 Comparing Markovian reliability assessment with the neural network
method
It is necessary to incorporate reliability into the model to ensure the level of
service for each machine in each shop and the AGVs. For modeling reliability, the
approach of Ball and Lin (1993) is adopted and further extended.
The reliability is defined as the probability that the system works until time t. If
a machine in a shop is broken down, it can be regarded as a failure. A desired
level of reliability can be achieved by limiting the failure probabilities. This
approach for handling reliability is called the method of chance constraints in the
context of mathematical programming. The use of chance constraints in vehicle
routing problem was illustrated in Stewart and Golden (1983). Carbone (1974)
used chance constraints for selecting multiple facilities under normally distributed
demand. The model minimized an upper bound on the total demand-weighted
distance while ensuring that constraint was satisfied with specified chance or
probability. Shiode and Drezner (2003) used a similar approach in a competitive
It is assumed that the reliability of each machine type and the AGV are
independently according to Exponential processes. Also, J is total types of
machines, i.e., drilling machines, turning machines, bending machines show three
machine types. We discuss the reliability based model as follows:
R j (t ) : The probability that machine type jth works until time t.
⎧⎛ J ⎞
⎪ ⎜⎜1 − ∏ (1 − R j (t )) ⎟⎟, when machines in each shop are in parallel case
⎪⎪ ⎝ j =1 ⎠
R(t ) system =⎨
⎪⎛ J ⎞
⎪⎜⎜ ∏ R j (t ) ⎟⎟, when machines in each shop are in series case
⎪⎩⎝ j =1 ⎠
(1)
In our proposed problem, AGVs are series and the machine types in each shop
may be in parallel or series cases and the shops are parallel, i.e., a composite
system is configured. Therefore, the reliability of the system is as follows:
⎛ J ⎞
⎜1 − ∏ (1 − R j (t )) ⎟ ≥ α , (2)
⎜ ⎟
⎝ j =1 ⎠
owhere α is the lower bound for a desirable reliability of the system until time t.
As previously assumed, the reliability of each machine type and AGV are
independently according to Exponential distribution:
−t
θj
R j (t ) = e , (3)
Where θj is the exponential parameter for machine type or AGV breakdown.

Then,
⎛ J
−t
⎞
⎜1 − θj ⎟
⎜ ∏
(1 − e ) ≥α . (4)
⎟
⎝ j =1
⎠
It is obvious that to obtain a higher level of reliability, more cost is incurred to
the system. Hence, a cost function ( C j (t ) ) is defined to keep machine type jth
reliable until time t. For the whole system we have:
J
∑C (t ) .
j =1
j (5)

In this section, we construct the proposed failure state diagrams and matrices for
machines and AGVs’ using Markov system, separately. A Markov system is a
system that can be in one of several (numbered) states, and can pass from one state
to another each time step according to fixed probabilities. If a Markov system is in
state i, there is a fixed probability, pij, of it going into state j the next time step, and
pij is called a transition probability. A Markov system can be illustrated by means
of a state transition diagram, which is a diagram showing all the states and
transition probabilities. The entries in each row add up to 1.
First, we configure the machines’ state diagram. As stated before, the machines
may be broken down in three states, namely, (A) amateur operator, (B) equipment
deficiency, and (C) inappropriate part specifications. Note that, the states refer to
the break down states causes, i.e., the machine is working or it is broken down due
to the failure states such as amateur operator, equipment deficiency, and
inappropriate part specifications. In another word, since we are modeling the
reliability of the system considering different state changes, it is common in
Markovian computations to monitor the state transition while all are the causes of
breakdown. The state transition diagram for machines is shown in Figure 1.
−ε
1−α
α 1− β − γ
β
ε
γ
ν
1 − δ −ν
Fig. 1 The state transition diagram for machines
As a result the corresponding transition matrix Pij is,
⎡1 − α − ε α ε ⎤
⎢
Pij = ⎢ β 1− β − γ γ ⎥⎥ . (6)
⎢⎣ ν δ 1 − δ − ν ⎥⎦
where α , β , γ , δ , ε , and ν are the transition probability from the three states
given in Figure 1. Using the probability transition matrix and the limiting
probability we obtain each state’s occurrence probability as follows.
3
π j = ∑ π i pij , for j=1,2,3. (7)
i =1
∑π
j =1
j =1. (8)
Using these probabilities, we can compute the reliability of each state helping
us to assess the total reliability of the system.
We also can compute the long run probability for each state using steady state
distribution given below.
⎡1 − α − ε α ε ⎤
⎢
[A B C ] ⎢ β 1− β − γ γ ⎥ = [A B C ]
⎥ , (9)
⎢⎣ ν δ 1 − δ − ν ⎥⎦
having A+B+C=1.
The same computations exist for AGVs different failure state, while we stated 2
states, i.e., we have two state probability and a 2 × 2 transition matrix.
Now, for reliability we have,
R(t)=1-F(t), (10)
where, F(t) is the failure probability computed above as states’ probabilities. Note
that, we can compute the reliability in two cases, first for current state, and second
for steady state. The numerical comparison of the two could be interesting.
Having the current state of the system by Markovian model and by the means
of neural network, we can compute the steady state probabilities. Next, we review
the artificial neural network and the backpropagation neural network for our
proposed work. The reason is to find the difference in accuracy of the two
methods and determine the most effective one. It is obvious that the neural
network can be more efficient due to using past data in training stage.
The aim to compute the steady state probability and reliability is to obtain an
estimation of the system availability for long run planning horizon. Therefore, it is
significant for a decision maker to determine steady state reliability using the
corresponding probability, accurately.
Artificial Neural Network

Neural networks are being widely used in many fields of study. This could be
attributed to the fact that these networks attempt to model the capabilities of
human brains. Since the last decade, neural networks have been used as a
theoretically sound alternative to traditional statistical models. Although neural
networks (NNs) originated in mathematical neurobiology, the rather simplified
practical models currently in use have moved steadily towards the field of
statistics. A number of researchers have illustrated the connection of neural
networks to traditional statistical models. For example, Gallinari et al. (1991)
presented analytical results establishing a link between discriminant analysis and
multilayer perceptrons (MLP) used for classification problems. Cheng and
Titterington (1994) made a detailed analysis and comparison of various neural
network models with traditional statistical models. They showed strong
associations of the feed-forward neural networks with discriminant analysis and
regression models, and unsupervised networks such as self-organizing neural
networks with clustering. Neural networks are being used in areas of prediction
and classification, areas where regression models and related statistical techniques
have traditionally been used. Ripley (1994) discusses the statistical aspects of
neural networks and classifies neural networks as one of a class of flexible
nonlinear regression models. Warner and Misra (1996) presented a comparison
between regression analysis and neural network computation in terms of notation
and implementation. They also discuss when it would be advantageous to use a
neural network model in place of a parametric regression model, as well as some
of the difficulties in implementation. Vach et al. (1996) presented a comparison
between feed-forward neural networks and the logistic regression. The conceptual
similarities and discrepancies between the two methods are also analyzed.
Artificial neural networks have been applied successfully to many
manufacturing and engineering areas. Zhengrong et al. (1996) used quadratic
regression to assess the results of neural network for improving the efficiency of
fermentation process development. The results show that different sizes of neural
nets within a certain range give an equally good prediction by using the ‘‘stopping
training” technique, while quadratic regressions are sensitive to the size of the data
sets. Smith and Mason (1997) mentioned that regression and neural network
modeling methods have become two competing empirical model-building
methods. They compared the predictive capabilities of NNs and regression
methods in manufacturing cost estimation problems.
The Backpropagation Neural Network

The backpropagation algorithm trains a given feed-forward multilayer neural
network for a given set of input patterns with known classifications. When each
entry of the sample set is presented to the network, the network examines its
output response to the sample input pattern. The output response is then compared
to the known and desired output and the error value is calculated. Based on the
error, the connection weights are adjusted. The backpropagation algorithm is
based on Widrow-Hoff delta learning rule in which the weight adjustment is done
through mean square error of the output response to the sample input. The general
steps of backpropagation are given below.
1. Propagate inputs forward in the usual way, i.e., all outputs are computed using
sigmoid thresholding of the inner product of the corresponding weight and
input vectors. All outputs at stage n are connected to all the inputs at stage n+1
2. Propagate the errors backwards by apportioning them to each unit according to
the amount of the error the unit is responsible for.
We now discuss how to develop the stochastic backpropagation algorithm for
the general case. The following notations and definitions are needed:
x j : input vector for unit j (xji = ith input to the jth unit)
w j : weight vector for unit j (wji = weight on xji)
z j = w j .x j : the weighted sum of inputs for unit j
oj : output of unit j ( o j = σ (z j ) )
tj : target for unit j
Downstream(j) : set of units whose immediate inputs include the output of j
Output : Set of output units in the final layer.
Since we update after each training example, we can simplify the notation
somewhat by assuming that the training set consists of exactly one example and so
the error can simply be denoted by E.
∂E
We want to calculate corresponding to each input weight wji of each
∂w ji
output unit j. Note first that since zj is a function of wji regardless of where in the
network unit j is located,
∂E ∂E ∂z j ∂E
= . = .x ji , (11)
∂w ji ∂z j ∂w ji ∂z j
Furthermore, ∂E is the same regardless of which input weight of unit j we are

∂z j
trying to update. So, we denote this quantity by δj.
Consider the case when j is an output unit. We know that
1
E= ∑ (t k − σ ( z k )) 2 .
2 k∈Outputs
(12)
Since the outputs of all units k ≠ j are independent of wji, we can then drop the
summation and consider just the contribution to E by j and we call it δj:
∂E ∂ 1 ∂o j ∂
δj = = (t j − o j ) 2 = −(t j − o j ) = − (t j − o j ) σ (z j )
∂z j ∂z j 2 ∂z j ∂z j (13)
= −(t j − o j )(1 − σ ( z j ))σ ( z j ) = −(t j − o j )(1 − o j )o j .
Thus,
∂E
Δw ji = −η = ηδ j x ji . (14)
∂w ji
Now, consider the case when j is a hidden unit. Like before, we make the
following two important observations:
1. For each unit k downstream from j, zk is a function of zj.

2. The contribution to error by all units l ≠ j , in the same layer as j, is
independent of wji.
∂E
We want to calculate for each input weight wji for each hidden unit j. Note
∂w ji
that wji influences just zj which influences oj which influences zk,
∀k ∈ Downstream( j ), each of which influences E. So, we can write,
∂E ∂E ∂z k ∂o j ∂z j ∂E ∂z k ∂o j
= ∑ . . . = ∑ . .
∂w ji k∈Downstream ( j ) ∂z k ∂o j ∂z j ∂w ji k∈Downstream ( j ) ∂z k ∂o j ∂z j
.x ji (15)
Again, note that all the terms except xji in (15) are the same regardless of which
input weight of unit j we are trying to update. Like before, we denote this common
∂E ∂z ∂o
quantity by δ j . Also, note that δ k , k wkj and j = o j (1 − o j ) .
∂z k ∂o j ∂z j
Substituting them in (13),
∂E ∂z k ∂o j
δj = ∑ . .
k∈Downstream ( j ) ∂z k ∂o j ∂z j
= ∑ δ k .wkj .o j (1 − o j ) ,
k∈Downstream ( j )
(16)
we obtain:
δ k = o j (1 − o j ) ∑δ k .w .
k∈Downstream( j )
kj (17)
Fig. 2 A configuration of the proposed neural network

To adapt the backpropagation algorithm on our proposed model, consider the

failure causes for machines and AGVs as inputs and the current state failure
probability of machines and AGVs as outputs. We train the network collecting
data in different time periods and compute the importance weight for each input
resulting the corresponding output. A configuration of the proposed neural
network is shown in Figure 2.
We are now in a position to state the backpropagation algorithm formally.
Algorithm 1: Formal statement of stochastic backpropagation.

(Training examples, η , ni, nh, no)
Each training example is of the form x, t , where x is the input vector and t is
the target vector, η is the learning rate (e.g., 0.05), ni, nh and no are the number of
input, hidden and output nodes, respectively. Input from unit i to unit j is denoted
by xji and its weight is denoted by wji. Create a feed-forward network with ni
inputs, nh hidden units, and no output units.
Initialize all the weights to small random values (e.g., between -0.05 and 0.05).
While termination condition is not met Do
For each training example x, t ,
1. Input the instance x and compute the output ou of every unit.

2. For each output unit k, calculate
δ k = ok (1 − ok )(t k − ok ) . (18)
3. For each hidden unit h, calculate
δ h = oh (1 − oh ) ∑δ k .w .
k∈Downstream( h )
kh (19)
4. Update each network weight wji as follows:
w ji ← w ji + Δw ji . (20)
where,
Δw ji = ηδ j x ji . (21)
This way, we can compare the performance of backpropagation neural network

and limiting distribution model for computing the steady state probabilities using
the current state probabilities.
Mathematical Optimization
Here, the mathematical optimization model is given. As stated before we aim to
maximize the total reliability of machines in shops in the whole jobshop system
and maximize the total reliability of the AGVs. Since the reliability model is
stochastic, one may think about simulation study. But considering multiple-
objectives and especially including cost factors in the form of a composite
mathematical function is difficult and requires tiring and complicated simulation
efforts. Also, as we will present further, we considered several 0/1 integer
variables which are easier to be modelled mathematically.
Maximizing Total Reliability of Machines

The following mathematical notations are employed to model this maximization
problem:
Mathematical notations:
k Index for machines, k=1,…,K.
l Index for jobs, l=1,…,L.
m Index for shops, m=1,…,M.
R Reliability of machine k.
k
⎧1 if machine k process job l.

τ kl = ⎨
⎩0 otherwise.
⎧1 if machine k in shop m is chosen.
ϕkm = ⎨
⎩0 otherwise.
The mathematical model:
Max ∑∑ ( R .ϕ ) ,
m k
k km
(22)
s.t.
∑ϕk
km .τ kl = 1, ∀l , m ,
(23)
ϕkm ∈{0,1} , ∀k, m. (24)
Maximizing Total Reliability of AGVs

The following mathematical notations are employed to model this maximization
problem:
Mathematical notations:
n Index for AGVs, n=1,…,N.
R Reliability of AGV n.
n
⎧1 if AGV n can service shop m.

ς nm = ⎨
⎩0 otherwise.
⎧1 if AGV n is chosen.
ζn = ⎨
⎩0 otherwise.
The mathematical model:
Max ∑ R .ζ
n
n n
, (25)
s.t.
∑ζn
n .ς nm = 1, ∀ m,
(26)
ζ n ∈{0,1} , ∀n. (27)
As a result a multi-objective mathematical model is configured as follows:
Max ∑∑ ( R .ϕ ) ,
m k
k km
(28)
Max ∑ R .ζ
n
n n
, (29)
s.t.
∑ϕk
km .τ kl = 1, ∀l , m ,
(30)
∑ζn
n .ς nm = 1, ∀ m,
(31)
∑ϕk
km = 1, ∀m, (32)
ζ n ∈{0,1} , ∀n. (33)
ϕkm ∈{0,1} , ∀k, m. (34)

Next, an approach to optimize the proposed multi-objective model is given. We

use objectives weighing method to integrated and optimize the model.
Analytic Hierarchy Process (AHP) for Multi-objective Optimization

To weight the objectives, we take a multi-criteria decision-making approach.
Multi-criteria decision-making (MCDM), dealing primarily with problems of
evaluation or selection, is a rapidly developing area in operations research and
management science. AHP is a technique of considering data or information for a
decision in a systematic manner. It is mainly concerned with a way of solving
based on three principles: constructing the hierarchy, priority setting, and logical
consistency. We apply AHP to weight the objectives.
Construction of the Hierarchy

A complicated decision problem, composed of various attributes of an objective,
is structured and decomposed into sub-problems (sub-objectives, criteria,
alternatives, etc.), within a hierarchy.
Priority Setting
The relative “priority” given to each element in the hierarchy is determined by pair-
wise comparisons of the contributions of elements at a lower level in terms of the
criteria (or elements) with a causal relationship. In AHP, multiple paired comparisons
are based on a standardized comparison scale of nine levels (see table 1).
Table 1 Scale of relative importance
Intensity of importance Definition of importance

1 Equal
2 Weak
3 Moderate
4 Moderate plus
5 Strong
6 Strong plus
7 Very strong or demonstrated
8 Very, very strong
9 Extreme
Let C = {c1 ,...,cn } be the set of criteria. The result of the pair-wise
comparisons on n criteria can be summarized in an n × n evaluation matrix A in
which every element aij is the quotient of weights of the criteria, as shown below:
A = (aij), i, j = 1, . . . , n. (35)
The relative priorities are given by the eigenvector (w) corresponding to the
largest eigenvalue
( λmax ) as:
Aw = λmaxw . (36)
When pair-wise comparisons are completely consistent, the matrix A has rank 1
and λmax = n . In that case, weights can be obtained by normalizing any of the
rows or columns of A.
The procedure described above is repeated for all subsystems in the hierarchy.
In order to synthesize the various priority vectors, these vectors are weighted with
the global priority of the parent criteria and synthesized. This process starts at the
top of the hierarchy. As a result, the overall relative priorities to be given to
the lowest level elements are obtained. These overall, relative priorities indicate
the degree to which the alternatives contribute to the objective. These priorities
represent a synthesis of the local priorities, and reflect an evaluation process that
permits integration of the perspectives of the various stakeholders involved.
Consistency Check
A measure of consistency of the given pair-wise comparison is needed. The
consistency is defined by the relation between the entries of A; that is, we say A is
consistent if aik= aij · ajk, for all i,j,k. The consistency index (CI) is:
(λmax − n)
CI = . (37)
(n − 1)
The final consistency ratio (CR), on the basis of which one can conclude
whether the evaluations are sufficiently consistent, is calculated to be the ratio of
the CI and the random consistency index (RI):
CI
CR = . (38)
RI
The value 0.1 is the accepted upper limit for CR. If the final consistency ratio
exceeds this value, the evaluation procedure needs to be repeated to improve
consistency. The measurement of consistency can be used to evaluate the
consistency of decision-makers as well as the consistency of all the hierarchies.
We are now ready to give an algorithm for computing objective weights using
the AHP. The following notations and definitions are used.
n: number of criteria
i: number of objectives
p: index for objectives, p=1or 2
d: index for criteria, 1≤ d ≤ D
R pd : the weight of pth item with respect to dth criterion
wd : the weight of dth criterion
Algorithm 2: OWAHP (compute objective weights using the AHP)

Step 1: Define the decision problem and the goal.
Step 2: Structure the hierarchy from the top through the intermediate to the lowest
level.
Step 3: Construct the objective-criteria matrix using steps 4 to 8 using the AHP.
(Steps 4 to 6 are performed for all levels in the hierarchy.)
Step 4: Construct pair-wise comparison matrices for each of the lower levels for
each element in the level immediately above by using a relative scale
measurement. The decision-maker has the option of expressing his or her intensity
of preference on a nine-point scale. If two criteria are of equal importance, a value
of 1 is set for the corresponding component in the comparison matrix, while a 9
indicates an absolute importance of one criterion over the other (table 1 shows the
measurement scale).
Step 5: Compute the largest eigenvalue by the relative weights of the criteria and
the sum taken over all weighted eigenvector entries corresponding to those in the
next lower level of the hierarchy.
Analyze pair-wise comparison data using the eigenvalue technique. Using these
pair-wise comparisons, estimate the objectives. The eigenvector of the largest
eigenvalue of matrix A constitutes the estimation of relative importance of the
attributes.
Step 6:Construct the consistency check and perform consequence weights analysis
as follows:
⎡ 1 w1
…
w1 ⎤
⎢ w2 wn⎥
⎢ ⎥
⎢w 2 1 …
w2 ⎥
A = ( aij ) = ⎢ w 1 wn⎥.
⎢ ⎥
⎢ ⎥
⎢w n wn
1 ⎥
⎣⎢ w 1 w2 ⎥⎦
Note that if the matrix A is consistent (that is, aik= aij · ajk, for all
i , j , k = 1, 2, ..., n ), then we have (the weights are already known),
wi
a ij = , i, j = 1,2,..., n.
wj
(39)
λmax = n.
If the pair-wise comparisons do not include any inconsistencies, then
The more consistent the comparisons are, the closer the value of computed λmax
is to n. Set the consistency index (CI), which measures the inconsistencies of pair-
wise comparisons, to be:
CI =
( λmax − n ) ,
( n − 1)
and let the consistency ratio (CR) be:
⎛ CI ⎞
CR = 100 ⎜ ⎟,
⎝ RI ⎠
where n is the number of columns in A and RI is the random index, being the
average of the CI obtained from a large number of randomly generated matrices.
Note that RI depends on the order of the matrix, and a CR value of 10% or less
is considered acceptable.
Step 7: Form the objective-criteria matrix as specified in table 2:
Table 2 The objective-criteria matrix
C1 C2 … Cd
objective 1 R11 R12 … R1d
objective 2 R21 R22 … R2d
Step 8: As a result, configure the pair-wise comparison for criteria-criteria

matrix as in table 3:
Table 3 The criteria-criteria pair-wise comparison matrix
C1 C2 … Cd wd
Criteria 1 1 a12 … a1d w1
Criteria 2 1/a12 1 … a2d w2
Criteria d 1/a1d 1/a2d … 1 wd
The wd are gained by a normalization process. The wd are the weights for
criteria.
Step 9: Compute the overall weights for the objectives, using tables 2 and 3, as
follows:
ψ = Total weight for objective 1 = R11 × w1 + R12 × w2 + ... + R1d × wd ,

(40)
ψ ' = Total weight for objective 2 = R21 × w1 + R22 × w2 + ... + R2d × wd ,
where ψ + ψ ' = 1 . Thus the integrated objective function is formed as:
⎛ ⎞
Max ⎜ψ .∑∑ ( Rk .ϕkm ) +ψ '.∑ ( Rn .ζ n ) ⎟ .
⎝ m k n ⎠
10.4 Conclusions
We proposed a Markovian model for flexible manufacturing systems (FMSs). The

model considered two features of automated flexible manufacturing systems
equipped with automated guided vehicle (AGV) namely, the reliability of
machines and the reliability of AGVs in a multiple AGV jobshop manufacturing
system. We made use of current state transition matrix for the failure of the
machines and AGVs in different states. Therefore, a Markovian model was
proposed for reliability assessment. Also, for steady state probability
computations, the limiting theorem was compared with adapted backpropagation
neural network showing neural network’s effectiveness. Using the reliabilities, we
worked out an optimization mathematical model. The optimization objectives in
the proposed model were maximizing the total reliability of machines in shops in
the whole jobshop system and maximizing the total reliability of the AGVs. The
computational results illustrated the applicability of our proposed model. A
strategic viewpoint of such computations was to enable the management to control
the failures of AGVs and Machines to satisfy the optimization purposes.
References
Barron, Y., Frostig, E., Levikson, B.: Analysis of R out of N systems with several
repairmen, exponential life times and phase type repair times: An algorithmic approach.
Beamon, B.M.: Performance, reliability, and performability of material handling systems.
Carbone, R.: Public facilities location under stochastic demand. INFOR 12(3), 261–270
(1974)
Cheng, B., Titterington, D.M.: Neural networks: A review from a statistical perspective.
Statistical Science 9(1), 2–30 (1994)
References 167
Choi, S.H., Lee, J.S.L.: A heuristic approach to machine loading problem in non-
preemptive flexible manufacturing systems. International Journal of Industrial
Engineering 5(2), 105–116 (1998)
Engineering 56(1), 283–293 (2009)
Egbehi, P.J.: Positioning of automated guided vehicles in a loop layout to improve response
time. European Journal of Operational Research 71, 32–44 (1993)
Gallinari, P., Thiria, S., Badran, F., Fogelman-Soulie, F.: On the relations between
discriminant analysis and multilayer perceptrons. Neural Networks 4(3), 349–360
(1991)
Grieco, A., Semeraro, Q., Tolio, T.: A Review of Different Approaches to the FMS
Loading Problem. International Journal of Flexible Manufacturing Systems 13(4), 361–
384 (2001)
Gupta, V., Dharmaraja, S.: Semi-Markov modeling of dependability of VoIP network in the
presence of resource degradation and security attacks. Reliability Engineering & System
Safety 96(12), 1627–1636 (2011)
Hilderbrant, R.R.: Scheduling flexible machining systems when machines are prone to
failure. Ph.D. Thesis, Massachusetts Institute of Technology, USA (1980)
Kančev, D., Čepin, M.: Evaluation of risk and cost using an age-dependent unavailability
modelling of test and maintenance for standby components. Journal of Loss Prevention
in the Process Industries 24(2), 146–155 (2011)
Ke, J.-C., Hsu, Y.-L., Liu, T.-H., George Zhang, Z.: Computational analysis of machine
repair problem with unreliable multi-repairmen. Computers & Operations
Research 40(3), 848–855 (2013)
Kim, K.H.: Positioning of automated guided vehicles in a loop layout to minimize the mean
vehicle response time. Int. J. of Production Economics 39, 201–214 (1995)
Kimemia, J.G.: Hierarchical control of production in flexible manufacturing systems. Ph.D.
Thesis, Massachusetts Institute of Technology, USA (1982)
Manufacturing 25(4-5), 829–838 (2009)
Liberopulos, G.: Flow control of failure prone manufacturing systems: Controller design
theory and applications. Ph.D. Thesis, Boston University, USA (1993)
Marquez, A.C., Heguedas, A.S.: Models for maintenance optimization: a study for
repairable systems and finite time periods. Reliability Engineering & System
Safety 75(3), 367–377 (2002)
Paraschidis, K., Fahantidis, N., Petridis, V., Doulgeri, Z., Petrou, L., Hasapis, G.: A robotic
system for handling textile and nonrigid flat materials. In: Proceedings of the Computer
Integrated Manufacturing and Industrial Automation (CIMIA), Athens, pp. 98–105
(1994)
Reibman, A.L.: Modeling the effect of reliability on performance. IEEE Transactions on
Reliability 39(3), 314–320 (1990)
Ripley, B.D.: Neural networks and related methods for classification. Journal of the Royal
Statistical Society, Series B (Methodological) 56(3), 409–456 (1994)
Sanchez, A.M.: FMS in Spanish industry – lessons from experience. Integrated

Manufacturing Systems 5(2), 28–36 (1994)
Smith, A.E., Mason, A.K.: Cost estimation predictive modeling: regression versus neural
network. The Engineering Economist 42(2), 137–161 (1997)
Stoop, P.P.M., Wiers, V.C.S.: The complexity of scheduling in practice. International
Journal of Operations and Production Management 16(10), 37–53 (1996)
Terkaj, W., Tolio, T.: A Stochastic approach to the FMS Loading Problem. CIRP Journal of
Manufacturing Systems 35(5), 481–490 (2006)
Terkaj, W., Tolio, T., Valente, A.: Designing Manufacturing Flexibility in Dynamic
Production Contexts. In: Tolio, T. (ed.) Design of Flexible Production Systems, pp. 1–
18. Springer (2009)
Tompkins, J.A., White, J.A., Bozer, Y.A., Frazelle, E.H., Tanchoco, J.M.A., Trevino, J.:
Facilities Planning. John Wiley and Sons Inc., New York (2002)
Vach, W., Robner, R., Schumacher, M.: Neural networks and logistic regression: Part II.
Vinod, B.: Queuing models for flexible manufacturing systems subject to resource failure.
Ph.D. Thesis, Purdue University, US (1983)
Vinod, B., Solberg, J.J.: Performance models for unreliable flexible manufacturing systems.
International Journal of Management Science 12(3), 299–308 (1984)
Viswanadham, N., Narahari, Y.: Performance modeling of Automated Manufacturing
systems. Prentice-Hall, Englewood Cliffs (1992)
Wang, Z.-M., Yang, J.-G.: Numerical method for Weibull generalized renewal process and
its applications in reliability analysis of NC machine tools. Computers & Industrial
Engineering 63(4), 1128–1134 (2012)
Warner, B., Misra, M.: Understanding neural networks as statistical tools. The American
Statistician 50(4), 284–293 (1996)
Ye, Z.-S., Xie, M., Tang, L.-C.: Reliability evaluation of hard disk drive failures based on
counting processes. Reliability Engineering & System Safety 109, 110–118 (2013)
Gu, Z., Lam, L.H., Dhurjati, P.S.: Feature correlation method for enhancing fermentation
development: A comparison of quadratic regression with artificial neural networks.
Computers and Chemical Engineering 20(suppl. 1), S407–S412 (1996)
Zhou, Y., Zhang, Z., Lin, T.R., Ma, L.: Maintenance optimization of a multi-state series-
parallel system considering economic dependence and state-dependent inspection
intervals. Reliability Engineering & System Safety 5 (2012)
Chapter 11
Producer's Behavior Analysis for AGV System
11.1 Summary
Traditional manufacturing has relied on dedicated mass-production systems to

achieve high production volumes at low costs. As living standards improve and
the demands for new consumer goods rise, manufacturing flexibility gains
prominence as a strategic tool for rapidly changing markets. Flexibility, however,
cannot be properly incorporated in the decision-making process if it is not well
defined and measured in a quantitative manner. Flexibility in its most rudimentary
sense is the ability of a manufacturing system to respond to changes and
uncertainties associated with the production process (Miettinen et al., 2010;
Kumar and Sridharan, 2009; Das et al., 2009). A comprehensive classification of
eight flexibility types was proposed in Browne et al. (1984). Resource and system
flexibilities were examined in Slack (1987), whereas global measures for flexible
manufacturing systems (FMSs) were defined in Gupta and Buzacott (1989).
Routing flexibility based on information theoretic concepts was examined by Yao
and Pei (1990) and Kumar (1987). Flexibility measures for one machine, a group
of machines, and the whole industry were presented in Brill and Mandelbaum
(1989), involving appropriate weights and machine efficiencies in carrying out
sets of tasks.
Artificial Intelligence (AI) methods seem appropriate in most practical
situations where numerical data are not readily available and linguistic variables
are more amenable to handling imprecise knowledge (Gen et al., 2009; Shin et al.,
2009). The flexibility of competing systems can be ranked appropriately using an
algorithmic approach (Abdel-Malek and Wolf, 1991) or a decision support system
(Suresh, 1991) based on performance and economic criteria. According to Turban
and Aronson (1998), a Decision Support System (DSS) is a computer-based
information system that combines models and data in an attempt to solve non-
structured problems with extensive user involvement. They believe that an Expert
System (ES) is a computer system that applies reasoning methodologies on
knowledge to render advice or recommendations much like a human expert. When
expert systems technology was first applied to decision- making problems, it fell
short in several respects. Early expert systems were rule-based. They were not

170 11 Producer's Behavior Analysis for AGV System
capable of handling the classical DSS functions being more computational than
logical. Recently, artificial intelligence researchers have seen the necessity of
using statistical techniques to build intelligent decision support systems (Nolan,
1998; Weiss and Kulikowski, 1991). Examples of such statistical techniques
include fuzzy logic, neural networks, rule induction and various Bayesian
techniques. Turban and Aronson (1998) believed that although uncertainty was
widespread in the real world, its treatment in the practical world of artificial
intelligence was very limited.
Manufacturing costs for products are very crucial in decision making and
strategic planning. And with respect to cost estimation, research and development
departments in the past could only estimate the final product’s total cost.
Moreover, rules of thumb of the engineers are often applied as the cost estimation
benchmarks, making the results controversial in terms of accuracy (Mostafaee et
al., 2010; Eklin et al., 2009; Qian and Ben-Arieh, 2008; Wang, 2007; Verlinden et
al., 2008; Wang, 2007). Although calculation by cost model has the advantage of
being timely, only representative values exclusive of indirect tasks cost and raw
materials cost are calculated resulting in an inadequate accuracy of the estimation.
Niazi et al. (2006) pointed out that Backpropagation Network (BPN) could be
applied for training to deduce unprecedented problems by accumulated knowledge
and information. Specially, it could find out solutions in uncertain circumstances
and have satisfying results in dealing with non-linear problems. Therefore, BPN is
the most popular neural network models being applied and it fits the nature of
product cost estimation the best. McKim (1993) discussed applying BPN in cost
estimation projects. Finally, by integrating a user interface, the demand of a fast
response cost estimation model was met at initial product development stage in his
work. Zhang et al. (1996) categorized cost estimation techniques into traditional
detailed breakdown, simplified-breakdown, group-technology-based, regression-
based and activity-based cost approaches. Ben- Arieh and Qian (2003) divided
cost estimation models into intuitive, analogical, parametric and analytical
approaches. Shehab and Abdalla (2001) proposed intuitive, parametric, variant-
based and generative cost estimating approaches. Cavalieria et al. (2004) proved
three cost analyses of analogy-based, parametric and engineering approaches.
Niazi et al. (2006) discussed, on the basis of the integrated cost estimation
approaches, qualitative and quantitative cost estimation techniques pointing out
the key advantages and limitations of each cost estimation technique.
Jobshop is a flexible, scalable and intelligent production planning and control
system offering advanced functionality and value in key areas of manufacturing
and assembly. The Flexible Jobshop Problem (FJP) is an extension of the classical
jobshop problem allowing for an operation to be processed by any machine in a
given set. Automated Guided Vehicles (AGVs) reduce the cost of manufacturing
and increase efficiency in a manufacturing system. These trailers can be used to
move raw materials in line to get them ready for manufacturing (Aized, 2009;
Hsueh, 2010). To conceptualize an AGV, it is necessary to understand the
fundamentals of Flexible Manufacturing Systems (FMSs).
Here, soft computing techniques are considered for configuring an expert

system in the presence of uncertainty. The purpose of this study is to design a
model for an AGV based automated manufacturing system. Three main factors
accompanied by their sub-factors are chosen and then propose a neural network
based approach to analyze the effects of the sub-factors on the main factors. Due
to uncertainty of cost elements in the manufacturing environment, fuzzy logic is
applied to determine an appropriate range for each of the linguistic variables.
Then, using the linguistic variables, the possible fuzzy rules are extracted and
using a multiple linear regression model the significant rules are investigated and
eliminate inappropriate rules. A sensitivity analysis is associated with the
regression method. Then, considering time as a triangular fuzzy number, an
aggregation process to combine both the obtained cost and the time parameters is
proposed. Then a dynamic program is proposed for finding an optimal path in the
proposed manufacturing system.
carries raw material, semi-produced and final products in batch sizes. Due to
shows up in the shop's specifications that affect the production cost. As a result,
the system would consider a flexible jobshop model where multi shops of the
same duty exist and each operation can be processed on any type of machine in
any shop. The sequences of jobs are specified and the jobs are assumed to be
independent.
rates of defect, etc.. Each flow path for the AGV is associated with a time
parameter and also a cost parameter related to each shop. The aim is to find a path
for the AGV minimizing an aggregate time and cost objective. Considering the
variability of the AGV flow among shops, the time of each flow path of the AGV
is a triangular fuzzy number. In each shop, different machines and operators are
working. Due to unpredictable events during working times a cost may incur. This
cost is inferred from an expert system via a fuzzy neural network. Cost parameters
for each shop are considered to be 3: (1) Equipment sensitivity, (2) operator
proficiency, and (3) product specifications, each being specified by one of the
three levels of low, moderate, and high. A configuration for the proposed problem
is presented in Figure 1.
ost
e, C
Tim
Fig. 1 A configuration of the proposed problem
As stated before, time is considered to be a triangular fuzzy number and cost is

inferred from an expert system. In our expert system a backpropagation neural
network is considered to estimate the cost factors (outputs) using their
corresponding sub-factors (inputs). The backpropagation algorithm trains a given
feed-forward multilayer neural network for a given set of input patterns with
known classifications. The backpropagation algorithm is based on Widrow-Hoff
delta learning rule in which the weight adjustment is done through mean square
error of the output response to the sample input (Abdi et al., 1996).
The cost factors and their related sub-factors are shown in Figure 2.
Fig. 2 The cost factors and their related sub-factors.

Using the existing data, the network is trained and then by the resulting pattern
one can apply the model to obtain the output with respect to the proposed sub-
factors. The numerical results would show a lower bound and an upper bound for
each cost factor. These bounds are utilized as our cost factors ranges for the fuzzy
rule base. Considering the uncertainty in the cost factors due to dynamic changes
in neural network inputs, the cost factors in three levels of low, moderate, and
high is specified. The problem is to specify the range of these levels; that is, the
numerical range of low, moderate or high for each of the cost factors.
An inductive reasoning technique to obtain the appropriate range for the levels is
applied. This method is based on an ideal scheme describing the input and output
relationships for a well-established data base. This method is called entropy. A
key goal of entropy minimization analysis is to determine the quantity of
information in a given data set. The entropy of a probability distribution is a
measure of the uncertainty of the distribution. This information measure estimates
the uncertain range of data using a predetermined inappropriate range to start the
process. The higher the prior estimate of the probability for an outcome to occur,
the lower will be the information gained by observing it to occur. The entropy on a
set of possible outcomes of a trial where exactly one outcome is possible is
defined by the sum of probabilities. In other words, the entropy is the expected
value of information. For a simple one-dimensional (one uncertain variable) case,
let us assume that the probability of the ith sample wito be true is p(wi ) . If the
sample wi is actually observed in the future and discover that it is true, then the
following information, I(wi) is gained:
I (wi ) = −k ln p(wi ) , (1)
where k is a normalizing parameter. If it is discovered to be false, we still gain

some information:
I (wi ) = −k ln[1 − p(wi )] . (2)
Thus, the entropy of the inner product of all the samples (N) is:
N
S = −k ∑ [ pi ln pi + (1 − pi ) ln(1 − pi )], (3)
i =1
where piis the probability of the ith sample to be true. Note that S ≥ 0 , because
ln x ≤ 0 , for 0 ≤ x ≤ 1 .
The entropy of a rule should be minimized. Minimum of entropy S is associated
with all the pi being as close to one or zero as possible, which in turn implies that
they have a very high probability of either happening or not happening,

respectively. Note in equation (3) that if pi=1, then S=0.
This result makes sense, since pi is the probability measure of whether a value
belongs to a partition or not. The precedence for partitioning the data is explained
next.
First, seeking a threshold value for a sample in the ranges [x1, x] and [x, x2] are
assumed. It is denoted that the first region p and the second region q. by moving
an imaginary threshold value x between x1 and x2, entropy is calculated for each
value of x.
An entropy with each value of x in the region containing x1 and x2 is:
S ( x) = p( x) S p ( x) + q( x) S q ( x) , (4)
where,
S p ( x) = −[ p1 ( x) ln p1 ( x) + p2 ( x) ln p 2 ( x)] , (5)
S q ( x) = −[q1 ( x) ln q1 ( x) + q2 ( x) ln q2 ( x)] , (6)
with pk(x) and qk(x) as conditional probabilities that the class k sample is in the
regions [x1, x1+x] and [x1+x, x2], respectively, p(x) and q(x) are probabilities that
all samples are in the regions [x1, x1+x] and [x1+x, x2], respectively. Also,
p ( x) + q ( x) = 1 . (7)
A value of x that gives the minimum entropy is the optimum threshold value.
The entropy estimates pk(x), qk(x), p(x), and q(x) are calculated as follows:
nk ( x) + 1
p k ( x) = , (8)
n( x ) + 1
N k ( x) + 1
q k ( x) = , (9)
N ( x) + 1
n( x)
p ( x) = , (10)
n
q( x) = 1 − p( x) , (11)
where nk(x) is the number of class k samples located in [x1, x1+x], n(x) is the total
number of samples located in [x1, x1+x], Nk(x) is the number of class k samples
located in [x1+x, x2], N(x) is the total number of samples located in [x1+x, x2], and
n is the total number of samples in [x1, x2]. While moving x in the region [x1, x2],
the values of entropy are calculated for each position of x.
This procedure finds the region for the levels of the cost factors. After each
update in input data, the ranges are specified using an entropy technique. After
determining the regions of the cost factors, the fuzzy rules are then composed
which clarify the total cost for each shop. To find the effective rules, a multiple
linear regression model is configured using the previous data and identify the
regression coefficients. Since we have three cost factors each with three levels of
low, moderate and high, then twenty seven (3*3*3) possible rules exist. Thus,
using the test of hypothesis, the effective rules are identified.
As stated before, some rules affect the cost of the system. Whether a rule is
significant for cost or not should be investigated. One way to survey the effect of
Therefore, the following equation is considered,
ui = β 0 + β1r1 + ... + β m rm + ε i , i = 1,...,n, (12)
whereui is the total cost in the ith period, β0 is the intercept, the βj are the
coefficients for the rj and εi is the error term (note that the β j , 0 ≤ j ≤ m ,
depend on the period i, but for simplicity the index i is omitted here). The aim is to
identify the β j which are not important for the total cost of the proposed system.
Here, a two-sided hypothesis testing is applied as follows;
H 0 : β j = 0;
(13)
H 1 : β j ≠ 0.
The test of the null hypothesis H0 against the two sided alternative proceeds in
three steps. The first is to compute the standard error of β j , SE ( β j ) . The
second step is to compute the t-statistic,
βj −0
t= , (14)
SE ( β j )
and the third step is to compute the p-value, the probability of observing a value of
act
assuming that the null hypothesis is correct.

A small value of the p-value, say less than 5%, provides evidence against the
null hypothesis in the sense that the chance of obtaining a value of βj by pure
random variation from one sample to the next is less than 5% if, in fact, the null
hypothesis is correct. If so, the null hypothesis is rejected at the 5% significance
level. Simply, one can reject the null hypothesis at the 5% significance level if
t act > 1.96 . Therefore, if the hypothesis is accepted, then it is ensued that the
corresponding rule is not important and thus is not effective on the cost. The
estimation of the coefficients is performed using a regression software.
By these rules the cost of each shop is estimated (arc length in the network).
Consequently, a dynamic program is applied to find the optimal path in the
proposed jobshop automated manufacturing network. A flowchart for the
proposed expert system is presented in Figure 3.
Fig. 3 A proposed expert system
Our decision model is to consider both time and cost parameters, and thus an
integration of the two parameters would be required. The integration is to have a
weighted sum of cost and time as an arc length in the proposed network. Time is a
triangular fuzzy number and cost is considered to be an indirect triangular fuzzy
number.
As stated before, time is a triangular fuzzy number, ensued from the experts'
knowledge. Brainstorming and expert knowledge versus meta-rule techniques
(neural networks, genetic algorithms, etc.) are two common approaches for
defining fuzzy rules and membership functions. While available empirical data is
requisite for using the second approach, due to unavailability of historical data for
cost, we make use of the first approach to obtain the membership functions and
fuzzy rules.
In Zadeh's words (Zadeh, 1965), fuzzy logic is a tool for ‘‘Computing with
Words’’. He stated that the main role of fuzzy logic was to serve as a methodology
for computing with words when no other methodology could attain such purpose
(Zadeh, 1996, 1999). Triangular membership functions are common (Pedrycz,
1994). Mamdani fuzzy system is being commonly used in the literature (Klir and
Yuan, 1995; Zimmermann, 1996). The fuzzy set acquired from aggregation of
rules’ results will be defuzzified using defuzzification methods like centroid
(center of gravity), max membership, mean-max, and weighted average. The
centroid method is very popular, in which the ‘‘center of mass’’ of the result
provides the crisp value. For our problem in which various possible conditions of
parameters are stated in forms of fuzzy sets, the Mamdani fuzzy system is utilized,
because the fuzzy rules representing the expert knowledge in Mamdani fuzzy
systems would consider fuzzy sets.
The input to our Mamdani type fuzzy system (Babuska, 1998) is composed of
equipment sensitivity, operator proficiency, and product specifications. For any of
the inputs, three linguistic terms of "low", "moderate", and "high" are defined. The
output of the system is cost being identified by any of the three linguistic terms,
"low", "moderate", or "high". The maximum membership grade of linguistic term
‘‘high’’ is 30.
As a result, a triangular fuzzy number as the time and a numerical value as the
cost are obtained. We intend to consider an integrated time-cost value as the value
of each proposed arc in the network. Time and cost having different scales, it
would not be possible to perform basic operations such as addition on their
original forms. Thus, to remove the scales, the time and cost values are normalized
throughout the network.
As stated, to perform basic operations on the time and cost parameters with
their different scales, we need to remove their scales. To do this, two
normalization processes are proposed for time and cost, separately. In (15), (16)
and (17) below, i is an index to show the node and j is an index to show the stage
(shop’s type) in the proposed network, rijis the value of either time or cost
parameter in each node, and nijis its corresponding normalized value. Considering
time as a triangular fuzzy number, the normal value with a positive view is given
by
rij − r jmin
nij = , ∀i, j, (15)
r jmax − r jmin
while the normal value with a negative view is given by,
r jmax − rij
nij = , ∀i, j , (16)
r jmax − r jmin
max
where r j and rjmin are maximum and minimum values in each column of an
assumed matrix of time or cost, respectively. Since time in our proposed model is
a criterion implicating a negative aspect in decision making, then (16) is chosen
for normalizing time in our approach.
Assuming cost as a crisp value, the cost values are normalized as follows:
rij
nij = , ∀i , j , (17)
∑ rkj2
k
To compute the minimum or maximum value for (15) and (16), comparisons
are needed to be made. This means that it is necessary to have a method for
ranking and comparing fuzzy numbers. To compare fuzzy numbers, the efficient
approach developed by Mahdavi et al. (2009) based on the distance between fuzzy
numbers is used. They used the distance function introduced by Sadeghpour
Gildeh and Gien (2001). The main advantages of this distance function are the
generality of its usage on various fuzzy numbers, and its reliability in
distinguishing unequal fuzzy numbers.
Then, a weighing method is considered to make the time and cost appropriate
for basic operations.
To weigh the parameters, a multi criteria decision making approach is taken.
Multi-Criteria Decision Making (MCDM), dealing primarily with problems of
evaluation or selection (Keeney and Raiffa, 1976; Teng, 2002), is a rapidly
developing area in operations research and management science. The Analytical
Hierarchy Process (AHP), developed by Saaty (1980), is a technique of
considering data or information for a decision in a systematic manner
(Schniederjans and Garvin, 1997). AHP is mainly concerned with the way to solve
based on three principles: (1) constructing the hierarchy, (2) priority setting, and
(3) logical consistency. AHP is applied to weigh the parameters. In AHP, multiple
paired comparisons are based on a standardized comparison scale of nine levels
(Saaty, 1980).
Here, the weights for the parameters of each arc in the proposed network are
n
obtained. If C n and T are considered as normalized cost and time, respectively,
then the total weighted normalized value of each arc is determined as follows:
P= (ψ × T n ) + (ψ '×C n ) . (18)
On the other hand, time is a triangular fuzzy number and cost is a crisp
numerical value. But, it is needed to have triangular numbers for both parameters
to perform the basic operations. For this, cost is considered as a trivial triangular
fuzzy number and show it by a triplet (0,C,0), where C is the numerical value
inferenced from the expert system for cost with the membership value equal to 1.
Several fuzzy ranking methods have been proposed (Bortalan and Degani,
1985; Luis and Antonio, 1989; Kim and Park, 1990). Since the graded mean
integration representation method (Chen and Hseih, 2000) not only alleviates
some drawbacks of the existing methods, but also possesses the advantages of
being easily implementable and quite effective in problem solving (see Lee et al.,
2007), we will use it to transform the total weighted normalized value of each arc
in our proposed network. If a = ( a1 , a2 , a3 ) is a triangular fuzzy number, then
~ ) ) is defined to be:
the graded mean integration ( R (a
a + 4a 2 + a 3
R (a~ ) = 1 . (19)
6
Equation (19) is applied to transform the fuzzy numbers to crisp values and use
them to find the optimal path.
Shortest Path in a Network

Let G = (V,A) be a graph, where V = {1,..., N } is the set of nodes, and A ⊂ V ×V is
in the graph. Next, a fuzzy dynamic programming approach for computing the
optimal path is described.
An Algorithm for Computing a Fuzzy Shortest Path

The following dynamic programming algorithm is for computing the shortest path
in a network. The algorithm is based on Floyd’s dynamic programming method to
find a shortest path, if it exists, between every pair of nodes i and j in the network
(see Floyd, 1962, and also Dreyfus and Law, 1977). The following optimal value
function f k (i, j) and the corresponding labeling function Pk (i, j) are used:
f k (i, j) : length of the shortest path from node i to node j when the path is
considered to use only the nodes from the set nodes {1,..., k }.
Pk (i, j) : the last intermediate node on the shortest path from node i to node j
using {1,..., k } as intermediate nodes.
The dynamic updating for the optimal path length and its corresponding
labeling are:
f k (i, j) = min{ f k −1 (i, j), f k −1 (i, k ) + f k −1 (k , j)},
⎧ Pk −1 (i, j ) if k is not on the shortest path from i to j using {1,..., k}

Pk (i, j ) = ⎨
⎩ Pk −1 ( k , j ) otherwise
We are now ready to give steps of the algorithm.
Algorithm 1:A dynamic programming method for computing a shortest path

in a fuzzy network with triangular fuzzy arc lengths.
Give G = (V , A ) , where V is the set of nodes with V = N , and A is the set of

arcs.
~ ~ ~
Step 0: Let k=0 and f k (i, j) = dij , for all (i , j ) ∈ A , f k = (i, j) = ∞ , for all
(i , j ) ∉ A . If an arc exists from node i to node jthen let Pk (i, j ) = i .
Step 1: Let k = k + 1 .
Do the following steps for i = 1,2,3,..., N , j = 1, 2,3,..., N , i ≠ j.
1.1 [
Compute the value of f k (i, j) = min f k −1 (i, j), f k −1 (i, k ) + f k −1 (k, j) , ]
(for comparison of fuzzy numbers, where the D p , q approach is used for
computing the distances).
1.2 If node k is not on the shortest path using {1,2,..., k } as intermediate nodes,
then let Pk (i, j) = Pk −1 (i, j ) else let Pk (i, j) = Pk −1 (k , j) .
Step 2: If k < N then go toStep 1.
~
Step 4: Obtain the shortest path using Pk (i, j) . If f N = (i, j) = ∞ , then there is
no path between i and j. the shortest path from node i to j, if it exists, is identified
backwards and read by the nodes: j, PN (i, j) = k followed by
PN (i, k ),...,PN (i, l ) = i , where l is the node immediately after i in the path.
Sensitivity Analysis
Analyzing Producer’s Behavior
Here, a comprehensive analysis is worked out for the obtained paths in last
section. We obtained a path and its corresponding value. These values are
considered as costs for each path. The aim is to find the quantity of products to be
carried by AGV to maximize the profit of a producer. Obviously, profit is the
difference between total revenue and total expenses in a perfect competitive
market. The notations of profit maximization model are as follows:
Index:
j Counter for paths; j=1,2,…,m.
Parameters:
pj Price of product in the jth path.
qj Quantity of product in the jth path.
rj Numerical value for path j.
xj Quantity of input material in path j.
Fj Product transformation function of q1,...,qm .
Then, the mathematical model for profit is:

m m m m
Max π = ∑ p j q j − ∑ r j x j ⇒ π = ∑ p j q j − ∑ r j Fi (q1 ,..., q m ) . (20)
j =1 j =1 j =1 j =1
Obviously, the maximization function can be turned into minimization as

follows:
m m
Min - π = ∑ ri Fi (q1 ,..., q m ) − ∑ pi qi , (21)
i =1 i =1
where the xi are functions of the quantities of the outputs qj ( xi = Fi (q1 ,...,qm ) ).
It means that, considering the inputs given to the manufacturing system, some
outputs are produced. The relationships among inputs and outputs are stated as
mathematical functions. The model is an unconstrained nonlinear problem. To
solve the profit maximization problem for a producer desiring to maximize the
profit using the above model, the partial derivatives are set equal to zero (first
order conditions):
∂ (−π ) m ∂Fi m
= ∑ ri − p j = ∑ ri f i − p j = 0 , (22)
∂q j i =1 ∂q j i =1
∂xi
where fi is the marginal cost, fi = , and the marginal productivity is
∂q j
1 1
MP = = .
fi ∂xi
∂q j
To investigate whether the obtained quantity of product is optimal or not, the
second order conditions are checked. The second order conditions for the
maximization of profit require that the relevant Hessian of π (or − π ) be
negative (or positive) semi-definite (if the Hessian of π (or − π ) is not negative
(or positive) semi-definite, then we are sure that the obtained path is not a
maximizer). Here, the Hessian of ( − π ) is:
∂ 2π
[∇ 2
]
(−π ) ij = −
∂qi ∂q j
, (23)
To be positive semi-definite, the eigenvalues of ∇ 2 (−π ) , the λi , must

satisfy:
λi ≥ 0, ∀i , (24)
If the solution point is so that ∇ 2 (−π ) is positive definite, that is,
λi > 0, ∀i , (25)
then the point is a local maximizer (or minimizer) of π (or − π ).

This way, an application of the obtained optimal paths have been seen. The
optimization scheme helps the producer to handle economic batch sizes for the
AGV to attain more profit considering the specified prices and costs.
Path Reliability
It is necessary to incorporate reliability into the model to ensure the level of
service for every AGV in every path. For modeling reliability, the approach of
Ball and Lin (1993) is adopted for extension.
The reliability is defined as the probability that the system functions until time
t. If an AGV in a path is broken down, it can be regarded as a failure. A desired
level of reliability can be achieved by limiting the failure probabilities. This

approach for handling reliability is called the method of chance constraints and
was initially suggested by Charnes and Cooper (1959) in the context of
mathematical programming. The use of chance constraints in vehicle routing
problem was illustrated by Stewart and Golden (1983). Carbone (1974) used
chance constraints for selecting multiple facilities under normally distributed
demands. The model minimized an upper bound on the total demand-weighted
distance while ensuring that constraints were satisfied with specified chance or
probability. Shiode and Drezner (2003) used a similar approach for a competitive
It is assumed that the reliability of moving an AGV on each arc is
independently exponentially distributed. We now discuss the reliability based
model. Let R j (t ) be the probability of an AGV functioning on arc j until time t.
Since the arcs composing a path are in series, then the corresponding reliability is
computes to be
m
R(t ) system = ∏ R j (t ). (26)
j =1
Therefore, the reliability of the system is determined to satisfy
⎛ m ⎞
⎜ ∏ R j (t ) ⎟ ≥ α , (27)
⎜ ⎟
⎝ j =1 ⎠
where α is the lower bound for a desirable reliability of the system until time t.
As previously assumed, the reliability of each arc is independently exponentially
distributed:
−t
θj
R j (t ) = e . (28)
Therefore, we need to have
⎛ m θ− t ⎞
⎜ e j ⎟ ≥α .
⎜∏
(29)
⎟
⎝ j =1 ⎠
Using the inequality (29), the θj is obtained considering a confidence level.
Now, we investigate the θj for arcs separately by data collection. The aim is to
analyze the θj obtained from data collection not having any interaction with other
arcs, as compared to the θj obtained by (29). Data in a specific working time

t=24 are collected. Then, for t>24 our data are censored (type I censored data).
Assume x1:n, x2:n, … , xr:nare the r censored data in a specific path. Then, θˆ , the
estimated θ , is obtained by maximum likelihood estimation (MLE) as follows.
r
∑x i :n + ( n − r )t 0
θˆMLE = i =1
, (30)
r
where t0 is the end time of observation (t=24) and r is the number of failures.
2rθˆ
Here, using ≈ χ 2 (2r ) as a pivot (2r is the degrees of freedom), a
θ
confidence interval for θ can be set.
2rθˆ
To set a confidence interval using ≈ χ 2 ( 2r ) ,
θ
⎡ 2rθˆ ⎤
P ⎢ χ 2 α (2r ) < < χ α2 ( 2r ) ⎥ = 1 − α . (31)
⎣ 1− 2 θ 2 ⎦
Inverting (31) and multiplying 2rθˆ into all fractions,

⎡ ⎤
⎢ 2rθˆ 2rθˆ ⎥
P⎢ 2 <θ < 2 =1−α , (32)
χ α (2r ) χ α ( 2r ) ⎥
⎢ 1− ⎥
⎣ 2 2 ⎦
where (32) is a two sided ( 1 − α )% confidence interval based on type I censored
data.
11.4 Conclusions
An approach for finding an optimal path in a flexible jobshop manufacturing

system considering two criteria of time and cost is proposed. The proposed
flexible jobshop system has more than one shop with the same duty. The
difference among shops with the same duty is in their machines with various
specifications. The shops configure a network in which they are considered as
References 185
nodes and the paths among them are considered as network arcs. An Automated
Guided Vehicle (AGV) functions as a material handling device through the
manufacturing network. Time is considered to be a triangular fuzzy number and
cost is inferred from an expert system considering three parameters of equipment
sensitivity, operator proficiency, and product specification via linguistic variables.
This cost is associated with some main factors and some sub-factors. To
approximate the cost, a backpropagation neural network was applied to estimate
the sub-factors with the corresponding main factors and compute a bound for the
main factors. These bounds were processed by the entropy technique to obtain
regions for the linguistic variables. After determining the regions for the linguistic
variables, we composed the rules and verified their significance using regression
analysis. The purified rules were used to infer the cost of the system. The aim was
to find a path which minimized an aggregate weighted unscaled time and cost
objective. Since time and cost had different scales, a normalization process was
used to remove the scales and because the model was bi-objective, the AHP
weighing method was applied to gain a single objective. A fuzzy dynamic
programming approach was used to compute a shortest path in the proposed
network. Also, some sensitivity testing on the obtained shortest paths are
performed. Finally, the proposed approach was fully illustrated by a numerical
example.
References
Abdel-Malek, L., Wolf, C.: Evaluating flexibility of alternative FMS designs-A
comparative measure. Int. J. Prod. Econ. 23(1), 3–10 (1991)
Abdi, H., Valentin, D., Edelman, B., O’Toole, A.J.: A Widrow-Hoff learning rule for a
generalization of the linear auto-associator. Journal of Mathematical Psychology 40,
175–182 (1996)
Industrial Engineering 57(3), 822–831 (2009)
Babuska, R.: Fuzzy Modeling for Control. Kluwer Academic Publishers, Norwell (1998)
Ben-Arieh, D., Qian, L.: Activity-based cost management for design and development
stage. International Journal of Production Economics 83, 169–183 (2003)
Bortolan, G., Degani, R.: A review of some method for ranking fuzzy subsets. Fuzzy Sets
and Systems 15, 1–19 (1985)
Brill, P.H., Mandelbaum, M.: On measures of flexibility in manufacturing systems. Int. J.
Prod. Res. 27(5), 747–756 (1989)
Carbone, R.: Public facilities location under stochastic demand. INFOR 12(3), 261–270
(1974)
Cavalieria, S., Maccarroneb, P., Pinto, R.: Parametric vs. neural network models for the
estimation of production costs: A case study in the automotive industry. International
Journal of Production Economics 91, 165–177 (2004)
Charnes, A., Cooper, W.: Chance-constraint programming. Management Science 6(1), 73–
79 (1959)
Chen, S.H., Hsieh, C.H.: Representation, ranking, distance, and similarity of L-R type
fuzzy number and application. Aust. J. Intell. Process. Syst. 6(4), 217–229 (2000)
Engineering 56(1), 283–293 (2009)
Delgado, M., Verdegay, J.L., Vila, M.A.: A procedure for ranking fuzzy numbers using
fuzzy relations. Fuzzy Sets and Systems 26, 49–62 (1988)
Eklin, M., Arzi, Y., Shtub, A.: Model for cost estimation in a finite-capacity stochastic
environment based on shop floor optimization combined with simulation. European
http://EViews.com
Engineering 56(3), 779–808 (2009)
Gupta, D., Buzacott, J.A.: A framework for understanding flexibility of manufacturing
systems. J. Manuf. Syst. 8(2), 89–97 (1989)
Hsueh, C.F.: A simulation study of a bi-directional load-exchangeable automated guided
vehicle system. Computers & Industrial Engineering 58(4), 594–601 (2010)
Keeney, R., Raiffa, H.: Decision with multiple objectives: Preference and value tradeoffs.
Wiley and Sons, New York (1976)
Kim, K., Park, K.S.: Ranking fuzzy numbers with index of optimism. Fuzzy Sets and
Systems 35, 143–150 (1990)
(1995)
Manufacturing 25(4-5), 829–838 (2009)
Kumar, V.: Entropic measures of manufacturing flexibility. Int. J. Prod. Res. 25(7), 957–
966 (1987)
Lee, K.L., Huang, W.C., Teng, J.Y.: Locating the competitive relation of global logistics
hub using quantitative SWOT analytical method. Quality & Quantity 43(1), 87–107
(2007)
Luis, M., Antonio, G.M.: A subjective approach for ranking fuzzy numbers. Fuzzy Sets and
Systems 29, 145–153 (1989)
Mahdavi, I., Nourifar, R., Heidarzade, A., Mahdavi-Amiri, N.: A dynamic programming
approach for finding shortest chains in a fuzzy network. Applied Soft Computing 9(2),
503–511 (2009)
McKim, R.A.: Neural network applications to cost engineering. Cost Engineering 35, 31–
35 (1993)
Mostafaee, A., Saljooghi, F.H.: Cost efficiency measures in data envelopment analysis with
data uncertainty. European Journal of Operational Research 202(2), 595–603 (2010)
References 187
Engineering 128, 563–575 (2006)
Nolan, A.: An intelligent system for case review and risk assessment in social services. AI
Mag. 19(1), 39–46 (1998)
Pedrycz, W.: Why triangular membership functions? Fuzzy Sets and Systems 64, 21–30
(1994)
Qian, L., Ben-Arieh, D.: Parametric cost estimation based on activity-based costing: A case
study for design and development of rotational parts. International Journal of
Production Economics 113(2), 805–818 (2008)
Sadeghpour Gildeh, B., Gien, D.: La distance-Dp,q et le cofficient de corrélation entre deux
variables aléatoires floues. In: Actes de LFA 2001, Monse-Belgium, pp. 97–102 (2001)
Schniederjans, M.J., Garvin, T.: Using the analytic hierarchy process and multi-objective
programming for the selection of cost drivers in activity based costing. European
Shehab, E.M., Abdalla, H.S.: Manufacturing cost modeling for concurrent product
development. Robotics and Computer-Integrated Manufacturing 17, 341–353 (2001)
Shin, M., Mun, J., Jung, M.: Self-evolution framework of manufacturing systems based on
fractal organization. Computers & Industrial Engineering 56(3), 1029–1039 (2009)
Slack, N.: The flexibility of manufacturing systems. Int. J. Prod. Oper. Manag. 7(4), 35–47
(1987)
Suresh, N.C.: An extended multi-objective replacement model for flexible automation
investments. Int. J. Prod. Res. 27(9), 1823–1844 (1991)
Teng, J.-Y.: Project Evaluation: Method and Applications. National Taiwan Ocean
University, Taiwan (2002)
Turban, E., Aronson, J.: Decision Support Systems and Intelligent Systems. Prentice-Hall,
London (1998)
Verlinden, B., Duflou, J.R., Collin, P., Cattrysse, D.: Cost estimation for sheet metal parts
using multiple regression and artificial neural networks: A case study. International
Journal of Production Economics 111(2), 484–492 (2008)
Wang, H.S.: Application of BPN with feature-based models on cost estimation of plastic
injection products. Computers & Industrial Engineering 53(1), 79–94 (2007)
Wang, Q.: Artificial neural networks as cost engineering methods in a collaborative
manufacturing environment. International Journal of Production Economics 109(1-2),
53–64 (2007)
Wang, X., Kerre, E.E.: Reasonable properties for the ordering of fuzzy quantities (I). Fuzzy
Sets and Systems 118, 375–385 (2001)
Weiss, N., Kulikowski, I.: Computer Systems That Learn. Morgan Kaufmann, California
(1991)
Yao, D.D., Pei, F.F.: Flexible parts routing in manufacturing systems. IIE Trans. 22(1), 48–
55 (1990)
Zadeh, L.A.: Fuzzy sets. Information and Control 8, 338–353 (1965)

Zadeh, L.A.: Fuzzy logic = computing with words. IEEE Transactions on Fuzzy Systems 4,
103–111 (1996)
Zadeh, L.A.: From computing with numbers to computing with words – From manipulation
of measurements to manipulation of perceptions. IEEE Transactions on Circuit
Systems 45, 105–119 (1999)
Zhang, Y.F., Fuh, J.Y.H.: A neural network approach for early cost estimation of packaging
products. Computers & Industrial Engineering 34, 433–450 (1998)
Zimmermann, H.J.: Fuzzy Set Theory and its Applications, 3rd edn. Kluwer Academic
Publishers (1996)
Chapter 12
Risk for Multiple AGV System
12.1 Summary
A standard formula for the quantitative definition of risk is, Risk = P (loss) × L(loss) ,
where risk is the function of the probability (P) of loss and the significance of its
consequences (L) (Manuj and Mentzer, 2008). Hetland (2003) and Diekmann et al.
(1988), on the other hand, view risk as the implication of an uncertain phenomenon.
Waters (2007) explains the difference: risk occurs because there is uncertainty about
the future, which means that unexpected events may occur. Knight’s (1921)
distinction between certainty, risk and uncertainty is probably the best known and
most used typology of uncertainty for risk management. In his definition of risk
Knight coined the terms (quantitative) ‘‘measurable’’ uncertainty and (non-
quantitative) ‘‘un-measurable’’ uncertainty when there is only partial knowledge of
outcomes in the form of beliefs and opinions (Vilko and Hallikas, 2012).
The researchers’ role in the discussions was to present the data and guide the
discussion in a holistic direction. Consensus was finally reached about the risks,
their categorization and impact. The risk drivers found from the manufacturing
system were classified by source. We found that an adaptation of Manuj and
Mentzer’s (2008) risk source classification with its wide perspective on
manufacturing risk management provided a solid framework for our case; the
classification is both qualitative and quantitative, taking into account both the
direct and indirect impacts, and this facilitates in depth understanding of the risk
sources without losing the holistic view. The risks were categorized as follows:
Supply Risks, Security Risks, Operational Risks, Macro Risks, Policy Risks and
Environmental Risks. No risk sources fitting the Demand Risks, Competitive
Risks and Resource Risks classification were identified in the group discussions
(Pons, 2010). Having identified the risks and made their semi-quantitative
assessment, the expert panel turned to the delay impact, which was modeled in the
form of triangular distributions representing the minimum, the most likely and the
highest impact. Risk effects can be categorized in three different types: time-
based, finance-based and quality-based. Time-based effects either delay or disrupt
the material flow of the manufacturing system. In this case the disruption was

190 12 Risk for Multiple AGV System
identified as a breakdown in the chain such that the goods do not reach their
destination by the time they are expected by the final customer. There are no clear
time limitations on the delay due to the fact that it had significantly different
consequences in different stages of the system. Manufacturing transportation
could be delayed for times without serious consequences. The risk impact in this
case varied highly depending on the goods. Risk likelihood and impact were
evaluated on a scale of 0,1,3,9, where 0 implies zero likelihood or no impact and 9
denotes a very high likelihood or impact. The scale is adapted from the Quality
Function Deployment design method (see e.g., Akao, 2004), which is commonly
used in the context of new product/ service development.
Manufacturing risk assessment is a supporting tool for the contractor and
program office decision-making process. It seeks to estimate the probabilities of
success or failure associated with the manufacturing alternatives available (Nau et
al., 2012). These risk assessments may reflect alternative manufacturing
approaches to a given design or may be part of the evaluation of design
alternatives, each of which has an associated manufacturing approach.Risk
management is an overarching process that begins during the earliest stages of a
program and continues throughout its entire life cycle. Risk encompasses the
following steps:
 Risk identification;
 Risk analysis;
 Risk mitigation planning;
 Risk mitigation plan implementation; and
 Risk tracking.
Assessing manufacturing risks is a requirement, and it is required as early as
pre-Milestone A where the Analysis of Alternatives (AoA) is required to assess
the "manufacturing feasibility" of the proposed approach.
As a system progresses through its definition, design, development, testing and
fielding, more information becomes available concerning the system's risk. If the
risk management process is conducted continuously, then new information will
lead to identifying and analyzing new risk root causes, and identifying and
implementing mitigation plans for them. It will also lead to re-analyzing
previously identified risk root causes, and re-evaluating and adjusting mitigation
plans already in place. This continuous activity allows the PM to focus valuable
program resources where they can be most effective, and shift resources as new
future root causes are discovered and others are re-evaluated.
Iterative Systems Engineering process is the perfect vehicle for helping
manufacturing managers to identify risk early through technical reviews and
audits and to support the development of plans and mitigations to reduce those
risks.
Critical success factors refer to identifying the factors that must be successfully
mastered to execute a successful risk management program. Some examples of
risk management critical success factors include:
12.1 Summary 191
 Clearly define and establish feasible, stable, and well-understood user

requirements;
 Establish a close partnership with users, industry, and other key
stakeholders;
 Comprehensively plan, formally document, and continuously apply the
risk management process, and ensure it is integral to all program
processes;
 Use continuous, event-driven technical reviews as part of the risk
management process; and
 Clearly define criteria for assessing the effectiveness of implemented risk
mitigation actions.
Risk is time phased and should be tied to appropriate maturity models such as the
Technology Readiness Level (TRL), Manufacturing Readiness Level (MRL) and
Sustainment Readiness Model (SRM) that are considered best practices. Other
chapters will discuss these models. These models provide for an assessment of a
technology, manufacturing process, logistics/sustainment considerations of a
component, subsystem, or weapon system. These models have been structured to:
 Define the current level of maturity;
 Identify maturity shortfalls and associated cost and risk; and
 Provide a basis for investments to mature the component, subsystem, or
weapon system and thereby manage risk.
The risks were analyzed in terms of their effects on the manufacturing system.
Some of the investigated organizations had severe problems informing a holistic
and clear view, and clear over estimations as well as underestimations were
evident in their assessment of the risk impact beyond their own functions. Expert
panel discussions were held in order to verify the risk values (Nau et al., 2011).
The first of these took place during the interview process in order to discuss the
preliminary findings from the interviews conducted thus far and to explore the
different viewpoints. The expert panel consisted of logistics field researchers and
port operations experts in the case manufacturing system.
Here, a manufacturing network is proposed and the defects are determined in
order to analyze risk and compute loss. The aim is to obtain paths with less loss
and more reliability. With respect to the multi stage decision making process of
the proposed network, we develop a dynamic program being a useful tool for
multi stage decision making. To counteract the dynamism of digital data in
different time periods, a Bayesian approach is employed to determine the loss
function of moving through the stages of the proposed network.
The main contribution of this work is to propose an integrated methodology
under uncertainty for risk evaluation in very high automated technology of
material handling in manufacturing. Since, in past works risk evaluation for high
technology systems neglected to consider uncertainty of dynamic changes in
automated manufacturing system. Also, considering risk for AGV systems is
novel.
Several AGVs are considered for material handling purpose in a manufacturing

system consisting of workstations to process parts to form products. To
conceptualize a network, the workstations are proposed to be nodes and the AGV
guide paths are the arcs. Due to demands sent to the AGVs, they carry the parts on
the arcs thus AGVs may have conflict. Also, due to high technology employed in
AGV configurations defects are inevitable.
Suppose that there are several kinds of parts that should be transported by AGVs
from node A to node B. The distance between A and B is t and the velocities of the
AGVs are given by 1 .Thus, it takes time t to transport the part from A to B. During
the operation of an AGV, certain fatalrisks which may cause the failure of the
transportation (e.g., defect in automation) arrive at the vehicle according to
ahomogeneous Poisson process with intensity λ . A fatalrisk, independently of all
others, causes the failure of the transportation with probability p. When more than
one vehicle operates at the same time, given afatalrisk, the failures between vehicles
are assumed to be independent. We have m kinds of parts (part1, part2, . . . , partm)
that should be transported and m identical AGVs (AGV1, AGV2, . . . , AGVn). We
define that the ‘mission’ is completed when all the m kinds of parts are safely
transported. A configuration of the proposed manufacturing network is depicted in
Figure 1.
Fig. 1 A configuration of the proposed manufacturing network
As shown in Figure 2, several AGVs carry parts amongst the work stations on
the guide paths to process the manufacturing plan and satisfy product demands.
Turning points are mounted as guide paths division centers to prohibit AGVs’
conflicts during movements. The intersection of guide paths is determined as the
turning point. In these points an AGV is directed according to the process plan
sent from the control unit and with respect to work stations’ demands. These
points are the state variables for the proposed network to be handled by a dynamic
program.
As a result, for an economic decision making, loss of the incurred risks can be
considered. The loss here is composed of the risk of defect when an AGV moves
in the manufacturing network. A loss concept is utilized to consider all aspects of
a reliable manufacturing network.
We consider loss as a function of defect elements, that is,
Loss ≡ f ( X 1 ,...X p ) , where X1,…,Xp are the elements of defects, i.e., part m
transported by AGV n. Clearly, defect data for AGVs are recovered in any time
periods causing uncertainty of data. Therefore, the proposed loss function should
be estimated. For loss function estimation, a Bayesian approach is utilized. To find
Fig. 2 A flowchart of the proposed methodology

an optimal path in the manufacturing network having less risk, and therefore less loss,
dynamic programming is applied. The reason is that the proposed network is
composed of nodes as state variable and part processing in different work stations as
stages of the dynamic program. A flowchart of the proposed methodology is shown in
Figure 2.

The model applied in the present work belongs to the applied measurement
approach (AMA) and it is called the loss distribution approach (LDA). It is
characterized by the categorization of the losses in terms of ‘Frequency’ (the
number of loss events during a certain time period) and ‘Severity’ (the impact of
the event in terms of economic loss) (Valle and Giudici, 2006). Formally, for each
intersection r (where r = 1, ... , R) and for a given number of AGV t, the total
operational loss could be defined as the sum of a random number ntof losses:
Lrt= Xr1+ Xr2+· · ·+Xr nt , (1)
where Lrt denotes the total operational loss, Xr1, ... , Xr nt denote individual loss
severities and ntdenotes the frequency, for t =1, ... , T, T representing the number
of AGVs. Note that, for each intersection and for each AGV, the total loss can be
expressed as
Lt= st· nt, (2)
where ntis the frequency, defined as before, and st(commonly referred to as the
severity) is the mean loss for that AGV. The LDA assumes that, for each AGV:
(1) the individual losses {Xrq}, where q = 1, ... , nt, are independent and
identically distributed random variables;
(2) the distribution of the frequency ntis independent of the distribution of the
severities {Xrq}, for q = 1,… , nt. This implies that ntis independent of st;
(3) besides, Lrt, for t = 1, ... ,T, are independent and identically distributed
random variables.
For a given intersection, we construct a discrete probability density of the
number of loss events ntfor the AGV and ntcontinuous probability densities of the
loss severities (Alexander, 2003). Now, if we express the likelihood function for
each intersection in a general way, we obtain the following equation, indicating
( )
the severity distribution with f x q ξ and the frequency distribution with
f (nt ψ ) , where ξ denotes the parameter vector of the severity distribution and
ψ denotes the parameter vector of the frequency distribution, we obtain the
following form:
T ⎡ nt ⎤
L(x, n ψ , ξ ) = ∏ ⎢∏ f ( x j ξ ⎥ f (nt ψ ) . (3)
t =1 ⎣ q =1 ⎦
Different functional forms for the frequency and severity distributions exist.
Frequency represents a discrete phenomenon. Since we want to determine the
probability that a certain number of loss events occur in a predetermined time
horizon, the most suitable probability distributions are the Poisson and the
Negative Binomial. Severity, instead, is a continuous phenomenon and we chose
to describe it by the Exponential, the Gamma and the Pareto distributions. For
more details about these distributions see Johnson et al. (1994).
In the proposed problem the number of defects occurs for an AGV is pointed
by frequency and the effect of loss is implied by severity. Our first problem is to
estimate, on the basis of data, the parameters of the frequency and severity
distributions, denoted by ψ and ξ , respectively, in equation (3). The classical
approach suggests the employment of the method of moments or the method of
maximum likelihood, as described, for example, by Gourieroux and Monfort
(1995). An alternative approach is the Bayesian method, which allows the
combination of quantitative data, coming from the time series of operational
losses, and prior information, represented by expert opinions. See Berger (1985),
Bernardo and Smith (1994) or Robert (1994) for an introduction to these methods.
In this chapter we estimate the parameters with the maximum likelihood method
and with the Bayesian method, making some comparisons between the classical
and the Bayesian approaches. In particular, we choose the maximum likelihood
method because of its good asymptotic properties, since the MLE converges
almost surely to the true value of the parameter, under fairly general conditions
(see Lehmann and Casella, 1998). Once the parameters have been estimated, the
marginals are defined and the operational loss distribution has to be identified.
Bayesian Approach
Suppose a continuous probability distribution with probability density function
(pdf) ƒΘ is assigned to an uncertain quantity Θ. In the conventional language of
mathematical probability theory Θ would be a "random variable" (Lehmann and
Casella, 1998). The probability that the event B will be the outcome of an
experiment depends on Θ; it is P(B | Θ). As a function of Θ this is the likelihood
function:
L(θ ) = P(B Θ = θ ) . (4)
Then the posterior probability distribution of Θ, i.e. the conditional probability

distribution of Θ given the observed data B, has probability density function
f Θ (θ B) = constant.f Θ (θ ) L(B θ ) , (5)
where the "constant" is a normalizing constant so chosen as to make the integral

of the function equal to 1, so that it is indeed a probability density function.
This is the form of Bayes' theorem actually considered by Thomas Bayes (Berger,
1985). More generally still, the new data B may be the value of an observed
continuously distributed random variable X. The probability that it has any
particular value is therefore 0. In such a case, the likelihood function is the
value of a probability density function of X given Θ, rather than a probability of B
given Θ:
L(θ ) = f X (x Θ = θ ) . (6)
Bayes's rule provides the framework for combining prior information with
sample data. In this reference, we apply Bayes's rule for combining prior
information on the assumed distribution's parameter(s) θ with sample data in order
to make inferences based on the model. The prior knowledge about the
parameter(s) is expressed in terms of a pdf f(θ), called the prior distribution. The
posterior distribution of θ given the sample data, using Bayes rule, provides
the updated information about the parameters θ. This is expressed with the
following posterior pdf:
L(Data θ )ϕ (θ )
f (θ Data ) = . (7)
∫ L(Data θ )ϕ (θ )dθ
ε
where,
θ is a vector of the parameters of the chosen distribution,
ε is the range of θ,
L(Data|θ) is the likelihood function based on the chosen distribution and data,
f(θ) is the prior distribution for each of the parameters.
The integral in equation (7) is often referred to as the marginal probability and
can be interpreted as the probability of obtaining the sample data given a prior
distribution and it's a constant number. Generally, the integral in equation (7) does
not have a closed form solution and numerical methods are needed for its solution.
As can be seen from equation (7), there is a significant difference between
classical and Bayesian statistics. First, the idea of prior information does not exist
in classical statistics. All inferences in classical statistics are based on the sample
data. On the other hand, in the Bayesian framework, prior information constitutes
the basis of the theory. Another difference is in the overall approach of making
inferences and their interpretation. For example, in Bayesian analysis the
parameters of the distribution to be "fitted" are the random variables. In reality,
there is no distribution fitted to the data in the Bayesian case.
In order to obtain an estimate of the random variables, a probability needs to be

specified or we can use the expected value of the posterior distribution. In order to
demonstrate the procedure of obtaining results from the posterior distribution, we
will rewrite equation (7) for a single parameter θ1:
L(Data θ1 )ϕ (θ 1 )
f (θ1 Data ) = . (8)
∫ L(Dataθ1 )ϕ (θ1 )dθ1
ε
The expected value (or mean value) of the parameter θ1 can be obtained as
follows:
E (θ1 ) = ∫ θ1 . f (θ1 Data )dθ1 . (9)

ε
Now, using the estimated loss function, we obtain the arc values being applied
in dynamic programming with respect to state and stage concepts to find the
optimal path.
A Dynamic Program
The output of the current state is F (w1,...,wl ) , being the obtained loss from the
Bayesian method. While the current state is an n-dimensional vector having
several nodes (work stations), we consider a state function δ l ( p), l = 1,...,n .
Also, a policy function is proposed to consider the AGV movement policy, and
hence assuming an m-dimensional vector of policies ( m ≤ n ), we have
β l ( p), l = 1,...,m , where βl ( p) ∈{0,1} , i.e., policy movement occurs (=1)
or does not occur (=0) for an AGV movement. Here, our problem becomes n
dynamical equations of the form
δ l ( p + 1) = δ l ( p ) + f l , p (δ 1 ( p ),..., δ n ( p ), β1 ( p ),..., β m ( p ) ), l = 1,..., n, p = 0,..., N − 1. (10)
Therefore, a dynamic programming approach can be used to model our multi-

dimensional problem. Let S p (δ 1 ( p),...,δ n ( p) ) bethe minimum loss of the
remaining nodes an AGV visits if we start stage p in state δ1 ( p),...,δ n ( p) .
Then, for p = 0,..., N − 1 , the strategy function is:
⎡ g p (δ 1 ( p ),..., δ n ( p ), β 1 ( p ),..., β m ( p ) ) + ⎤
⎢ ⎥ .(11)
S p (δ 1 ( p ),..., δ n ( p ) ) = min ⎢ ⎛ δ ( p ), f (δ ( p ),..., δ ( p ), β ( p ),..., β ( p ) ),..., δ ( p ) + ⎞ ⎥
β1 ( p ),..., β m ( p ) S ⎜ 1 1, p 1 n 1 m n
⎟
⎢ p +1 ⎜ f (δ ( p ),..., δ ( p ), β ( p ),..., β ( p ) ) ⎟⎥
⎣ ⎝ n, p 1 n 1 m ⎠⎦
with the boundary condition,

S N (δ1 ( N ),...,δ n ( N )) = h(δ1 ( N ),...,δ n ( N )) . (12)
This way, the state and the policy direction for the AGV is at hand. Solving
(11) for the appropriate indices, we obtain the state and the policy direction in
each time period.
Considering the boundary condition, we can compute hOPT (the optimum states,
i.e., the optimal arcs leading to minimum loss of AGV movement) for given
policy functions. Due to nonlinear format and operations on mathematical
functions, it is easy to see that exact dynamic program (DP) requires O(n( δ * )m))
calculations for the multi-dimensional dynamic program where
δ * = min{δ1,..., δl } . Space requirements and value function computations
become impractical for even moderate m. Thus, exact DP is not a practical
methodology particularly for large-scale optimization problems. The basic idea
behind approximate DP is to approximate the optimal value function
S p (δ1 ( p),..., δn ( p)) and to construct a suboptimal solution with respect to
h (δ1 ( p),..., δn ( p)) . We introduce a heuristic algorithm to approximate the

optimal value function in the following sections.
A Heuristic Solution Approach

This section introduces our proposed Adaptive Dynamic Program Heuristic
(ADPH) approach for the multi-dimensional problem. The basic idea of ADPH is
estimating the optimal value function by the solution value of a suboptimal
methodology to the corresponding subproblem and constructing a solution through
boundary condition equation. This suboptimal methodology is called the heuristic
approach (Bertsimasand Demir, 2002).
Let h (δl ( p) ) be a heuristic for the subproblem h (δ1 ( p),..., δn ( p)) . Let
δl ( p) be the corresponding heuristic solution and S p (δl ( p)) be the heuristic
value, i.e., an estimate of the optimal value S p (δ1( p),...,δn ( p)) . The ADPH
algorithm starts by applying h (δl ( p) ) to the problem h (δ1 ( p),..., δn ( p))
and getting δ l ( p) . If the solution δl ( p) is optimal, the algorithm terminates
with an optimal solution. If the problem is infeasible, the algorithm terminates
without a solution. Otherwise, the algorithm sets the variables best-solution
δ BEST = δ l ( p) and best-solution-value S p = δ l ( p) (Step 2 in Figure 4).
The algorithm proceeds by applying reduced loss fixing as described below to

fix some the β l ( p), l = 1,...,m variables to the corresponding values (0 or 1) in
an optimal solution to the problem h (δ1 ( p),..., δn ( p)) _ (Step 3 in Figure 4).
Reduced loss fixing might effectively reduce the number of variables in advance.
Let M denote the set of indices of the fixable variables by the reduced loss
criterion. We denote by xM the corresponding fixed values, that is,
xlM = 0 or 1 , for all l∈M .
The algorithm iteratively sets the variables to 0 or 1 as described in Step 4 of
Figure 4. At each iteration, we update the best-solution, δ BEST = δ l ( p) and best-
solutionvalue, S p = δ l ( p) . We check if early termination is possible (true) or

not (false). We also allow fixing a set of variables, which we call lag-variable-
fixing. At the final step, we set the first variable to the optimal solution of the
reduced problem (Step 5 of Figure 4) and the algorithm returns δ BEST = δ l ( p)
and S p = δ l ( p) . We present the details of ADPH algorithm in Figure 3.
Variable Assignment: To calculate S p for a variable l ∉ M , we employ an

estimated value, M (l, p) , instead of optimal ones, M (l, p) . This is the basic
idea in ADPH. We denote by U(l,p) an upper bound to the problem h (δl ( p) ) . In
our computations, we set the upper bound U(l,p) to h OPT
. Let ε denote the
percentage deviation of δ (l −1, p + 1(βl ( p))) with respect to the upper bound
U(l-1,(p+1) βl ( p) ) for βl ( p) =0, 1 and also let ε * = min {ε 0 , ε1} . We
estimate the optimal values, M(l−1,(p+1) βl ( p) ), by
M (l − 1, ( p + 1)βl ( p)) = (1 − ε *)U (l − 1, ( p + 1) βl ( p)) for βl ( p) =0, 1.

By the definition of ε * , the optimal values are approximated such that the
percentage deviations of the estimate values are the same with respect to
the associated upper bounds. Finally, we set Sp replacing the optimal values by the
estimated ones.
1: Initialization:
l = n, p =0, Sp=0
early−termination = false
2: Apply h ( δ l ( p) ) and get h (δl ( p) ) , S p = δ l ( p)

Problem infeasible → exit, the problem is infeasible
δ l ( p) optimal → exit, an optimal solution available
δ ← δ l ( p) , S p ← δ l ( p)
BEST
3: Apply reduced loss fixing and get M, xM

4: While (l > 2)and (early−termination = false)do
if (l ∈M)then
M
Sp← xl
else
Calculate Sp by calling variable assignment
Update best-solution and best-solution-value and get δ BEST , S p
Check early termination criterion and get the status (true or false)
Fix variables by lag variable fixing and get M, SM
p←p+1
l←l−1
5: Set Sp to the optimal solution of h (δl ( p) )
6: Output: δ BEST and Sp
Fig. 3 The ADPH algorithm for the h (δ1 ( p),..., δ n ( p) )
Update Best-Known Solution and Value: Once we set Sp, if necessary, we

update the best-solution, δ BEST , and best-solution-value, Sp as follows. Let
n
p=p+1( ∑S j =l
j
p ). Because of the construction of Spjfor j in [l,p],
δ c = (δl ( p)(l −1, p),S pl ,...,S pn ) is a feasible solution to the problem h (δl ( p) ) ,
n
where the value of the solution δ c equals S pc = h(l − 1, p)∑ S pj . We update
j =l
δ BEST to
c c
δ c and Sp to S p , respectively, if S p is larger than Sp.
Early Termination: The ADPH algorithm can be terminated early while setting
Sp if we determine that we find an optimal solution to the problem h (δl ( p) ) ,
n
where p=p+1( ∑S
j =l
j
p ). We apply the following tests to determine whether we
have an optimal solution or not to the problem h (δl ( p) ) :

(a) We check if δ (l −1, p) and δ (l − 1, p + S p ) are optimal solutions to
the subproblems h (δ (l −1, p) ) and h (δ (l − 1, p + S p ) ) ,
respectively.
We conclude that an optimal solution exists if h (δ (l −1, p) ) = U(l-1, p) and
h (δ (l − 1, p + S p ) ) = U(l-1,p+Sp);
(b) We conclude that δ (l, p) = (δ (l −1, p),0) is an optimal solution if
h (δ (l −1, p) ) is greater than U(l-1,p+Sp) and h (δ (l −1, p) ) = U(l-1,

p);
(c) Similarly, we conclude that δ (l , p ) = (δ (l − 1, p + S p ),1) is an optimal
solution if h (δ (l − 1, p + S p ) ) is greater than U(l-1,p) and
h (δ (l − 1, p + S p ) ) = U(l-1,p+Sp).
Lag Variable Fixing: To allow the ADPH to generate solutions in significantly

shorter times (possibly of lower quality), we allow fixing a set of variables Spjfor j
in [l-n, l-1], and j ∉M . We call the method lag-variable-fixing and those
variables as lag-fixable variables. We denote by
M lag,l
= { j : j ∈[l − n, l −1 and j ∈ M } . Once the lth variable is assigned to
S Mj are set to S p (l −1, p) or S p (l −1, p + S p ) ,
j j
0 or 1, the variables
respectively, for all j ∈ M lag,l . We update the set M by including those indices
M lag,l . Spj are assigned to SMj for all j ∈ M
lag,l
in in Step 4 of Figure 4. Lag-size
parameter q specifies the number of variables to be fixed lagging from the lth
variable while setting Spl through the relationship, q=[l/lag-time], where lag-time
is a user-specified parameter.
12.4 Conclusions
We proposed a new automated manufacturing network considering the existing

data on states and stages aiming to find the arcs with more convergence for value
added purposes. The arcs were the guide paths that AGVs move through to carry
parts. Regarding the multi stage decision making process of the AGVs, we
develop a multi-dimensional dynamic program as a useful tool for multi stage
decision making. To counteract the dynamism of the digital data in different time
periods, two concepts of state and policy direction are introduced to determine the
loss of moving through the stages of the proposed manufacturing network. Using
this function, a dynamic model was proposed to consider the AGV's movement
state and policy direction in each time period. A dynamic program was proposed
to minimize loss of the remaining nodes assuming that the analysis started at an
appropriate stage of the problem. Since the space requirements and value function
computations become impractical for even moderate size, we approximated the
optimal value function developing a heuristic algorithm. A comprehensive
heuristic solution methodology was developed to approximate the proposed
dynamic program due to nonlinearity and computational difficulties. The solution
methodology was verified in comparison with some famous methods in the
literature. In summary, the computational evidence suggests that the ADPH
( )
approach for the h δl ( p) seems an attractive alternative to existing
methodologies as it produces near optimal solutions fast.
The advantages of the proposed methodology are:
 Controlling uncertainty and dynamism of future analysis for AGV

movement
 Breakthrough the traditional obstacles of loss function estimation by
Bayesian analysis using prior and posterior distributions
 Usefulness of the collected data (past data) for future analysis of AGV
risk assessment
 Maximizing the expected benefit form an AGV movement process
 Competitive advantage of the management equipped with the proposed
methodology
 Reducing the risk exist in the AGV based transportation
To sum up, the proposed methodology with the above capabilities is a
significant tool for the automated manufacturing network since the current market
is very competitive and absorbing large amount of markets requires subtle analysis
and decision making. Also, the obtained results show the more reliablepaths for
AGV movements. This is useful for redesigning a routing structure for
optimization of extension of shop purposes. Another element of the results
obtained is the conflict free notion in the AGV movement on guide paths being
very important in designing an AGV system. Clearly, due to dynamism the model
is capable to determine low risk conflict free AGV routings.
References 203
References
Akao, Y.: QFD: Quality Function Deployment—Integrating Customer Requirements into
Product Design. Productivity Press (2004)
Alexander, C. (ed.): Operational Risk. Regulation, Analysis and Management. Financial
Times Prentice Hall, London (2003)
Berger, J.O.: Statistical decision theory and Bayesian analysis, 2nd edn. Springer, NewYork
(1985)
Bernardo, J.M., Smith, A.F.M.: Bayesian theory. Wiley, Chichester (1994)
Bertsimas, D., Demir, R.: An Approximate Dynamic Programming Approach to
Multidimensional Knapsack Problems. Management Science 48(4), 550–565 (2002)
Diekmann, J.E., Sewester, E.E., Taher, K.: Risk Managesment in Capital Projects.
Construction Industry Institute, Austin (1988)
Gourieroux, C., Monfort, A.: Statistics and Econometric Models. Cambridge University
Press, Cambridge (1995)
Hetland, P.W.: Uncertainty Management. In: Smith, N.J. (ed.) Appraisal, Risk and
Uncertainty, ch. 8, pp. 59–88. Thomas Telford, London (2003)
Johnson, N.L., Kotz, S., Balakrishnan, N.: Continuous Univariate Distributions, 2nd edn.,
vol. 1. Wiley, NewYork (1994)
Knight, F.H.: Risk, Uncertainty and Profit. Houghton Mifflin Company, Boston (1921)
Kochenberger, G.A., McCarl, B.A., Wymann, F.P.: A heuristic for general integer
programming. Decision Sci. 5, 36–44 (1974)
Lehmann, E.L., Casella, G.: Theory of Point Estimation, 2nd edn. Springer (1998) ISBN 0-
387-98502-6
Loulou, R., Michaelides, E.: New greedy-like heuristics for the multidimensional 0-1
knapsack problem. Oper. Res. 27, 1101–1114 (1979)
Manuj, I., Mentzer, J.: Global supply chain risk management. Journal of Business
Logistics 29(1), 133–155 (2008)
Nau, B., Roderburg, A., Klocke, F.: Ramp-up of hybrid manufacturing technologies. CIRP
Journal of Manufacturing Science and Technology 3, 313–316 (2011)
Nau, B., Roderburg, A., Klocke, F., Park, H.S.: Risk assessment of hybrid manufacturing
technologies for ramp-up projects. CIRP Journal of Manufacturing Science and
Technology 5, 228–234 (2012)
Pons, D.J.: Strategic Risk Management: Application to Manufacturing. The Open Industrial
& Manufacturing Engineering Journal 3, 13–29 (2010)
Senju, S., Toyoda, Y.: An approach to linear programming with 0-1 linear variables.
Management Sci. 15, 196–207 (1968)
Toyoda, Y.: A simplified algorithm for obtaining approximate solutions to zero-one
programming problems. Management Sci. 21, 1417–1427 (1975)
Valle, L.D., Giudici, P.: A Bayesian approach to estimate the marginal loss distributions in
operational risk management. Computational Statistics & Data Analysis 52, 3107–3127
(2006)
Vilko, J.P.P., Hallikas, J.M.: Risk assessment in multimodal supply chains. International
Journal of Production Economics 140(2), 586–595 (2012)
Waters, D.: Supply chain risk management: vulnerability and resilience in logistics. Kogan
Page Limited (2007) ISBN-13: 978-0-7494-4854-7

Autonomous Guided Vehicles: Hamed Fazlollahtabar Mohammad Saidi-Mehrabad

Uploaded by

Copyright:

Available Formats

Autonomous Guided Vehicles: Hamed Fazlollahtabar Mohammad Saidi-Mehrabad

Uploaded by

Document Information

Original Title

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

Autonomous Guided Vehicles: Hamed Fazlollahtabar Mohammad Saidi-Mehrabad

Uploaded by

Copyright:

Available Formats

What is the topic of the document?

What is the topic of the document?

What areas does the series cover?

What areas does the series cover?

Studies in Systems, Decision and Control 20

More information about this series at http://www.springer.com/series/13304

Autonomous Guided Vehicles

ISSN 2198-4182 ISSN 2198-4190 (electronic)

Springer Cham Heidelberg New York Dordrecht London

Printed on acid-free paper

Springer International Publishing AG Switzerland is part of Springer Science+Business Media

unmanned vehicle that is capable of transporting goods in an industrial

January 2015 Hamed Fazlollahtabar

1 Models for AGVs’ Scheduling and Routing ......................................... 1

2 Analytical Material Flow Model for AGV System .............................. 17

3 Nonlinear Stochastic Model for AGV System ...................................... 27

4 Reliability Model for AGV .................................................................... 41

5 Uncertain Optimal Path for AGV System ............................................ 57

5.3 Mathematical Modeling .................................................................. 60

6 Cross Entropy Model for AGV Routing Time ..................................... 79

7 Neuro-Fuzzy-Regression Expert System for AGV Optimal Path ...... 93

9 Earliness/Tardiness for a Multiple AGV System ................................. 133

10 Markovian Model for Multiple AGV System ...................................... 147

11 Producer's Behavior Analysis for AGV System .................................. 169

12 Risk for Multiple AGV System ............................................................. 189

An automated guided vehicle (AGV) is a driverless material handling system used

© Springer International Publishing Switzerland 2015 1

1.2 Scheduling Models

The literature discussed so far on scheduling of AGVs hardly considers (side)

1.2.1 Artificial Intelligence

optimization techniques. While these methods can give an optimal solution to

1.3 Routing Models

partitioning based heuristic. In a systematic way fragments of routes are combined

1.3.1 Artificial Intelligent

multi-line FMS. The author examined the impact of guide-path flexibility on

A TAGV system is obtained by partitioning all workstations into multiple

Nowadays, so as to adapt to the global market, where competition is getting

Summarizing, in the context of manufacturing areas, static and dynamic

Manufacturing automation has become increasingly important as the need to

© Springer International Publishing Switzerland 2015 17

and has become more widely used in modern manufacturing environments

2.2 Statement of the Problem

Fig. 1 The Schematic presentation of the Proposed Model

2.3 Mathematical Model

m Index for shops, m=1,2, … , M.

C.Ti Cycle Time for product ith

Cipmn − Cip −1kh ≥ P.Timn + Wimn + T .Timn kh , ∀i, p, m, n, k , h, (3)

z ipmn .Cipmn ≤ C.Ti , ∀i, p, m, n, (5)

z ipmn ∈ {0,1} (8)

Equation (1), is the objective function of the proposed problem which

2.3.1 Regression Model for Performance Analysis

Dt = β 0 + β1 X 1 + ... + β j X j + ε j , t = 1,...,T ; j = 1,..., J . (9)

sampling distribution of βj.

Although the formula for σ β2 j

assuming that the null hypothesis is correct. Stated mathematically,

equality follows by dividing by SE ( β j ) , and t act is the value of the t-statistic

p − value = Pr( Z > t act ) = 2Φ (− t act ) , (15)

Advanced automated manufacturing systems are widely used in industrial

© Springer International Publishing Switzerland 2015 27