Introduction to Operations Research

Matthew Galati
magh@lehigh.edu

Department of Industrial and Systems Engineering, Lehigh University Service Parts Solutions, IBM Corporation

Introduction to Operations Research p.1

Outline
What is Operations Research? Optimization Problems and Applications Personal Examples

Introduction to Operations Research p.2

What is Operations Reseach?

Operations Research (OR) started just before World War II in Britain with the establishment of teams of scientists to study the strategic and tactical problems involved in military operations. The objective was to nd the most effective utilization of limited military resources by the use of quantitative techniques. Following the war, numerous peacetime applications emerged, leading to the use of OR and management science in many industries and occupations. Occupational Outlook Handbook - http://stats.bls.gov/oco/ocos044.htm.

Introduction to Operations Research p.3

Denitions
Model: A schematic description of a system,theory, or phenomenon that accounts for its known or inferred properties and may be used for further study of its characteristics. System: A functionally related group of elements, especially:
The human body regarded as a functional physiological unit. An organism as a whole, especially with regard to its vital processes or functions. A group of physiologically or anatomically complementary organs or parts: the nervous system; the skeletal system. A group of interacting mechanical or electrical components. A network of structures and channels, as for communication, travel, or distribution. A network of related computer software, hardware, and data transmission devices.

Introduction to Operations Research p.4

Denitions
Operations Research (OR) is the study of mathematical models for complex organizational systems. Optimization is a branch of OR which uses mathematical techniques such as linear and nonlinear programming to derive values for system variables that will optimize performance.

Introduction to Operations Research p.5

Models
Linear Programming
Typically, a single objective function, representing either a prot to be maximized or a cost to be minimized, and a set of constraints that circumscribe the decision variables. The objective function and constraints all are linear functions of the decision variables. Software has been developed that is capable of solving problems containing millions of variables and tens of thousands of constraints.

Network Flow Programming

A special case of the more general linear program. Includes such problems as the transportation problem, the assignment problem, the shortest path problem, the maximum ow problem, and the minimum cost ow problem. Very efcient algorithms exist which are many times more efcient than linear programming in the utilization of computer time and space resources.

Introduction to Operations Research p.6

Models
Integer Programming
Some of the variables are required to take on discrete values. NP-hard: Most problems of practical size are very difcult or impossible to solve.

Nonlinear Programming
The objective and/or any constraint is nonlinear. In general, much more difcult to solve than linear. Most (if not all) real world applications require a nonlinear model. In order to be make the problems tractable, we often approximate using linear functions.

Introduction to Operations Research p.7

Models
Dynamic Programming
A DP model describes a process in terms of states, decisions, transitions and returns. The process begins in some initial state where a decision is made. The decision causes a transition to a new state. Based on the starting state, ending state and decision a return is realized. The process continues through a sequence of states until nally a nal state is reached. The problem is to nd the sequence that maximizes the total return. Objectives with very general functional forms may be handled and a global optimal solution is always obtained. "Curse of dimensionality" - the number of states grows exponentially with the number of dimensions of the problem.

Introduction to Operations Research p.8

Models
Stochastic Processes
In many practical situations the attributes of a system randomly change over time. Examples include the number of customers in a checkout line, congestion on a highway, the number of items in a warehouse, and the price of a nancial security to name a few. The model is described in part by enumerating the states in which the system can be found. The state is like a snapshot of the system at a point in time that describes the attributes of the system. Events occur that change the state of the system. Consider an Automated Teller Machine (ATM) system. The state is the number of customers at or waiting for the machine. Time is the linear measure through which the system moves. Events are arrivals and departures.

Introduction to Operations Research p.9

Models
Markov Chains
A stochastic process that can be observed at regular intervals such as every day or every week can be described by a matrix which gives the probabilities of moving to each state from every other state in one time interval. Assuming this matrix is unchanging with time, the process is called a Markov Chain. Computational techniques are available to compute a variety of system measures that can be used to analyze and evaluate a Markov Chain model.

Markov Processes
A continuous time stochastic process in which the duration of all state changing activities are exponentially distributed. Time is a continous parameter. The process is entirely described by a matrix showing the rate of transition from each state to every other state. The rates are the parameters of the associated exponential distributions. The analytical results are very similar to those of a Markov Chain.

Introduction to Operations Research p.10

Models
Simulation
It is often difcult to obtain a closed form expression for the behavior of a stochastic system. Simulation is a very general technique for estimating statistical measures of complex systems. A system is modeled as if the random variables were known. Then values for the variables are drawn randomly from their known probability distributions. Each replication gives one observation of the system response. By simulating a system in this fashion for many replications and recording the responses, one can compute statistics concerning the results. The statistics are used for evaluation and design.

Introduction to Operations Research p.11

Models
Time Series and Forecasting
A time series is a sequence of observations of a periodic random variable. Typically serve as input to OR decision models. Example - inventory model requires estimates of future demands. Example - a course scheduling and stafng model for the university department requires estimates of future student inow. Example - A model for providing warnings to the population in a river basin requires estimates of river ows for the immediate future.

Introduction to Operations Research p.12

Models
Inventory Theory
Inventories are materials stored, waiting for processing, or experiencing processing. When and how much raw material should be ordered? When should a production order should be released to the plant? What level of safety stock should be maintained at a retail outlet? How is in-process inventory maintained in a production process?

Reliability Theory
Attempts to assign numbers to the propensity of systems to fail. Estimating reliability is essentially a problem in probability modeling. Extremely important in the telecommunications and networking industry.

Introduction to Operations Research p.13

Mathematical Programming
A mathematical model consists of:
Decision Variables, Constraints, Objective Function, Parameters and Data

The general form of a math programming model is: min or max f (x1 , . . . , xn ) s.t. gi (x1 , . . . , xn ) xX Linear program (LP): all functions f and gi are linear and X is continuous. Integer program (IP): X is discrete. = bi

Introduction to Operations Research p.14

Mathematical Programming
A solution is an assignment of values to variables. A feasible solution is an assignment of values to variables such that all the constraints are satised. The objective function value of a solution is obtained by evaluating the objective function at the given solution. An optimal solution (assuming minimization) is one whose objective function value is less than or equal to that of all other feasible solutions.

Introduction to Operations Research p.15

Outline
What is Operations Research? Optimization Problems and Applications Personal Examples

Introduction to Operations Research p.16

LP Example: Two Crude Petroleum

Two Crude Petroleum distills crude from two sources:
Saudi Arabia,Venezuela

They have three main products:

Gasoline,Jet Fuel,Lubricants

Yields
Gasoline Saudi Arabia Venezuela Availability and cost Availability Saudi Arabia Venezuela 9000 barrels 6000 barrels Cost $20/barrel $15/barrel 0.3 barrels 0.4 barrels Jet Fuel 0.4 barrels 0.2 barrels Lubricants 0.2 barrels 0.3 barrels

Introduction to Operations Research p.17

LP Example: Two Crude Petroleum

Production Requirements (per day) Gasoline 2000 barrels Jet fuel 1500 barrels Lubricants 500 barrels

Objective: Minimize production cost. This yields the following LP formulation: min 20x1 + 15x2

s.t. 0.3x1 + 0.4x2 2.0 0.4x1 + 0.2x2 1.5 0.2x1 + 0.3x2 0.5 0 x1 9 0 x2 6 Graphical Demo

Introduction to Operations Research p.18

Simplex Method
The simplex method generates a sequence of feasible iterates by repeatedly moving from one vertex of the feasible set to an adjacent vertex with a lower value of the objective function . When it is not possible to nd an adjoining vertex with a lower value, the current vertex must be optimal, and termination occurs. Exponential (worst case) run time; in practice, runs very fast.

Introduction to Operations Research p.19

Optimization - Linear Programming

Simplex (Dantzig 1947) Ellipsoid (Khachian 1979) - the "rst" polynomial-time algorithm Interior Point - the "rst" practical polynomial-time algorithm Projective Method (Karmarkar 1984) Afne Method (Dikin 1967) Logarithmic Barrier Method (Frisch 1955, Fiacco 1968, Gukk et. al 1986)

Introduction to Operations Research p.20

Integer Programming
Let X be a discrete set (integers).

Convex hull of integer solutions Linear programming relaxation

Introduction to Operations Research p.21

Integer Programming
Why not just solve the LP and round? max 1.00x1 + 0.64x2 s.t. 50x1 + 31x2 250 3x1 2x2 4 x1 0 x2 0

Introduction to Operations Research p.22

LP-based Branch and Bound

Consider problem P : min cT x s.t. Ax b xi Z i I where (A, b) Rmn+1 , c Rn . Let P = conv{x Rn : Ax b, xi Z i I}. Basic Algorithmic Approach Use LP relaxations to produce lower bounds. Branch using hyperplanes. Basic Algorithmic Elements A method for producing and tightening the LP relaxations. A method for branching.

Introduction to Operations Research p.23

Branch, Cut, and Price

Weyl-Minkowski b) (A, Rmn+1 s.t. P = {x Rn : Ax b} We want the solution to min{cT x : Ax b}. Solving this LP isnt practical (or necessary).

BCP Approach Form LP relaxations using submatrices of A. The submatrices are dened by sets V [1..n] and C [1..m]. Forming/managing these relaxations efciently is one of the primary challenge of BCP.

Introduction to Operations Research p.24

Cutting Plane Method

Basic cutting plane algorithm
Relax the integrality constraints. Solve the relaxation. Infeasible STOP. If x integral STOP. Separate x from P. No cutting planes algorithm fails.

The key is good separation algorithms.

Introduction to Operations Research p.25

Branch and Cut Methods

If the cutting plane approach fails, then we divide and conquer (branch).

Introduction to Operations Research p.26

The Challenge of BCP

The efciency of BCP depends heavily on the size (number of rows and columns) and tightness of the LP relaxations. Tradeoff Small LP relaxations faster LP solution. Big LP relaxations better bounds. The goal is to keep relaxations small while not sacricing bound quality. We must be able to easily move constraints and variables in and out of the active set. This means dynamic generation and deletion.

Introduction to Operations Research p.27

Outline
What is Operations Research? Optimization Problems and Applications Personal Examples

Introduction to Operations Research p.28

Common Problems in OR
Set Covering, Packing, Partitioning
Let S be a set of objects and a set of subsets of S. Let aij = 1, if i j and dene variable xj = 1, if the j th member of is used.

Set Covering Set Packing Set Partitioning

: : :

x{0,1}

min {
i

ci xi :
j

aij xj 1, i S} aij xj 1, i S}
j

x{0,1}

max {
i

ci xi :

x{0,1}

min

or max {
x{0,1} i

ci xi :
j

aij xj = 1, i S}

Air Crew Scheduling (Covering): Consider S to be a set of "legs" that the airline has to cover and the members of are possible rotations involving particular combinations of ights.

Introduction to Operations Research p.29

Common Problems in OR
Quadratic Assignment Problem
Given two sets of objects S and T where |S| = |T |, each member of S must be assigned to exactly one member of T , and vice versa. Costs are incurred for an assignment of i S to j T and k S to l T .
x{0,1}

min {
i,kS j,lT k>i

tik djl xij xkl :

jT

xij = 1, i S,
iS

xij = 1, j T }

Facility Location: Consider S to be a set of n factories and T to be a set of n cities. Locate one factory in each city and minimize the total communication cost between factories. Interpret tik as the frequency of communication between factories i and k and djl as the cost per unit of communication between cities j and l. Circuit Design: Consider S to be a set of n electronic modules and T to be a set of n predetermined positions on a backplate. The modules have to be connected to each other by a series of wires. Interpret tik as the number of wires which must connect module i to the module k and djl as the distance between position j and position l on the backplate.

Introduction to Operations Research p.30

Common Problems in OR
Routing problems. Finding a path or cycle in a network. An easy routing problem is the shortest path; a hard one is the travelling salesman problem. One prevalent class, with many variations, is vehicle routing.
Shortest path. In a graph or network, this is a path from one node to another whose total cost is the least among all such paths. The "cost" is usually the sum of the arc costs, but it could be another function (e.g., the product for a reliability problem, or max for a fuzzy measure of risk). Vehicle routing problem (VRP). Find optimal delivery routes from one or more depots to a set of geographically scattered points (e.g., population centers). In its most complex form, the VRP is a generalization of the TSP, as it can include additional time and capacity constraints, precedence constraints, plus more.

Introduction to Operations Research p.31

Common Problems in OR
Production Scheduling Problem
To determine levels of production over time. Constraints include demand requirements (possibly with backordering), capacity limits (including warehouse space for inventory), and resource limits. Dene xt = level of production in period t (before demand); yt = level of inventory at the end of period t; Ut = production capacity in period t; Wt = warehouse capacity in period t; ht = holding cost (per unit of inventory); pt = production cost (per unit of production); Dt = demand at the end of period t. min{
x,y

px + hy : yt+1 = yt + xt Dt , t, 0 x U, 0 y W }

Introduction to Operations Research p.32

Common Problems in OR
Portfolio Selection Problem.
The objective is to minimize the variance on returns. Let xj be the percent of capital invested in the j th opportunity (e.g., stock or bond), so x must satisfy x 0 and j xj = 1. Let vj be the expected return per unit of investment in the j th opportunity, so that vx is the sum total rate of return per unit of capital invested. It is required to have a lower limit on this rate: vx b (where b is between min vj and max vj ). Subject to this rate of return constraint, the objective is the quadratic form, xT Qx, where Q is the variance-covariance matrix associated with the investments (i.e., if the actual return rate is V j , then Qi j = E[(Vi vi )(Vj vj )]

Chemical Equilibrium Problem.

The problem is to minx cx + j xj log xj : x > 0, j xj = M, Ax = b, where the objective is the Gibbs free energy function for xj = number of moles of species j, bi = atomic weight of atom of type i, and aij = number of atoms of type i in one mole of species j. The equation, Ax = b, is mass balance.

Introduction to Operations Research p.33

Outline
What is Operations Research? Optimization Problems and Applications Personal Examples

Introduction to Operations Research p.34

TRL Trans Refrigerated Lines

Dynamic Load Assignment Problem (Trucks to Loads) Parametric Multi-Criteria Objective
Early/Late Delivery, Early/Late Pickup Reduction in Empty Travel, Vehicle Maintenance Schedule Driver Vacation, Driver Hours Balancing (Union Labor)

GUI Interface for Load Planners (Real Time Dispatch) Lesson in Industry - Systems Development Paradox
In order to create the system correctly we needed the knowledge base of the load planners. Load planners are well aware that system would serve as their replacement. Upper management was willing to fund the development due to the potential cost reduction.

Introduction to Operations Research p.35

IBM SPS - Industry Standard

Rigid Structured Echelon Stocking/Order Routing Model
Stocking/order routing predetermined via pass-up schemes.

Service Delivery Performance Based on Parts Availability

No time/distance component linked to "Point-of-Demand"

Stock Network Locations Independently

Request for Emergency Order

Not Allowed ! (formally)

85%

75%

95%

90%

Customer Base

Introduction to Operations Research p.36

IBM SPS - Neighborhood

Regional Contained Stocking/Ordering
Neighborhoods are created around the time-based requirements of customers. All locations may serve as emergency stocking points. Stocking/order routing based on "pooled risk". Pass-up converted to pass-along

Introduction to Operations Research p.37

IBM SPS - Neighborhood

P(Part is required from location i) P(Part is available in the neighborhood) = = i
jN1 (i)

(1-j )

0.99994

Introduction to Operations Research p.38

IBM SPS - Neighborhood

For a given stocking level at each location and demand made at location 1. P (Part is required from location i) P (Part is available in the neighborhood) = = i
jN1 (i)

(1 j )

0.99994

2 hour radius around location 1

5 3 4

Location 1 2

P (N oP arts) = 1 0.10 0.08 0.09 0.11 0.10 0.10

P (P artRequired) 0.90 0.092 0.0073 0.0064 0.0007 0.000007

3 4 5 6

Introduction to Operations Research p.39

IBM SPS - Neighborhood

Mixed Integer Programming Model Variables
Demand ow rate from facilities to customers (continuous). Stocking levels at facilities, distribution centers, and hubs (integral).

Constraints
Satisfy demand to all customers (ow balance). Probabilistically meet all service contract requirements (parts procurement).

Size (United States)

Customers: over 2,000,000 individual across 40,000 zipcodes - aggregated to 4,000 regions. Facilities: inside 110, outside 3000. Distribution Centers: 3 Variables in MIP: 400,000 (Continuous), 122,000 (Integer) Constraints in MIP: 100,000 (Original), 660,00 (Additional Valid Cuts)

Introduction to Operations Research p.40

IBM SPS - Neighborhood

Inventory Implications
Lowers inventory investment for given PPT criteria when compared to hierarchical structure Locations not stocked independently of other locations in various vitality class neighborhoods Churn : Regional view of available inventory lessens need to "instantly" reestablish stocking levels Returns : Real-time inventory visibility pickup/dropoff parts using scheduled runs no need to return parts to original location due to regional visibility

Introduction to Operations Research p.41

IBM SPS - Neighborhood

Service Implications
Customers are assigned to closest primary location If no stock at primary location, pass-up strategy converts to pass-along Demand satised directly from least cost/least critical secondary location contained in PPT class "neighborhood" Regional orientation for providing customer service through "pooled risk"

Introduction to Operations Research p.42

Node Routing
We are given an undirected graph G = (V, E). The edges represent transportation arteries or communication links. Each edge has an associated cost or length. The nodes represent supply/demand points. Assume one supply point (the depot). A node routing is a directed subgraph G of G satisfying the following properties: G is (weakly) connected. The in-degree of each non-depot node is 1.
1 1 1 3 7 4 4 9 5 32 (depot) 1 5 5 6 1 5 1 2 7 7

Introduction to Operations Research p.43

Optimal Node Routing

Properties of a node routing. It is a spanning arborescence plus (possibly) some edges returning to the depot. There is a unique path from the depot to each demand point. We wish to construct a least cost routing. Cost Measures Sum the lengths of all edges in G . Sum the length of all paths from the depot. Some linear combination of these two.
1 1 1 3 7 4 4 9 5 32 (depot) 1 5 5 6 1 5 1 2 7 7

Introduction to Operations Research p.44

IP Formulation
IP formulation for this routing problem: M in
{i,j}E

cij xij + cij yij x((V \ {i})) = 1 y((V \ {i})) y(({i})) = di i V i V

s.t.

yij M xij {i, j} E yij 0 {i, j} E

xij {0, 1} {i, j} E where:

V = V \ {0}. xij , xji (xed-charge variables) indicate whether {i, j} is in the routing and its orientation. and yij (ow variable) represents demand ow.

Introduction to Operations Research p.45

Complexity
This node routing problem is NP-complete in general. Polynomially solvable special cases. = 0 Minimum Spanning Tree Problem. = 0 Shortest Paths Tree Problem. Note that demands are irrelevant. Other special cases. , > 0 Cable-Trench Problem (CTP). = 0 and x(({i}) = 1 Traveling Salesman Problem (TSP). > 0 and x(({i}) = 1 Variable Cost TSP (VCTSP). x((V \ {0}) = x(({0}) = k k-TSP.

Introduction to Operations Research p.46

Complexity
1100 700

800 1400

400

2100

800 100 500 600

1200

1300 1300 22500

300 2100

900

1000 2500 1800 700

900

1100 700

1100 700

800 1400

800 1400

400

2100

400

2100

800 100 500 600 600 100

800

500

1200

1200

1300 1300 22500 1300 22500

1300

300 2100

900 2100

300

900

1000 2500 1800 700

900 2500

1000

900

1800 700

Introduction to Operations Research p.47