In this course we will concentrate on optimization, especially linear opti-mization. ORMottoandLinearProgramming The most common OR tool is Linear Opt...

0 downloads 10 Views 1MB Size

Loading...

Operations Research with GNU Linear Programming Kit

Tommi Sottinen [email protected] www.uwasa.fi/∼tsottine/orms1020 August 27, 2009

Contents

I

Introduction

1 On 1.1 1.2 1.3

5

Operations Research What is Operations Research . . . . . . . . . . . . . . . . . . . History of Operations Research* . . . . . . . . . . . . . . . . . Phases of Operations Research Study . . . . . . . . . . . . . . .

6 6 8 10

2 On Linear Programming 2.1 Example towards Linear Programming . . . . . . . . . . . . . . 2.2 Solving Linear Programs Graphically . . . . . . . . . . . . . . .

13 13 15

3 Linear Programming with GNU Linear Programming Kit 3.1 Overview of GNU Linear Programming Kit . . . . . . . . . . 3.2 Getting and Installing GNU Linear Programming Kit . . . . . 3.3 Using glpsol with GNU MathProg . . . . . . . . . . . . . . . 3.4 Advanced MathProg and glpsol* . . . . . . . . . . . . . . . .

21 21 23 24 32

II

. . . .

Theory of Linear Programming

39

4 Linear Algebra and Linear Systems 4.1 Matrix Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2 Solving Linear Systems . . . . . . . . . . . . . . . . . . . . . . . 4.3 Matrices as Linear Functions* . . . . . . . . . . . . . . . . . . .

40 40 48 50

5 Linear Programs and Their Optima 5.1 Form of Linear Program . . . . . . . . 5.2 Location of Linear Programs’ Optima 5.3 Karush–Kuhn–Tucker Conditions* . . 5.4 Proofs* . . . . . . . . . . . . . . . . .

55 55 61 64 65

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

CONTENTS

2

6 Simplex Method 6.1 Towards Simplex Algorithm . . . . . . . . . . . . . . . . . . . . 6.2 Simplex Algorithm . . . . . . . . . . . . . . . . . . . . . . . . .

68 68 75

7 More on Simplex Method 87 7.1 Big M Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 87 7.2 Simplex Algorithm with Non-Unique Optima . . . . . . . . . . 94 7.3 Simplex/Big M Checklist . . . . . . . . . . . . . . . . . . . . . 102 8 Sensitivity and Duality 103 8.1 Sensitivity Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 103 8.2 Dual Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121 8.3 Primal and Dual Sensitivity . . . . . . . . . . . . . . . . . . . . 136

III

Applications of Linear Programming

9 Data Envelopment Analysis 9.1 Graphical Introduction to Data Envelopment Analysis . . 9.2 Charnes–Cooper–Rhodes Model . . . . . . . . . . . . . . . 9.3 Charnes–Cooper–Rhodes Model’s Dual . . . . . . . . . . . 9.4 Strengths and Weaknesses of Data Envelopment Analysis

137

. . . .

138 138 152 160 167

10 Transportation Problems 10.1 Transportation Algorithm . . . . . . . . . . . . . . . . . . . . . 10.2 Assignment Problem . . . . . . . . . . . . . . . . . . . . . . . . 10.3 Transshipment Problem . . . . . . . . . . . . . . . . . . . . . .

168 168 179 184

IV

190

Non-Linear Programming

. . . .

. . . .

11 Integer Programming 191 11.1 Integer Programming Terminology . . . . . . . . . . . . . . . . 191 11.2 Branch-And-Bound Method . . . . . . . . . . . . . . . . . . . . 192 11.3 Solving Integer Programs with GNU Linear Programming Kit . 199

Preface

These lecture notes are for the course ORMS1020 “Operations Research” for fall 2009 in the University of Vaasa. The notes are a slightly modified version of the notes for the fall 2008 course ORMS1020 in the University of Vaasa. The chapters, or sections of chapters, marked with an asterisk (*) may be omitted — or left for the students to read on their own time — if time is scarce. The author wishes to acknowledge that these lecture notes are collected from the references listed in Bibliography, and from many other sources the author has forgotten. The author claims no originality, and hopes not to be c laws. sued for plagiarizing or for violating the sacred Vaasa August 27, 2009

T. S.

Bibliography

[1] Rodrigo Ceron: The GNU Linear Programming Kit, Part 1: Introduction to linear optimization, Web Notes, 2006. http://www-128.ibm.com/developerworks/linux/library/l-glpk1/. [2] Matti Laaksonen: TMA.101 Operaatioanalyysi, Lecture Notes, 2005. http://lipas.uwasa.fi/ ∼ mla/orms1020/oa.html. [3] Hamdy Taha: Operations Research: An Introduction (6th Edition), Prentice Hall, Inc, 1997. [4] Wayne Winston: Operations Research: Applications and Algorithms, International ed edition, Brooks Cole, 2004.

Part I

Introduction

Chapter 1

On Operations Research

This chapter is adapted from Wikipedia article Operations Research and [4, Ch. 1].

1.1

What is Operations Research

Definitions To define anything non-trivial — like beauty or mathematics — is very difficult indeed. Here is a reasonably good definition of Operations Research: 1.1.1 Definition. Operations Research (OR) is an interdisciplinary branch of applied mathematics and formal science that uses methods like mathematical modeling, statistics, and algorithms to arrive at optimal or near optimal solutions to complex problems. Definition 1.1.1 is problematic: to grasp it we already have to know, e.g., what is formal science or near optimality. From a practical point of view, OR can be defined as an art of optimization, i.e., an art of finding minima or maxima of some objective function, and — to some extend — an art of defining the objective functions. Typical objective functions are • • • • • • •

profit, assembly line performance, crop yield, bandwidth, loss, waiting time in queue, risk.

From an organizational point of view, OR is something that helps management achieve its goals using the scientific process.

What is Operations Research

7

The terms OR and Management Science (MS) are often used synonymously. When a distinction is drawn, management science generally implies a closer relationship to Business Management. OR also closely relates to Industrial Engineering. Industrial engineering takes more of an engineering point of view, and industrial engineers typically consider OR techniques to be a major part of their tool set. Recently, the term Decision Science (DS) has also be coined to OR. More information on OR can be found from the INFORMS web page http://www.thescienceofbetter.org/ (If OR is “the Science of Better” the OR’ists should have figured out a better name for it.) OR Tools Some of the primary tools used in OR are • • • • • • • •

statistics, optimization, probability theory, queuing theory, game theory, graph theory, decision analysis, simulation.

Because of the computational nature of these fields, OR also has ties to computer science, and operations researchers regularly use custom-written software. In this course we will concentrate on optimization, especially linear optimization. OR Motto and Linear Programming The most common OR tool is Linear Optimization, or Linear Programming (LP). 1.1.2 Remark. The “Programming” in Linear Programming is synonym for “optimization”. It has — at least historically — nothing to do with computerprogramming. LP is the OR’ists favourite tool because it is • simple, • easy to understand,

History of Operations Research*

8

• robust. “Simple” means easy to implement, “easy to understand” means easy to explain (to you boss), and “robust” means that it’s like the Swiss Army Knife: perfect for nothing, but good enough for everything. Unfortunately, almost no real-world problem is really a linear one — thus LP is perfect for nothing. However, most real-world problems are “close enough” to linear problems — thus LP is good enough for everything. Example 1.1.3 below elaborates this point. 1.1.3 Example. Mr. Quine sells gavagais. He will sell one gavagai for 10 Euros. So, one might expect that buying x gavagais from Mr. Quine would cost — according to the linear rule — 10x Euros.

The linear rule in Example 1.1.3 may well hold for buying 2, 3, or 5, or even 50 gavagais. But: • One may get a discount if one buys 500 gavagais. • There are only 1,000,000 gavagais in the world. So, the price for 1,000,001 gavagais is +∞. • The unit price of gavagais may go up as they become scarce. So, buying 950,000 gavagais might be considerably more expensive than =C9,500,000. • It might be pretty hard to buy 0.5 gavagais, since half a gavagai is no longer a gavagai (gavagais are bought for pets, and not for food). • Buying −10 gavagais is in principle all right. That would simply mean selling 10 gavagais. However, Mr. Quine would most likely not buy gavagais with the same price he is selling them. 1.1.4 Remark. You may think of a curve that would represent the price of gavagais better than the linear straight line — or you may even think as a radical philosopher and argue that there is no curve. Notwithstanding the problems and limitations mentioned above, linear tools are widely used in OR according to the following motto that should — as all mottoes — be taken with a grain of salt: OR Motto. It’s better to be quantitative and naïve than qualitative and profound.

1.2

History of Operations Research*

This section is most likely omitted in the lectures. Nevertheless, you should read it — history gives perspective, and thinking is nothing but an exercise of perspective.

History of Operations Research*

9

Prehistory Some say that Charles Babbage (1791–1871) — who is arguably the “father of computers” — is also the “father of operations research” because his research into the cost of transportation and sorting of mail led to England’s universal “Penny Post” in 1840. OR During World War II The modern field of OR arose during World War II. Scientists in the United Kingdom including Patrick Blackett, Cecil Gordon, C. H. Waddington, Owen Wansbrough-Jones and Frank Yates, and in the United States with George Dantzig looked for ways to make better decisions in such areas as logistics and training schedules. Here are examples of OR studies done during World War II: • Britain introduced the convoy system to reduce shipping losses, but while the principle of using warships to accompany merchant ships was generally accepted, it was unclear whether it was better for convoys to be small or large. Convoys travel at the speed of the slowest member, so small convoys can travel faster. It was also argued that small convoys would be harder for German U-boats to detect. On the other hand, large convoys could deploy more warships against an attacker. It turned out in OR analysis that the losses suffered by convoys depended largely on the number of escort vessels present, rather than on the overall size of the convoy. The conclusion, therefore, was that a few large convoys are more defensible than many small ones. • In another OR study a survey carried out by RAF Bomber Command was analyzed. For the survey, Bomber Command inspected all bombers returning from bombing raids over Germany over a particular period. All damage inflicted by German air defenses was noted and the recommendation was given that armor be added in the most heavily damaged areas. OR team instead made the surprising and counter-intuitive recommendation that the armor be placed in the areas which were completely untouched by damage. They reasoned that the survey was biased, since it only included aircraft that successfully came back from Germany. The untouched areas were probably vital areas, which, if hit, would result in the loss of the aircraft. • When the Germans organized their air defenses into the Kammhuber Line, it was realized that if the RAF bombers were to fly in a bomber stream they could overwhelm the night fighters who flew in individual cells directed to their targets by ground controllers. It was then a matter of calculating the statistical loss from collisions against the statistical

Phases of Operations Research Study

10

loss from night fighters to calculate how close the bombers should fly to minimize RAF losses.

1.3

Phases of Operations Research Study

Seven Steps of OR Study An OR project can be split in the following seven steps: Step 1: Formulate the problem The OR analyst first defines the organization’s problem. This includes specifying the organization’s objectives and the parts of the organization (or system) that must be studied before the problem can be solved. Step 2: Observe the system Next, the OR analyst collects data to estimate the values of the parameters that affect the organization’s problem. These estimates are used to develop (in Step 3) and to evaluate (in Step 4) a mathematical model of the organization’s problem. Step 3: Formulate a mathematical model of the problem The OR analyst develops an idealized representation — i.e. a mathematical model — of the problem. Step 4: Verify the model and use it for prediction The OR analyst tries to determine if the mathematical model developed in Step 3 is an accurate representation of the reality. The verification typically includes observing the system to check if the parameters are correct. If the model does not represent the reality well enough then the OR analyst goes back either to Step 3 or Step 2. Step 5: Select a suitable alternative Given a model and a set of alternatives, the analyst now chooses the alternative that best meets the organization’s objectives. Sometimes there are many best alternatives, in which case the OR analyst should present them all to the organization’s decision-makers, or ask for more objectives or restrictions. Step 6: Present the results and conclusions The OR analyst presents the model and recommendations from Step 5 to the organization’s decision-makers. At this point the OR analyst may find that the decisionmakers do not approve of the recommendations. This may result from incorrect definition of the organization’s problems or decision-makers may disagree with the parameters or the mathematical model. The OR analyst goes back to Step 1, Step 2, or Step 3, depending on where the disagreement lies.

Phases of Operations Research Study

11

Step 7: Implement and evaluate recommendation Finally, when the organization has accepted the study, the OR analyst helps in implementing the recommendations. The system must be constantly monitored and updated dynamically as the environment changes. This means going back to Step 1, Step 2, or Step 3, from time to time. In this course we shall concentrate on Step 3 and Step 5, i.e., we shall concentrate on mathematical modeling and finding the optimum of a mathematical model. We will completely omit the in-between Step 4. That step belongs to the realm of statistics. The reason for this omission is obvious: The statistics needed in OR is way too important to be included as side notes in this course! So, any OR’ist worth her/his salt should study statistics, at least up-to the level of parameter estimization. Example of OR Study Next example elaborates how the seven-step list can be applied to a queueing problem. The example is cursory: we do not investigate all the possible objectives or choices there may be, and we do not go into the details of modeling.

1.3.1 Example. A bank manager wants to reduce expenditures on tellers’ salaries while still maintaining an adequate level of customer service.

Step 1: The OR analyst describes bank’s objectives. The manager’s vaguely stated wish may mean, e.g., • The bank wants to minimize the weekly salary cost needed to ensure that the average waiting a customer waits in line is at most 3 minutes. • The bank wants to minimize the weekly salary cost required to ensure that only 5% of all customers wait in line more than 3 minutes. The analyst must also identify the aspects of the bank’s operations that affect the achievement of the bank’s objectives, e.g., • On the average, how many customers arrive at the bank each hour? • On the average, how many customers can a teller serve per hour? Step 2: The OR analyst observes the bank and estimates, among others, the following parameters: • On the average, how many customers arrive each hour? Does the arrival rate depend on the time of day?

Phases of Operations Research Study

12

• On the average, how many customers can a teller serve each hour? Does the service speed depend on the number of customers waiting in line? Step 3: The OR analyst develops a mathematical model. In this example a queueing model is appropriate. Let Wq = Average time customer waits in line λ = Average number of customers arriving each hour µ = Average number of customers teller can serve each hour A certain mathematical queueing model yields a connection between these parameters: (1.3.2)

Wq

=

λ . µ(µ − λ)

This model corresponds to the first objective stated in Step 1. Step 4: The analyst tries to verify that the model (1.3.2) represents reality well enough. This means that the OR analyst will estimate the parameter Wq , λ, and µ statistically, and then she will check whether the equation (1.3.2) is valid, or close enough. If this is not the case then the OR analyst goes either back to Step 2 or Step 3. Step 5: The OR analyst will optimize the model (1.3.2). This could mean solving how many tellers there must be to make µ big enough to make Wq small enough, e.g. 3 minutes. We leave it to the students to wonder what may happen in steps 6 and 7.

Chapter 2

On Linear Programming

This chapter is adapted from [2, Ch. 1].

2.1

Example towards Linear Programming

Very Naïve Problem

2.1.1 Example. Tela Inc. manufactures two product: #1 and #2. To manufacture one unit of product #1 costs =C40 and to manufacture one unit of product #2 costs =C60. The profit from product #1 is =C30, and the profit from product #2 is =C20. The company wants to maximize its profit. How many products #1 and #2 should it manufacture?

The solution is trivial: There is no bound on the amount of units the company can manufacture. So it should manufacture infinite number of either product #1 or #2, or both. If there is a constraint on the number of units manufactured then the company should manufacture only product #1, and not product #2. This constrained case is still rather trivial. Less Naïve Problem Things become more interesting — and certainly more realistic — when there are restrictions in the resources.

Example towards Linear Programming

14

2.1.2 Example. Tela Inc. in Example 2.1.1 can invest =C40, 000 in production and use 85 hours of labor. To manufacture one unit of product #1 requires 15 minutes of labor, and to manufacture one unit of product #2 requires 9 minutes of labor. The company wants to maximize its profit. How many units of product #1 and product #2 should it manufacture? What is the maximized profit?

The rather trivial solution of Example 2.1.1 is not applicable now. Indeed, the company does not have enough labor to put all the =C40,000 in product #1. Since the profit to be maximized depend on the number of product #1 and #1, our decision variables are: x1 = number of product #1 produced, x2 = number of product #2 produced. So the situation is: We want to maximize (max) profit:

30x1 + 20x2

subject to (s.t.) the constraints money: 40x1 + 60x2 ≤ 40,000 labor: 15x1 + 9x2 ≤ 5,100 non-negativity: x1 , x2 ≥ 0 Note the last constraint: x1 , x2 ≥ 0. The problem does not state this explicitly, but it’s implied — we are selling products #1 and #2, not buying them. 2.1.3 Remark. Some terminology: The unknowns x1 and x2 are called decision variables. The function 30x1 +20x2 to be maximized is called the objective function. What we have now is a Linear problem, max z = 30x1 s.t. 40x1 15x1

Program (LP), or a Linear Optimization + 20x2 + 60x2 ≤ 40,000 + 9x2 ≤ 5,100 x1 , x2 ≥ 0

We will later see how to solve such LPs. For now we just show the solution. For decision variables it is optimal to produce no product #1 and thus put all

Solving Linear Programs Graphically

15

the resource to product #2 which means producing 566.667 number of product #2. The profit will then be =C11,333.333. In other words, the optimal solution is x1 = 0, x2 = 566.667, z = 11,333.333. 2.1.4 Remark. If it is not possible to manufacture fractional number of products, e.g. 0.667 units, then we have to reformulate the LP-problem above to an Integer Program (IP) max z = 30x1 + 20x2 s.t. 40x1 + 60x2 15x1 + 9x2 x1 , x2 x1 , x2

≤ 40,000 ≤ 5,100 ≥ 0 are integers

We will later see how to solve such IPs (which is more difficult than solving LPs). For now we just show the solution: x1 = 1, x2 = 565, z = 11,330. In Remark 2.1.4 above we see the usefulness of the OR Motto. Indeed, although the LP solution is not practical if we cannot produce fractional number of product, the solution it gives is close to the true IP solution: both in terms of the value of the objective function and the location of the optimal point. We shall learn more about this later in Chapter 8.

2.2

Solving Linear Programs Graphically

From Minimization to Maximization We shall discuss later in Chapter 5, among other things, how to transform a minimization LP into a maximization LP. So, you should skip this subsection and proceed to the next subsection titled “Linear Programs with Two Decision Variables” — unless you want to know the general, and rather trivial, duality between minimization and maximization. Any minimization problem — linear or not — can be restated as a maximization problem simply by multiplying the objective function by −1: 2.2.1 Theorem. Let K ⊂ Rn , and let g : Rn → R. Suppose w∗

=

min g(x)

x∈K

Solving Linear Programs Graphically

16

and x∗ ∈ Rn is a point where the minimum w∗ is attained. Then, if f = −g and z ∗ = −w∗ , we have that z∗

= max f (x), x∈K

and the maximum z ∗ is attained at the point x∗ ∈ Rn . The mathematically oriented should try to prove Theorem 2.2.1. It’s not difficult — all you have to do is to not to think about the constraint-set K or any other specifics, like the space Rn , or if there is a unique optimum. Just think about the big picture! Indeed, Theorem 2.2.1 is true in the greatest possible generality: It is true whenever it makes sense! Linear Programs with Two Decision Variables We shall solve the following LP:

2.2.2 Example. max z = s.t.

4x1 + 3x2 2x1 + 3x2 −3x1 + 2x2 2x2 2x1 + x2 x1 , x2

≤ ≤ ≤ ≤ ≥

6 3 5 4 0

(1) (2) (3) (4) (5)

The LP in Example 2.2.2 has only two decision variables: x1 and x2 . So, it can be solved graphically on a piece of paper like this one. To solve graphically LPs with three decision variables would require three-dimensional paper, for four decision variables one needs four-dimensional paper, and so forth. Four-Step Graphical Algorithm Step 1: Draw coordinate space Tradition is that x1 is the horizontal axis and x2 is the vertical axis. Because of the non-negativity constraints on x1 and x2 it is enough to draw the 1st quadrant (the NE-quadrant). Step 2: Draw constraint-lines Each constraint consists of a line and of information (e.g. arrows) indicating which side of the line is feasible. To draw, e.g., the line (1), one sets the inequality to be the equality 2x1 + 3x2 = 6.

Solving Linear Programs Graphically

17

To draw this line we can first set x1 = 0 and then set x2 = 0, and we see that the line goes through points (0, 2) and (3, 0). Since (1) is a ≤-inequality, the feasible region must be below the line. Step 3: Define feasible region This is done by selecting the region satisfied by all the constraints including the non-negativity constraints. Step 4: Find the optimum by moving the isoprofit line The isoprofit line is the line where the objective function is constant. In this case the isoprofit lines are the pairs (x1 , x2 ) satisfying z = 4x1 + 3x2 = const. (In the following picture we have drawn the isoprofit line corresponding to const = 2 and const = 4, and the optimal isoprofit line corresponding to const = 9.) The further you move the line from the origin the better value you get (unless the maximization problem is trivial in the objective function, cf. Example 2.2.3 later). You find the best value when the isoprofit line is just about to leave the feasible region completely (unless the maximization problem is trivial in constraints, i.e. it has an unbounded feasible region, cf. Example 2.2.4 later).

Solving Linear Programs Graphically

x2 4

18

(2)

(4)

Redundant

3

(3)

2

D

E 1

Optimum Feasible region

C

Isoprofit lines (1) 0 A0

B 2

1

3

4 x1

From the picture we read — by moving the isoprofit line away from the origin — that the optimal point for the decision variables (x1 , x2 ) is C

= (1.5, 1).

Therefore, the optimal value is of the objective is z = 4×1.5 + 3×1 = 9. Example with Trivial Optimum Consider the following LP maximization problem, where the objective function z does not grow as its arguments x1 and x2 get further away from the origin:

Solving Linear Programs Graphically

19

2.2.3 Example. max z = −4x1 − 3x2 s.t. 2x1 + 3x2 −3x1 + 2x2 2x2 2x1 + x2 x1 , x2

≤ ≤ ≤ ≤ ≥

6 3 5 4 0

(1) (2) (3) (4) (5)

In this case drawing a graph would be an utter waste of time. Indeed, consider the objective function under maximization: z = −4x1 − 3x2 Obviously, given the standard constraints x1 , x2 ≥ 0, the optimal solution is x1 = 0, x2 = 0, z = 0. Whenever you have formulated a problem like this you (or your boss) must have done something wrong! Example with Unbounded Feasible Region

2.2.4 Example. max z = 4x1 + 3x2 s.t. −3x1 + 2x2 ≤ 3 x1 , x2 ≥ 0

(1) (2)

From the picture below one sees that this LP has unbounded optimum, i.e., the value of objective function z can be made as big as one wishes.

Solving Linear Programs Graphically

20

x2 4

(1) 3

Feasible region

2

1 Isoprofit lines 0

0

1

2

3

4 x1

Whenever you have formulated a problem like this you (or your boss) must have done something wrong — or you must be running a sweet business, indeed!

Chapter 3

Linear Programming with GNU Linear Programming Kit

This chapter is adapted from [1].

3.1

Overview of GNU Linear Programming Kit

GNU Linear Programming Kit The GNU Linear Programming Kit (GLPK) is a software package intended for solving large-scale linear programming (LP), mixed integer programming (MIP), and other related problems. GLPK is written in ANSI C and organized as a callable library. GLPK package is part of the GNU Project and is released under the GNU General Public License (GPL). The GLPK package includes the following main components: • • • • • •

Revised simplex method (for LPs). Primal-dual interior point method (for LPs). Branch-and-bound method (for IPs). Translator for GNU MathProg modeling language. Application Program Interface (API). Stand-alone LP/MIP solver glpsol.

glpsol GLPK is not a program — it’s a library. GLPK can’t be run as a computer program. Instead, client programs feed the problem data to GLPK through the GLPK API and receive results back. However, GLPK has a default client: The glpsol program that interfaces with the GLPK API. The name “glpsol” comes from GNU Linear P rogram Sol ver. Indeed, usually a program like glpsol is called a solver rather than a client, so we shall use this nomenclature from here forward.

Overview of GNU Linear Programming Kit

22

There is no standard Graphical User Interface (GUI) for glpsol that the author is aware of. So one has to call glpsol from a console. If you do not know how to use to a console in your Windows of Mac, ask your nearest guru now! Linux users should know how to use a console. 3.1.1 Remark. If you insist on having a GUI for GLPK, you may try • http://bjoern.dapnet.de/glpk/ (Java GUI) • http://qosip.tmit.bme.hu/∼retvari/Math-GLPK.html (Perl GUI) • http://www.dcc.fc.up.pt/∼jpp/code/python-glpk/ (Python GUI) The author does not know if any of these GUIs are any good. To use glpsol we issue on a console the command glpsol -m inputfile.mod -o outputfile.sol or the command glpsol –model inputfile.mod –output outputfile.sol The commands above mean the same. Indeed, -m is an abbreviation to –model, and -o is an abbreviation to –output. The option -m inputfile.mod tells the glpsol solver that the model to be solved is described in the file inputfile.mod, and the model is described in the GNU MathProg language. The option -o outputfile.sol tells the glpsol solver to print the results (the solution with some sensitivity information) to the file outputfile.sol. GNU MathProg The GNU MathProg is a modeling language intended for describing linear mathematical programming models. Model descriptions written in the GNU MathProg language consists of: • • • •

Problem decision variables. An objective function. Problem constraints. Problem parameters (data).

As a language the GNU MathProg is rather extensive, and can thus be a bit confusing. We shall not give a general description of the language in this course, but learn the elements of it through examples.

Getting and Installing GNU Linear Programming Kit

3.2

23

Getting and Installing GNU Linear Programming Kit

GLPK, like all GNU software, is open source: It is available to all operating systems and platforms you may ever use. This is the reason we use GLPK in this course instead of, say, LINDO. General information on how to get GLPK and other GNU software can be found from http://www.gnu.org/prep/ftp.html. The GLPK source code can be downloaded from ftp://ftp.funet.fi/pub/gnu/prep/glpk/. If you use this method you have to compile GLPK by yourself. If you do not know what this means try following the instructions given in the links in one of the following subsections: Windows, Mac, or Linux. Windows From http://gnuwin32.sourceforge.net/packages/glpk.htm you should find a link to a setup program that pretty much automatically installs the GLPK for you. Some installation instructions can be found from http://gnuwin32.sourceforge.net/install.html. The instructions given above may or may not work with Windows Vista. Mac From http://glpk.darwinports.com/ you find instruction how to install GLPK for Mac OS X. Linux If you are using Ubuntu then just issue the command sudo apt-get install glpk in the console (you will be prompted for your password). Alternatively, you can use the Synaptic Package Manager: Just search for glpk. If you use a RedHat based system, consult

Using glpsol with GNU MathProg

24

http://rpmfind.net/ with the keyword glpk. Debian users can consult http://packages.debian.org/etch/glpk. For other Linux distributions, consult http://www.google.com/.

3.3

Using glpsol with GNU MathProg

We show how to build up an LP model, how to describe the LP problem by using the GNU MathProg language, and how to solve it by using the glpsol program. Finally, we discuss how to interpret the results. Giapetto’s Problem Consider the following classical problem:

3.3.1 Example. Giapetto’s Woodcarving Inc. manufactures two types of wooden toys: soldiers and trains. A soldier sells for =C27 and uses =C10 worth of raw materials. Each soldier that is manufactured increases Giapetto’s variable labor and overhead costs by =C14. A train sells for =C21 and uses =C9 worth of raw materials. Each train built increases Giapetto’s variable labor and overhead costs by =C10. The manufacture of wooden soldiers and trains requires two types of skilled labor: carpentry and finishing. A soldier requires 2 hours of finishing labor and 1 hour of carpentry labor. A train requires 1 hour of finishing and 1 hour of carpentry labor. Each week, Giapetto can obtain all the needed raw material but only 100 finishing hours and 80 carpentry hours. Demand for trains is unlimited, but at most 40 soldier are bought each week. Giapetto wants to maximize weekly profits (revenues - costs).

To summarize the important information and assumptions about this problem: 1. There are two types of wooden toys: soldiers and trains. 2. A soldier sells for =C27, uses =C10 worth of raw materials, and increases variable labor and overhead costs by =C14.

Using glpsol with GNU MathProg

25

3. A train sells for =C21, uses =C9 worth of raw materials, and increases variable labor and overhead costs by =C10. 4. A soldier requires 2 hours of finishing labor and 1 hour of carpentry labor. 5. A train requires 1 hour of finishing labor and 1 hour of carpentry labor. 6. At most, 100 finishing hours and 80 carpentry hours are available weekly. 7. The weekly demand for trains is unlimited, while, at most, 40 soldiers will be sold. The goal is to find: 1. the numbers of soldiers and trains that will maximize the weekly profit, 2. the maximized weekly profit itself. Mathematical Model for Giapetto To model a linear problem (Giapetto’s problem is a linear one — we will see this soon), the decision variables are established first. In Giapetto’s shop, the objective function is the profit, which is a function of the amount of soldiers and trains produced each week. Therefore, the two decision variables in this problem are: x1 : Number of soldiers produced each week x2 : Number of trains produced each week Once the decision variables are known, the objective function z of this problem is simply the revenue minus the costs for each toy, as a function of x1 and x2 : z = (27 − 10 − 14)x1 + (21 − 9 − 10)x2 = 3x1 + 2x2 . Note that the profit z depends linearly on x1 and x2 — this is a linear problem, so far (the constraints must turn out to be linear, too). It may seem at first glance that the profit can be maximized by simply increasing x1 and x2 . Well, if life were that easy, let’s start manufacturing trains and soldiers and move to the Jamaica! Unfortunately, there are restrictions that limit the decision variables that may be selected (or else the model is very likely to be wrong). Recall the assumptions made for this problem. The first three determined the decision variables and the objective function. The fourth and sixth assumption say that finishing the soldiers requires time for carpentry and finishing. The limitation here is that Giapetto does not have infinite carpentry and finishing hours. That’s a constraint! Let’s analyze it to clarify. One soldier requires 2 hours of finishing labor, and Giapetto has at most 100 hours of finishing labor per week, so he can’t produce more than 50 soldiers per

Using glpsol with GNU MathProg

26

week. Similarly, the carpentry hours constraint makes it impossible to produce more than 80 soldiers weekly. Note here that the first constraint is stricter than the second. The first constraint is effectively a subset of the second, thus the second constraint is redundant. The previous paragraph shows how to model optimization problems, but it’s an incomplete analysis because all the necessary variables were not considered. It’s not the complete solution of the Giapetto problem. So how should the problem be approached? Start by analyzing the limiting factors first in order to find the constraints. First, what constrains the finishing hours? Since both soldiers and trains require finishing time, both need to be taken into account. Suppose that 10 soldiers and 20 trains were built. The amount of finishing hours needed for that would be 10 times 2 hours (for soldiers) plus 20 times 1 hour (for trains), for a total of 40 hours of finishing labor. The general constraint in terms of the decision variables is: 2x1 + x2 ≤ 100. This is a linear constraint, so we are still dealing with a linear program. Now that the constraint for the finishing hours is ready, the carpentry hours constraint is found in the same way to be: x1 + x2 ≤ 80. This constraint is a linear one. There’s only one more constraint remaining for this problem. Remember the weekly demand for soldiers? According to the problem description, there can be at most 40 soldiers produced each week: x1 ≤ 40, again a linear constraint. The demand for trains is unlimited, so no constraint there. The modeling phase is finished, and we have the following LP: max z = 3x1 + 2x2 s.t. 2x1 + x2 x1 + x2 x1 x1 , x2

≤ 100 ≤ 80 ≤ 40 ≥ 0

(objective function) (finishing constraint) (carpentry constraint) (demand for soldiers) (sign constraints)

Note the last constraint. It ensures that the values of the decision variables will always be positive. The problem does not state this explicitly, but it’s still important (and obvious). The problem also implies that the decision variables are integers, but we are not dealing with IPs yet. So, we will just hope that

Using glpsol with GNU MathProg

27

the optimal solution will turn out to be an integer one (it will, but that’s just luck). Now GLPK can solve the model (since GLPK is good at solving linear optimization problems). Describing the Model with GNU MathProg The mathematical formulation of Giapetto’s problem needs to be written with the GNU MathProg language. The key items to declare are: • • • •

The The The The

decision variables objective function constraints problem data set

The following code, written in the (ASCII) text file giapetto.mod, shows how to solve Giapetto’s problem with GNU MathProg. The line numbers are not part of the code itself. They have been added only for the sake of making references to the code. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19

# # Giapetto’s problem # # This finds the optimal solution for maximizing Giapetto’s profit # /* Decision variables */ var x1 >=0; /* soldier */ var x2 >=0; /* train */ /* Objective function */ maximize z: 3*x1 + 2*x2; /* Constraints s.t. Finishing s.t. Carpentry s.t. Demand

*/ : 2*x1 + x2 <= 100; : x1 + x2 <= 80; : x1 <= 40;

end;

Lines 1 through 5 are comments: # anywhere on a line begins a comment to the end of the line. C-style comments can also be used, as shown on line 7. They even work in the middle of a declaration. Comments are there to make the code more readable for a human reader. Computer reader, i.e. the GNU MathProg translator, will ignore comments. Empty lines, like line 6, are simply ignored by the MathProg translator. So, empty lines can be used to structure the code more readable.

Using glpsol with GNU MathProg

28

The first MathProg step is to declare the decision variables. Lines 8 and 9 declare x1 and x2. A decision variable declaration begins with the keyword var. To simplify sign constraints, GNU MathProg allows a non-negativity constraint >=0 in the decision variable declaration, as seen on lines 8 and 9. Every sentence in GNU MathProg must end with a semicolon (;). Be careful! It is very easy to forget those little semicolons at the end of declarations! (Even moderately experienced programmers should be very familiar with this problem.) Recall that x1 represents soldier numbers, and x2 represents train numbers. These variables could have been called soldiers and trains, but that would confuse the mathematicians in the audience. In general, it is good practice to use x for decision variables and z for the objective function. That way you will always spot them out quickly. Line 12 declares the objective function. LP’s can be either maximized or minimized. Giapetto’s mathematical model is a maximization problem, so the keyword maximize is appropriate instead of the opposite keyword, minimize. The objective function is named z. It could have been named anything, e.g. profit, but this is not good practice, as noted before. Line 12 sets the objective function z to equal 3*x1+2*x2. Note that: • The colon (:) character separates the name of the objective function and its definition. Don’t use equal sign (=) to define the objective function! • The asterisk (*) character denotes multiplication. Similarly, the plus (+), minus (-), and forward slash (/) characters denote addition, subtraction, and division as you’d expect. If you need powers, then use either the circumflex (ˆ) or the double-star (**). Lines 15, 16, and 17 define the constraints. The keyword s.t. (subject to) is not required, but it improves the readability of the code (Remember good practice!). The three Giapetto constraints have been labeled Finishing, Carpentry, and Demand. Each of them is declared as in the mathematical model. The symbols <= and >= express the inequalities ≤ and ≥, respectively. Don’t forget the ; at the end of each declaration. Every GNU MathProg text file must end with end;, as seen on line 19. 3.3.2 Remark. There was no data section in the code. The problem was so simple that the problem data is directly included in the objective function and constraints declarations as the coefficients of the decision variables in the declarations. For example, in the objective function, the coefficients 3 and 1 are part of the problem’s data set. Solving the Model with glpsol Now, glpsol can use the text file giapetto.mod as input. It is good practice to use the .mod extension for GNU MathProg input files and redirect the solution

Using glpsol with GNU MathProg

29

to a file with the extension .sol. This is not a requirement — you can use any file name and extension you like. Giapetto’s MathProg file for this example will be giapetto.mod, and the output will be in the text file giapetto.sol. Now, run glpsol in your favorite console: glpsol -m giapetto.mod -o giapetto.sol This command line uses two glpsol options: -m The -m or –model option tells glpsol that the input in the file giapetto.mod is written in GNU MathProg (the default modeling language for glpsol). -o The -o or –output option tells glpsol to send its output to the file giapetto.sol. 3.3.3 Remark. Some people prefer to use the extension .txt to indicate that the file in question is a (ASCII) text file. In that case giapetto.mod would be, e.g., giapetto_mod.txt. Similarly, giapetto.sol would be, e.g., giapetto_sol.txt. The command would be glpsol -m giapetto_mod.txt -o giapetto_sol.txt The solution report will be written into the text file giapetto.sol (unless you used the .txt extension style command of Remark 3.3.3), but some information about the time and memory GLPK consumed is shown on the system’s standard output (usually the console window): Reading model section from giapetto.mod... 19 lines were read Generating z... Generating Finishing... Generating Carpentry... Generating Demand... Model has been successfully generated glp_simplex: original LP has 4 rows, 2 columns, 7 non-zeros glp_simplex: presolved LP has 2 rows, 2 columns, 4 non-zeros lpx_adv_basis: size of triangular part = 2 * 0: objval = 0.000000000e+00 infeas = 0.000000000e+00 (0) * 3: objval = 1.800000000e+02 infeas = 0.000000000e+00 (0) OPTIMAL SOLUTION FOUND Time used: 0.0 secs Memory used: 0.1 Mb (114537 bytes) lpx_print_sol: writing LP problem solution to ‘giapetto.sol’...

The report shows that glpsol reads the model from the file giapetto.mod that has 19 lines, calls a GLPK API function to generate the objective function, and then calls another GLPK API function to generate the constraints.

Using glpsol with GNU MathProg

30

After the model has been generated, glpsol explains briefly how the problem was handled internally by GLPK. Then it notes that an optimal solution is found. Then, there is information about the resources used by GLPK to solve the problem (Time used 0.0 secs, wow!). Finally the report tells us that the solution is written to the file giapetto.sol. Now the optimal values for the decision variables x1 and x1, and the optimal value of the objective function z are in the giapetto.sol file. It is a standard text file that can be opened in any text editor (e.g. Notepad in Windows, gedit in Linux with Gnome desktop). Here are its contents: Problem: Rows: Columns: Non-zeros: Status: Objective:

giapetto 4 2 7 OPTIMAL z = 180 (MAXimum)

No. -----1 2 3 4

Row name -----------z Finishing Carpentry Demand

St Activity Lower bound Upper bound Marginal -- ------------- ------------- ------------- ------------B 180 NU 100 100 1 NU 80 80 1 B 20 40

No. -----1 2

Column name -----------x1 x2

St Activity Lower bound Upper bound Marginal -- ------------- ------------- ------------- ------------B 20 0 B 60 0

Karush-Kuhn-Tucker optimality conditions: KKT.PE: max.abs.err. = 0.00e+00 on row 0 max.rel.err. = 0.00e+00 on row 0 High quality KKT.PB: max.abs.err. = 0.00e+00 on row 0 max.rel.err. = 0.00e+00 on row 0 High quality KKT.DE: max.abs.err. = 0.00e+00 on column 0 max.rel.err. = 0.00e+00 on column 0 High quality KKT.DB: max.abs.err. = 0.00e+00 on row 0 max.rel.err. = 0.00e+00 on row 0 High quality End of output

Using glpsol with GNU MathProg

31

Interpreting the Results The solution in the text file giapetto.sol is divided into four sections: 1. Information about the problem and the optimal value of the objective function. 2. Precise information about the status of the objective function and about the constraints. 3. Precise information about the optimal values for the decision variables. 4. Information about the optimality conditions, if any. Let us look more closely: Information about the optimal value of the objective function is found in the first part: Problem: Rows: Columns: Non-zeros: Status: Objective:

giapetto 4 2 7 OPTIMAL z = 180 (MAXimum)

For this particular problem, we see that the solution is OPTIMAL and that Giapetto’s maximum weekly profit is =C180. Precise information about the status of the objective function and about the constraints are found in the second part: No. -----1 2 3 4

Row name -----------z Finishing Carpentry Demand

St Activity Lower bound Upper bound Marginal -- ------------- ------------- ------------- ------------B 180 NU 100 100 1 NU 80 80 1 B 20 40

The Finishing constraint’s status is NU (the St column). NU means that there’s a non-basic variable (NBV) on the upper bound for that constraint. Later, when you know more operation research theory you will understand more profoundly what this means, and you can build the simplex tableau and check it out. For now, here is a a brief practical explanation: Whenever a constraint reaches its upper or lower boundary, it’s called a bounded, or active, constraint. A bounded constraint prevents the objective function from reaching a better value. When that occurs, glpsol shows the status of the constraint as either NU or NL (for upper and lower boundary respectively), and it also shows the value of the marginal, also known as the shadow price. The marginal is the value by which the objective function would improve if the constraint were relaxed by one unit. Note that the improvement depends on whether the goal is to minimize or maximize the target function. For instance, in Giapetto’s problem, which seeks maximization, the marginal value 1 means that the objective

Advanced MathProg and glpsol*

32

function would increase by 1 if we could have one more hour of finishing labor (we know it’s one more hour and not one less, because the finishing hours constraint is an upper boundary). The carpentry and soldier demand constraints are similar. For the carpentry constraint, note that it’s also an upper boundary. Therefore, a relaxation of one unit in that constraint (an increment of one hour) would make the objective function’s optimal value become better by the marginal value 1 and become 181. The soldier demand, however, is not bounded, thus its state is B, and a relaxation in it will not change the objective function’s optimal value. Precise information about the optimal values for the decision variables is found in the third part: No. -----1 2

Column name -----------x1 x2

St Activity Lower bound Upper bound Marginal -- ------------- ------------- ------------- ------------B 20 0 B 60 0

The report shows the values for the decision variables: x1 = 20 and x2 = 60. This tells Giapetto that he should produce 20 soldiers and 60 trains to maximize his weekly profit. (The solution was an integer one. We were lucky: It may have been difficult for Giapetto to produce, say, 20.5 soldiers.) Finally, in the fourth part, Karush-Kuhn-Tucker optimality conditions: KKT.PE: max.abs.err. = 0.00e+00 on row 0 max.rel.err. = 0.00e+00 on row 0 High quality KKT.PB: max.abs.err. = 0.00e+00 on row 0 max.rel.err. = 0.00e+00 on row 0 High quality KKT.DE: max.abs.err. = 0.00e+00 on column 0 max.rel.err. = 0.00e+00 on column 0 High quality KKT.DB: max.abs.err. = 0.00e+00 on row 0 max.rel.err. = 0.00e+00 on row 0 High quality

there is some technical information mostly for the case when the algorithm is not sure if the optimal solution is found. In this course we shall omit this topic completely.

3.4

Advanced MathProg and glpsol*

This section is not necessarily needed in the rest of the course, and can thus be omitted if time is scarce.

Advanced MathProg and glpsol*

33

glpsol Options To get an idea of the usage and flexibility of glpsol here is the output of the command glpsol -h or the command glpsol –help which is the same command (-h is just an abbreviation to –help): Usage: glpsol [options...] filename General options: --mps --freemps --cpxlp --math

read LP/MIP problem in Fixed MPS format read LP/MIP problem in Free MPS format (default) read LP/MIP problem in CPLEX LP format read LP/MIP model written in GNU MathProg modeling language -m filename, --model filename read model section and optional data section from filename (the same as --math) -d filename, --data filename read data section from filename (for --math only); if model file also has data section, that section is ignored -y filename, --display filename send display output to filename (for --math only); by default the output is sent to terminal -r filename, --read filename read solution from filename rather to find it with the solver --min minimization --max maximization --scale scale problem (default) --noscale do not scale problem --simplex use simplex method (default) --interior use interior point method (for pure LP only) -o filename, --output filename write solution to filename in printable format -w filename, --write filename write solution to filename in plain text format --bounds filename write sensitivity bounds to filename in printable format (LP only) --tmlim nnn limit solution time to nnn seconds --memlim nnn limit available memory to nnn megabytes --check do not solve problem, check input data only --name probname change problem name to probname --plain use plain names of rows and columns (default)

Advanced MathProg and glpsol*

--orig

try using original names of rows and columns (default for --mps) --wmps filename write problem to filename in Fixed MPS format --wfreemps filename write problem to filename in Free MPS format --wcpxlp filename write problem to filename in CPLEX LP format --wtxt filename write problem to filename in printable format --wpb filename write problem to filename in OPB format --wnpb filename write problem to filename in normalized OPB format --log filename write copy of terminal output to filename -h, --help display this help information and exit -v, --version display program version and exit LP basis factorization option: --luf LU + Forrest-Tomlin update (faster, less stable; default) --cbg LU + Schur complement + Bartels-Golub update (slower, more stable) --cgr LU + Schur complement + Givens rotation update (slower, more stable) Options specific to simplex method: --std use standard initial basis of all slacks --adv use advanced initial basis (default) --bib use Bixby’s initial basis --bas filename read initial basis from filename in MPS format --steep use steepest edge technique (default) --nosteep use standard "textbook" pricing --relax use Harris’ two-pass ratio test (default) --norelax use standard "textbook" ratio test --presol use presolver (default; assumes --scale and --adv) --nopresol do not use presolver --exact use simplex method based on exact arithmetic --xcheck check final basis using exact arithmetic --wbas filename write final basis to filename in MPS format Options specific to MIP: --nomip consider all integer variables as continuous (allows solving MIP as pure LP) --first branch on first integer variable --last branch on last integer variable --drtom branch using heuristic by Driebeck and Tomlin (default) --mostf branch on most fractional varaible --dfs backtrack using depth first search --bfs backtrack using breadth first search --bestp backtrack using the best projection heuristic --bestb backtrack using node with best local bound (default) --mipgap tol set relative gap tolerance to tol --intopt use advanced MIP solver --binarize replace general integer variables by binary ones (assumes --intopt) --cover generate mixed cover cuts

34

Advanced MathProg and glpsol*

--clique --gomory --mir --cuts

generate generate generate generate

35

clique cuts Gomory’s mixed integer cuts MIR (mixed integer rounding) cuts all cuts above (assumes --intopt)

For description of the MPS and CPLEX LP formats see Reference Manual. For description of the modeling language see "GLPK: Modeling Language GNU MathProg". Both documents are included in the GLPK distribution. See GLPK web page at

Using Model and Data Sections Recall Giapetto’s problem. It was very small. You may be wondering, in a problem with many more decision variables and constraints, would you have to declare each variable and each constraint separately? And what if you wanted to adjust the data of the problem, such as the selling price of a soldier? Do you have to make changes everywhere this value appears? MathProg models normally have a model section and a data section, sometimes in two different files. Thus, glpsol can solve a model with different data sets easily, to check what the solution would be with this new data. The following listing, the contents of the text file giapetto2.mod, states Giapetto’s problem in a much more elegant way. Again, the line numbers are here only for the sake of reference, and are not part of the actual code. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22

# # Giapetto’s problem (with data section) # # This finds the optimal solution for maximizing Giapetto’s profit # /* Set of toys */ set TOY; /* Parameters */ param Finishing_hours param Carpentry_hours param Demand_toys param Profit_toys

{i {i {i {i

in in in in

TOY}; TOY}; TOY}; TOY};

/* Decision variables */ var x {i in TOY} >=0; /* Objective function */ maximize z: sum{i in TOY} Profit_toys[i]*x[i]; /* Constraints */

Advanced MathProg and glpsol*

23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49

36

s.t. Fin_hours : sum{i in TOY} Finishing_hours[i]*x[i] <= 100; s.t. Carp_hours : sum{i in TOY} Carpentry_hours[i]*x[i] <= 80; s.t. Dem {i in TOY} : x[i] <= Demand_toys[i];

data; /* data

section */

set TOY := soldier train; param Finishing_hours:= soldier 2 train 1; param Carpentry_hours:= soldier 1 train 1; param Demand_toys:= soldier 40 train 6.02E+23; param Profit_toys:= soldier 3 train 2; end;

Rather than two separate files, the problem is stated in a single file with a modeling section (lines 1 through 27) and a data section (lines 28 through 49). Line 8 defines a SET. A SET is a universe of elements. For example, if I declare mathematically x, for all i in {1;2;3;4}, I’m saying that x is an array, or vector, that ranges from 1 to 4, and therefore we have x[1], x[2], x[3], x[4]. In this case, {1;2;3;4} is the set. So, in Giapetto’s problem, there’s a set called TOY. Where are the actual values of this set? In the data section of the file. Check line 31. It defines the TOY set to contain soldier and train. Wow, the set is not a numerical range. How can that be? GLPK handles this in an interesting way. You’ll see how to use this in a few moments. For now, get used to the syntax for SET declarations in the data section: set label := value1 value2 ... valueN; Lines 11, 12, and 13 define the parameters of the problem. There are three: Finishing_hours, Carpentry_hours, and Demand_toys. These parameters make up the problem’s data matrix and are used to calculate the constraints, as you’ll see later. Take the Finishing_hours parameter as an example. It’s defined on the TOY set, so each kind of toy in the TOY set will have a value for Finishing_hours. Remember that each soldier requires 2 hours of finishing work, and each train requires 1 hour of finishing work. Check lines 33,

Advanced MathProg and glpsol*

37

34, and 35 now. There is the definition of the finishing hours for each kind of toy. Essentially, those lines declare that Finishing_hours[soldier]=2 and Finishing_hours[train]=1. Finishing_hours is, therefore, a matrix with 1 row and 2 columns, or a row vector of dimension 2. Carpentry hours and demand parameters are declared similarly. Note that the demand for trains is unlimited, so a very large value is the upper bound on line 43. Line 17 declares a variable, x, for every i in TOY (resulting in x[soldier] and x[train]), and constrains them to be greater than or equal to zero. Once you have sets, it’s pretty easy to declare variables, isn’t it? Line 20 declares the objective (target) function as the maximization of z, which is the total profit for every kind of toy (there are two: trains and soldiers). With soldiers, for example, the profit is the number of soldiers times the profit per soldier. The constraints on lines 23, 24, and 25 are declared in a similar way. Take the finishing hours constraint as an example: it’s the total of the finishing hours per kind of toy, times the number of that kind of toy produced, for the two types of toys (trains and soldiers), and it must be less than or equal to 100. Similarly, the total carpentry hours must be less than or equal to 80. The demand constraint is a little bit different than the previous two, because it’s defined for each kind of toy, not as a total for all toy types. Therefore, we need two of them, one for trains and one for soldiers, as you can see on line 25. Note that the index variable ( {i in TOY} ) comes before the :. This tells GLPK to create a constraint for each toy type in TOY, and the equation that will rule each constraint will be what comes after the :. In this case, GLPK will create Dem[soldier] : x[soldier] <= Demand[soldier] Dem[train] : x[train] <= Demand[train] Solving this new model must yield the same results. So issue the command glpsol -m giapetto2.mod -o giapetto2.sol and the text file giapetto2.sol should read: Problem: Rows: Columns: Non-zeros: Status: Objective:

giapetto2 5 2 8 OPTIMAL z = 180 (MAXimum)

No. Row name St Activity Lower bound Upper bound Marginal ------ ------------ -- ------------- ------------- ------------- ------------1 z B 180 2 Fin_hours NU 100 100 1

Advanced MathProg and glpsol*

3 Carp_hours NU 4 Dem[soldier] B 5 Dem[train] B No. -----1 2

Column name -----------x[soldier] x[train]

80 20 60

38

80 40 6.02e+23

1

St Activity Lower bound Upper bound Marginal -- ------------- ------------- ------------- ------------B 20 0 B 60 0

Karush-Kuhn-Tucker optimality conditions: KKT.PE: max.abs.err. = 0.00e+00 on row 0 max.rel.err. = 0.00e+00 on row 0 High quality KKT.PB: max.abs.err. = 0.00e+00 on row 0 max.rel.err. = 0.00e+00 on row 0 High quality KKT.DE: max.abs.err. = 0.00e+00 on column 0 max.rel.err. = 0.00e+00 on column 0 High quality KKT.DB: max.abs.err. = 0.00e+00 on row 0 max.rel.err. = 0.00e+00 on row 0 High quality End of output

Note how the constraints and the decision variables are now named after the TOY set, which looks clean and organized.

Part II

Theory of Linear Programming

Chapter 4

Linear Algebra and Linear Systems

Most students should already be familiar with the topics discussed in this chapter. So, this chapter may be a bit redundant, but it will at least serve us as a place where we fix some notation.

4.1

Matrix Algebra

Matrices, Vectors, and Their Transposes 4.1.1 Definition. A matrix is an array of numbers. We say that A is an (m × n)-matrix if it has m rows and n columns: a11 a12 · · · a1n a21 a22 · · · a2n A = . .. .. . .. .. . . . am1 am2 · · ·

amn

Sometimes we write A = [aij ] where it is understood that the index i runs from 1 through m, and the index j runs from 1 through n. We also use the denotation A ∈ Rm×n to indicate that A is an (m × n)-matrix. 4.1.2 Example. A =

5 2 −3 6 0 0.4

is a (2 × 3)-matrix, or A ∈ R2×3 . E.g. a12 = 2 and a23 = 0.4; a32 does not exist.

4.1.3 Definition. The transpose A0 of a matrix A is obtained by changing

Matrix Algebra

41

its rows to columns, or vice versa: a0ij = aji . So, if A is an (m × n)-matrix a11 a21 .. .

··· ··· .. .

a1n a2n .. .

am1 am2 · · ·

amn

A =

a12 a22 .. .

then its transpose A0 is the (n × m)-matrix 0 a11 a012 · · · a01m 0 0 a0 21 a22 · · · a2m 0 A = . .. .. = .. .. . . . 0 0 0 an1 an2 · · · anm

a11 a12 .. .

,

a21 a22 .. .

··· ··· .. .

am1 am2 .. .

a1n a2n · · ·

amn

.

4.1.4 Example. If A = then

5 2 −3 6 0 0.4

,

5 6 0 . = 2 −3 0.4

A0 So, e.g., a012 = 6 = a21 .

4.1.5 Remark. If you transpose twice (or any even number of times), you are back where you started: A00 = (A0 ) 0 = A. 4.1.6 Definition. A vector is either an (n × 1)-matrix or a (1 × n)-matrix. (n × 1)-matrices are called column vectors and (1 × n)-matrices are called row vectors.

Matrix Algebra

42

4.1.7 Example. If x is the column vector 1 −0.5 x = −8 , 11 then x0 is the row vector x0 = [1 − 0.5 − 8 11].

4.1.8 Remark. We will always assume that vectors are column vectors. So, e.g., a 3-dimensional vector x will be x1 x2 , x3 and not [x1 x2 x3 ] . Matrix Sums, Scalar Products, and Matrix Products Matrix sum and scalar multiplication are defined component-wise: 4.1.9 Definition. Let a11 a12 · · · a21 a22 · · · A= . .. .. .. . . am1 am2 · · ·

a1n a2n .. . amn

and B =

Then the matrix sum A + B is defined as a11 + b11 a12 + b12 a21 + b21 a22 + b22 A+B = .. .. . .

b11 b21 .. .

b12 b22 .. .

··· ··· .. .

b1n b2n .. .

bm1 bm2 · · ·

bmn

··· ··· .. .

a1n + b1n a2n + b2n .. .

am1 + bm1 am2 + bm2 · · ·

amn + bmn

Let λ be a real number. Then the scalar multiplication λa11 λa12 · · · λa1n λa21 λa22 · · · λa2n λA = . .. .. .. .. . . . λam1 λam2 · · ·

λamn

.

.

λA is defined as .

Matrix Algebra

43

4.1.10 Example. Let 5 2 A= 33 20 Then A − 100I =

and I =

1 0 0 1

−95 2 33 −80

and let B be a (n × p)-matrix B =

b11 b21 .. .

b12 b22 .. .

amn

··· ··· .. .

b1p b2p .. .

bn1 bn2 · · ·

bnp

.

.

4.1.11 Definition. Let A be a (m × n)-matrix a11 a12 · · · a1n a21 a22 · · · a2n A = . .. .. .. .. . . . am1 am2 · · ·

,

.

Then the product matrix [cij ] = C = AB is the (m × p)-matrix defined by cij

=

n X

aik bkj .

k=1

4.1.12 Example. 1 5 3 −1 2 1 5 9 16 13 33 7 −4 2 5 = , 0 3 −1 21 −14 5 9 0 2 1 6 since, e.g., 9 = c11 = a11 b11 + a12 b21 + a13 b31 = 2×1 + 1×7 + 5×0.

Matrix Algebra

44

4.1.13 Remark. Note that while matrix sum is commutative: A+B = B+A, the matrix product is not: AB 6= BA. Otherwise the matrix algebra follows the rules of the classical algebra of the real numbers. So, e.g., (A + B)(C + D) = (A + B)C + (A + B)D = AC + BC + AD + BD = A(C + D) + B(C + D). Inverse Matrices 4.1.14 Definition. The identity matrix In is an (n × n)-matrix (a square matrix) with 1s on the diagonal and 0s elsewhere: 1 0 0 ··· 0 0 0 1 0 ··· 0 0 0 0 1 ··· 0 0 In = . . . . . . . ... ... .. .. .. 0 0 0 ··· 1 0 0 0 0 ··· 0 1 We shall usually write shortly I instead of In , since the dimension n of the matrix is usually obvious. 4.1.15 Definition. The inverse matrix A−1 of a square matrix A, if it exists, is such a matrix that A−1 A = I = AA−1 .

4.1.16 Example. Let A = Then A−1 =

1 2 1 3

3 −2 −1 1

.

.

The inverse matrix A−1 in Example 4.1.16 above was found by using the Gauss–Jordan method that we learn later in this lecture. For now, the reader is invited to check that A−1 satisfies the two criteria of an inverse matrix: A−1 A = I = AA−1 .

Matrix Algebra

45

4.1.17 Example. Let A =

1 1 0 0

.

This matrix has no inverse. Indeed, if the inverse A−1 existed then, e.g., the equation 1 Ax = 1 would have a solution in x = [x1 x2 ]0 : 1 −1 . x = A 1 But this is impossible since Ax =

x1 + x2 0

6=

1 1

no matter what x = [x1 x2 ]0 you choose.

Dot and Block Matrix Notations When we want to pick up rows or columns of a matrix A we use the dotnotation: 4.1.18 Definition. Let A be the matrix a11 a12 · · · a21 a22 · · · A = . .. .. .. . . am1 am2 · · ·

a1n a2n .. . amn

.

Then its ith row is the n-dimensional row vector ai• = [ai1 ai2 · · · ain ] . Similarly, A’s j th column is the m-dimensional column vector a1j a2j a•j = . . .. amj

Matrix Algebra

46

4.1.19 Example. If

A =

5 2 −3 6 0 0.4

,

then a2• = [6 0 0.4] and a•3 =

−3 0.4

.

4.1.20 Remark. In statistical literature Pn the dot-notation is used for summation: There ai• means the row-sum j=1 aij . Please don’t be confused about this. In this course the dot-notation does not mean summation! When we want to combine matrices we use the block notation: 4.1.21 Definition. Let A be a (m × n)-matrix a11 a12 · · · a1n a21 a22 · · · a2n A = . .. .. .. .. . . . and let B be a (m × k)-matrix B =

am1 am2 · · ·

amn

b11 b21 .. .

··· ··· .. .

b1k b2k .. .

bm1 bm2 · · ·

bmk

Then the block matrix [A B] is a11 a12 a21 a22 [A B] = . .. .. . am1 am2

b12 b22 .. .

,

.

the (m × (n + k))-matrix ··· ··· .. .

a1n a2n .. .

b11 b21 .. .

b12 b22 .. .

··· ··· .. .

b1k b2k .. .

···

amn bm1 bm2 · · ·

bmk

Similarly, if C is an (p × n)-matrix, then the block matrix

A C

.

is defined

Matrix Algebra

47

as

a11 a21 .. .

a12 a22 .. .

am1 am2 A = c11 c12 C c21 c22 .. .. . . cp1 cp2

4.1.22 Example. Let 5.1 2.1 A = 6.5 −0.5 , 0.1 10.5 Then

c=

20 30

··· ··· .. .

a1n a2n .. .

··· ··· ··· .. .

amn c1n c2n .. .

···

cpn

.

0 and 0 = 0 . 0

,

1 −20 −30 0 5.1 2.1 = 0 6.5 −0.5 . 0 0.1 10.5

1 −c0 0 A

Block matrices of the type that we had in Example 4.1.22 above appear later in this course when we solve LPs with the simplex method in lectures 6 and 7. 4.1.23 Example. By combining the dot and block notation we have: a11 a12 · · · a1n a1• a21 a22 · · · a2n a2• A = . = .. = [a•1 a•2 · · · a•n ] . .. .. .. .. . . . . am1 am2 · · · amn am•

Order Among Matrices 4.1.24 Definition. If A = [aij ] and B = [bij ] are both (m × n)-matrices and aij ≤ bij for all i and j then we write A ≤ B. In this course we use the partial order introduced in Definition 4.1.24 mainly in the form x ≥ b, which is then a short-hand for: xi ≥ bi for all i.

Solving Linear Systems

4.2

48

Solving Linear Systems

Matrices and linear systems are closely related. In this section we show how to solve a linear system by using the so-called Gauss–Jordan method. This is later important for us when we study the simplex method for solving LPs in lectures 6 and 7. Matrices and Linear Systems A linear system is the system of linear equations

(4.2.1)

a11 x1 a21 x1

+ +

a12 x2 a22 x2

+ ··· + ···

am1 x1 + am2 x2 + · · ·

+ +

a1n xn a2n xn

= = .. .

b1 , b2 ,

+ amn xn = bm .

Solving the linear system (4.2.1) means finding the variables x1 , x2 , . . . , xn that satisfy all the equations in (4.2.1) simultaneously. The connection between linear systems and matrices is obvious. Indeed, let a11 a12 · · · a1n x1 b1 a21 a22 · · · a2n x2 b2 A= . , x = , and b = .. , .. .. .. .. .. . . . . . am1 am2 · · ·

amn

xn

bm

then the linear system (4.2.1) may be rewritten as Ax = b. Elementary Row Operations We develop in the next subsection the Gauss–Jordan method for solving linear systems. Before studying it, we need to define the concept of an Elementary Row Operation (ERO). An ERO transforms a given matrix A into a new ˜ via one of the following operations: matrix A ˜ is obtained by multiplying a row of A by a non-zero number λ: ERO1 A λa•i

˜•i . a

˜ are the same as in A. All other rows of A ˜ is obtained by multiplying a row of A by a non-zero number λ (as ERO2 A in ERO1), and then adding some other row, multiplied by a non-zero number µ of A to that row: λa•i + µa•j

˜•i a

˜ are the same as in A. (i 6= j ). All other rows of A

Solving Linear Systems

49

˜ is obtained from A by interchanging two rows: ERO3 A a•i

˜•j a

a•j

˜•i . a

˜ are the same as in A. All other rows of A

4.2.2 Example. We want to solve the following linear system: x1 + x2 = 2, 2x1 + 4x2 = 7.

To solve Example 4.2.2 by using EROs we may proceed as follows: First we replace the second equation by −2(first equation)+second equation by using ERO2. We obtain the system x1 +

x2 = 2, 2x2 = 3.

Next we multiply the second equation by 1/2 by using ERO1. We obtain the system x1 + x2 = 2, x2 = 3/2. Finally, we use ERO2: We replace the first equation by −(second equation)+first equation. We obtain the system x1

= 1/2, x2 = 3/2.

We have solved Example 4.2.2: x1 = 1/2 and x2 = 3/2. Now, let us rewrite what we have just done by using augmented matrix notation. Denote 1 1 2 A= and b = , 2 4 7 and consider the augmented matrix [A | b] =

1 1 2 2 4 7

.

This is the matrix representation of the linear system of Example 4.2.2, and the three steps we did above to solve the system can be written in the matrix representation as 1 1 2 1 1 2 1 1 2 1 0 1/2 2 4 7 0 2 3 0 1 3/2 0 1 3/2

Matrices as Linear Functions*

50

The usefulness of EROs and the reason why the Gauss–Jordan methdod works come from the following theorem: ˜ is obtained from the ˜ b] 4.2.3 Theorem. Suppose the augmented matrix [A| augmented matrix [A|b] via series of EROs. Then the linear systems Ax = b ˜ are equivalent. ˜ =b and Ax We shall not prove Theorem 4.2.3. However, Example 4.2.2 above should make Theorem 4.2.3 at least plausible. Gauss–Jordan Method We have already used the Gauss–Jordan method in this course — look how we solved Example 4.2.2. Now we present the general Gauss–Jordan method for solving x from Ax = b as a three-step algorithm (steps 1 and 3 are easy; Step 2 is the tricky one): Step 1 Write the problem Ax = b in the augmented matrix form [A|b]. Step 2 Using EROs transform the first column of [A|b] into [1 0 · · · 0]0 . This may require the interchange of two rows. Then use EROs to transform the second column of [A|b] into [0 1 · · · 0]0 . Again, this may require the change of two rows. Continue in the same way using EROs to transform successive columns so that ith column has the ith element equal to 1 and all other elements equal to 0. Eventually you will have transformed all the columns of the matrix A, or you will reach a point where there are one or more rows of the form [00 |c] on the bottom of the matrix. ˜ denote the augmented matrix at this ˜ b] In either case, stop and let [A| point. ˜ and read the solutions x — or the ˜ =b Step 3 Write down the system Ax lack of solutions — from that. 4.2.4 Remark.* the Gauss–Jordan method can also be used to calculate inverse matrices: To calculate the inverse of A construct the augmented matrix [A|I] and transform it via EROs into [I|B]. Then B = A−1 .

4.3

Matrices as Linear Functions*

This section is not needed later in the course; it just explains a neat way of thinking about matrices. If you are not interested in thinking about matrices as functions you may omit this section.

Matrices as Linear Functions*

51

Function Algebra We only consider function between Euclidean spaces Rm although the definitions can be easily extended to any linear spaces. 4.3.1 Definition. Let f, g : Rn → Rm , and let α ∈ R. (a) The function sum f + g is defined pointwise as (f + g)(x) = f (x) + g(x). (b) The scalar multiplication αf is defined pointwise as (αf )(x) = αf (x).

4.3.2 Example. Let f : R2 → R map a point to its Euclidean distance from the origin and let g : R2 → R project the point to its nearest point in the x1 -axis. So, q f (x) = kxk = x21 + x22 , g(x) = x2 . Then (f + 2g)(x) = kxk + 2x2 = So, e.g., (f + 2g)

1 1

=

q x21 + x22 + 2x2 . √

2 + 2.

4.3.3 Definition. Let f : Rn → Rm and g : Rk → Rn . The composition f ◦ g is the function from Rk to Rm defined as (f ◦ g)(x) = f (g(x)). The idea of composition f ◦ g is: The mapping f ◦ g is what you get if you first map x through the mapping g into g(x), and then map the result g(x) through the mapping f into f (g(x)).

Matrices as Linear Functions*

52

4.3.4 Example. Let g : R2 → R2 reflect the point around the x1 -axis, and let f : R2 → R be the Euclidean distance of the point from the point [1 2]0 . So, x1 x1 g = , x2 −x2 p x1 f = (x1 − 1)2 + (x2 − 2)2 . x2 Then (f ◦ g)

x1 x2

x1 = f g x2 x1 = f −x2 p (x1 − 1)2 + (−x2 − 2)2 . =

So, e.g., (f ◦ g)

1 1

= 3.

4.3.5 Definition. The identity function idm : Rm → Rm is defined by idm (x) = x. Sometimes we write simply id instead of idm . The idea of an identity map id is: If you start from x, and map it through id, you stay put — id goes nowhere. 4.3.6 Definition. Let f : Rm → Rn . If there exists a function g : Rn → Rm such that g ◦ f = idm and f ◦ g = idn then f is invertible and g is its inverse function. We denote g = f −1 . The idea of an inverse function is: If you map x through f , then mapping the result f (x) through f −1 gets you back to the starting point x.

Matrices as Linear Functions*

53

4.3.7 Example. Let f : R → R be f (x) = x3 . Then f −1 (x) = Indeed, now 3 √ √ 3 x f (f −1 (x)) = f 3 x = = x,

√ 3

x.

and in the same way one can check that f −1 (f (x)) = x.

Matrix Algebra as Linear Function Algebra Recall the definition of a linear function: 4.3.8 Definition. A function f : Rn → Rm is linear if (4.3.9)

f (αx + βy) = αf (x) + βf (y)

for all x, y ∈ Rn and α, β ∈ R. The key connection between linear functions and matrices is the following theorem which says that for every linear function f there is a matrix A that defines it, and vice versa. 4.3.10 Theorem. Let f : Rn → Rm . The function f is linear if and only if there is a (m × n)-matrix A such that f (x) = Ax. Proof. The claim of Theorem 4.3.10 has two sides: (a) if a function f is defined by f (x) = Ax then it is linear, and (b) if a function f is linear, then there is a matrix A such that f (x) = Ax. Let us first prove the claim (a). We have to show that (4.3.9) holds for f . But this follows from the properties of the matrix multiplication. Indeed, f (αx + βy) = A(αx + βy) = A(αx) + A(βy) = αAx + βAy = αf (x) + βf (y). Let us then prove the claim (b). This is a bit more difficult than the claim (a), since mow we have to construct the matrix A from the function f . The trick is to write any vector x ∈ Rn as x =

n X i=1

xi ei ,

Matrices as Linear Functions*

54

where ei = [ei1 ei2 · · · ein ]0 is the ith coordinate vector: eij = 1 if i = j and 0 otherwise. Then, since f is linear, we have ! n n X X f (x) = f = xi ei xi f (ei ) . i=1

So, if we can write

n X

i=1

xi f (ei ) = Ax

i=1

for some matrix A we are done. But this is accomplished by defining the matrix A by its columns as a•i = f (ei ) . This finishes the proof of Theorem 4.3.10. Since a row vector [c1 c2 · · · cn ] is an (1×n)-matrix, we have the following corollary: 4.3.11 Corollary. A function f : Rn → R is linear if and only if there is a vector c ∈ Rn such that f (x) = c0 x. Finally, let us interpret the matrix operations as function operations. 4.3.12 Theorem. Let Ah denote the matrix that corresponds to a linear function h. Let f and g be linear functions, and let λ be a real number. Then Âă Aλf

= λAf ,

Af +g = Af + Ag , Af ◦g = Af Ag , Aid = I, Af −1

= A−1 f .

The proof of Theorem 4.3.12 is omitted.

Chapter 5

Linear Programs and Their Optima

The aim of this chapter is, one the one hand, to give a general picture of LPs, and, on the other hand, to prepare for the Simplex method introduced in Chapter 6. If there is a one thing the student should remember after reading this chapter, that would be: The optimal solution of an LP is found in one of the corners of the region of all feasible solutions. There are many theoretical results, i.e. theorems, in this lecture. The proofs of the theorems are collected in the last subsection, which is omitted in the course. This lecture is adapted from [2, Ch. 2].

5.1

Form of Linear Program

Linear Program as Optimization Problem Let us start by considering optimization in general. Optimization problems can be pretty diverse. The next definition is, for most practical purposes, general enough. 5.1.1 Definition. An optimization problem is: maximize (or minimize) the objective function z = f (x1 , . . . , xn ) subject to the constraints l1

≤

g1 (x1 , . . . , xn ) .. .

≤ u1

lm ≤ gm (x1 , . . . , xn ) ≤ um 5.1.2 Remark. In some optimization problems some of the lower bounds l1 , . . . , lm may be missing. In that case we may simply interpret the missing lower bounds to be −∞. Similarly, some of the upper bounds u1 , . . . , um may be missing, and in that case we may interpret the missing upper bounds to be +∞. Also, when one formulates an optimization problem, it may turn out

Form of Linear Program

56

that the lower or upper bounds depend on the decision variables x1 , . . . , xn . In that case one can remove this dependence easily by using the following transformation: li (x1 , . . . , xn ) ≤ gi (x1 , . . . , xn ) ≤ ui (x1 , . . . , xn ) 0 ≤ gi (x1 , . . . , xn ) − li (x1 , . . . , xn ) gi (x1 , . . . , xn ) − ui (x1 , . . . , xn ) ≤ 0 So, the constraint i becomes two constraints, neither of which has bounds that depend on the decision x1 , . . . , xn . The variables x1 , . . . , xn in Definition 5.1.1 are called decision variables: They are the ones the optimizer seeks for — together with the value of the objective function, or course. Indeed, solving the optimization problem 5.1.1 means: 1. Finding the optimal decision x∗1 , . . . , x∗n under which the objective z ∗ = f (x∗1 , . . . , x∗n ) is optimized — maximized or minimized, depending on the problem — among all possible decisions x1 , . . . , xn that satisfy the constraints of the problem. 2. Finding, under the constraints, the optimal value z ∗ = f (x∗1 , . . . , x∗n ) — maximum or minimum, depending on the problem — of the objective. It may look silly that we have split the solution criterion into two points, but sometimes it is possible to find the optimal value z ∗ = f (x∗1 , . . . , x∗n ) without finding the optimal decision x∗1 , . . . , x∗n — or vice versa. In this course, however, we shall not encounter this situation. 5.1.3 Example. Mr. K. wants to invest in two stocks, #1 and #2. The following parameters have been estimated statistically: r1 r2 σ1 σ2 ρ

= 10% is the return of stock #1, = 5% is the return of stock #2, = 4 is the standard deviation of stock #1, = 3 is the standard deviation of stock #2, = −0.5 is the correlation between the stocks #1 and #2.

Mr. K. wants to maximize the return of his portfolio while keeping the risk (measured as standard deviation) of the portfolio below 3.5. How should Mr. K. distribute his wealth between the two stocks?

Mr. K.’s problem in Example 5.1.3 is an optimization problem. Indeed, let w1 and w2 denote the portions of Mr. K.’s wealth put in stocks #1 and #2,

Form of Linear Program

57

respectively. So, w1 and w2 are the decision variables of this problem. Then Mr. K.’s objective function to be maximized is the total return of his portfolio: z = f (w1 , w2 ) = r1 w1 + r2 w2 = 10w1 + 5w2 . The constraints are: q g1 (w1 , w2 ) = σ12 w12 + 2ρσ1 σ2 w1 w2 + σ22 w2 q = 16w12 − 12w1 w2 + 9w22 ≤ 3.5, for the risk, g2 (w1 , w2 ) = w1 + w2 ≤ 1, for the total wealth to be invested, and — if short-selling is not allowed — then there are the sign constraints 0 ≤ g3 (w1 , w2 ) = w1 , 0 ≤ g4 (w1 , w2 ) = w2 . Let us then consider linear optimization problems. Let us denote x1 x = ... , l = xn

l1 .. , u = . lm

u1 g1 (x) .. , and g(x) = .. . . . um gm (x)

Then Definition 5.1.1 can be written in a compact form as: 5.1.4 Definition. A general optimization problem is to either maximize or minimize the objective function z = f (x) subject to the constraints l ≤ g(x) ≤ u. An optimization problem is a linear optimization problem — or a linear program — if the objective function f and the constraint function g are both linear. Now, any linear form h(x) can be written as Ax, where A is a matrix uniquely determined by the function h (if you want to see how to construct the matrix A from the function h, see part (b) of the proof of Theorem 4.3.10). So, we arrive at the following definition:

Form of Linear Program

58

5.1.5 Definition. A linear optimization problem, or a linear program (LP) is to either maximize or minimize the objective function z = c0 x subject to the constraints l ≤ Ax ≤ u, and to the sign constraints x ≥ 0. 5.1.6 Remark. The sign constraints x ≥ 0 in Definition 5.1.5 are somewhat particular (as opposed to general), and not in line with Definition 5.1.1. However, in practice the sign constraints are so prevalent that we make it a standing assumption. Mr. K.’s optimization problem in Example 5.1.3 was not an LP, since the constraint function g1 was not linear (everything else in Mr. K.’s problem was linear). Assumptions of Linear Programs Definition 5.1.5 is the mathematical description of an LP. As such it is complete and perfect, as is the nature of mathematical definitions. Definition 5.1.5 is also very Laconic and not directly related to the “real world”, as is also the nature of mathematical definitions. The list below explains the consequences — or assumptions, if you like — of Definition 5.1.5 for the non-Spartans living in the “real world”: Proportionality The contribution to the objective function from each decision variable is proportional to the value of the decision variable: If, say, decision variable x2 is increased by ∆ then the value of objective function is increased by c2 ∆. Similarly, the contribution of each decision variable in restrictions is also proportional to the value of the said variable. So, e.g., if you double the value of the decision variable x2 the resources consumed by that decision will also double. Additivity The contribution to the objective function for any variable is independent of the values of the other decision variables. For example, no matter what the value of x1 is increasing x2 to x2 + ∆ will increase the value of the objective function by c2 ∆. Similarly, the resources used by decision x2 will increase independently of the value of x1 . Divisibility It is assumed that the decision variables can take fractional values. For example x1 may be π . This assumption is in many practical cases not true, but a reasonably good approximation of the reality. In case this assumption is violated, we have an Integer Program (IP). We shall learn about IPs in Chapter 11.

Form of Linear Program

59

Certainty It is assumed that all the parameters of the program are known with certainty. For example, in Giapetto’s problem 3.3.1 it was assumed that the demands for soldiers and trains were known. This is certainly almost never the case in practice. Indeed, typically in practice one has to estimate the parameters of the LP statistically. We shall not talk about statistical estimation in this course. 5.1.7 Remark. Unlike proportionality, additivity, and divisibility, the certainty assumption is not particularly “linear”. Standard Form of Linear Programs The LP 5.1.5 can be represented in many equivalent forms. In this course we consider three forms: 1. standard form, 2. slack form, 3. canonical slack form. The standard form is good for theoretical considerations. The slack form and the canonical slack form are food (no typo here) for the Simplex algorithm. In this subsection we consider the standard form. The slack form and the canonical form will be introduced in Chapter 6 where we study the Simplex algorithm. 5.1.8 Remark. When you solve LPs with GLPK there is usually no need to transform them to standard, slack, or canonical forms: GLPK will internally transform the LP into any form that is suitable for it (which is probably some kind of a slack form as GLPK uses a revised Simplex algorithm). Also, note that there is no universal consensus on what is a “standard form”, or a “slack form”, or a “canonical slack form” of an LP. So, in different textbooks you are likely to find different definitions. Indeed, e.g. [4] calls the slack form a standard form. 5.1.9 Definition. A standard form LP is: max z = c0 x s.t. Ax ≤ b x ≥ 0 So, an LP is in standard form if: 1. It is a maximization problem 2. There are no lower bound constraints Any LP of Definition 5.1.5 can be transformed into a standard form LP of Definition 5.1.9 by using the following three-step algorithm:

Form of Linear Program

60

Step 1: Change into maximization If the LP is a minimization problem, change it to a maximization problem by multiplying the objective vector c by −1: min c0 x max −c0 x. Step 2: Remove double inequalities If there are both lower and upper bound in a single constraint, change that constraint into two constraints: li ≤ ai1 x1 + · · · + ain xn ≤ ui li ≤ ai1 x1 + · · · + ain xn . ai1 x1 + · · · + ain xn ≤ ui Step 3: Remove lower bounds If there is a lower bound constraint li , change it to an upper bound constraint by multiplying the corresponding inequality by −1: li ≤ ai1 x1 + · · · ain xn

−ai1 x1 − · · · − ain xn ≤ −li .

5.1.10 Example. Let us find the standard form of the LP min z = −2x1 + 3x2 s.t.

1 ≤

x1 + x2 2x1 − x2 2 ≤ 7x1 + x2 x1 , x2

≤ 9 ≤ 4 ≤ 100 ≥ 0

(1) (2) (3) (4)

Step 1: We turn the LP into a maximization problem, and get the objective max z = 2x1 − 3x2 . Step 2: We remove the double inequalities (1) and (3). From the constraint (1) we get the constraints 1 ≤ x1 + x2 x1 + x2 ≤ 9

(1.a) (1.b)

and from the constraint (3) we get the constraints 2 ≤ 7x1 + x2 7x1 + x2 ≤ 100

(3.a) (3.b)

Before going to Step 3 let us check the status of the LP now: max z = −2x1 + 3x2

Location of Linear Programs’ Optima s.t.

1 ≤

x1 x1 2x1 2 ≤ 7x1 7x1

+ x2 + x2 − x2 + x2 + x2 x1 , x2

61

≤ ≤

9 4

≤ 100 ≥ 0

(1.a) (1.b) (2) (3.a) (3.b) (4)

Step 3: We remove the lower bounds for the inequalities (1.a) and (3.a). We obtain the standard form max z = −2x1 + 3x2 s.t.

5.2

−x1 x1 2x1 −7x1 7x1

− x2 + x2 − x2 − x2 + x2 x1 , x2

≤ −1 ≤ 9 ≤ 4 ≤ −2 ≤ 100 ≥ 0

(1.a) (1.b) (2) (3.a) (3.b) (4)

Location of Linear Programs’ Optima

In this section we consider the region of the admissible decisions in an LP problem — the so-called feasible region. We also consider the location of the optimal decision of an LP, which must of course be in the feasible region. The main result — and a problem — to be remembered is: The optimal solution of an LP is found in one of the corners of the feasible region. The decision variables corresponding to the corners are called Basic Feasible Solutions (BFS). So, the problem is to find the best BFS.

Shape of Feasible Region Since we now know how to transform any LP into a standard form, we shall state LPs in their standard forms in the definitions. 5.2.1 Definition. The feasible region K of an LP max z = c0 x s.t. Ax ≤ b x ≥ 0 is the set of decisions x that satisfy the constraints Ax ≥ b and x ≥ 0: K

= {x ∈ Rn ; Ax ≤ b, x ≥ 0} .

Location of Linear Programs’ Optima

62

Note that the feasible region K is determined by the technology matrix A and the constraints b. The objective c has no effect on the feasible region. 5.2.2 Remark. In what follows we use the convention that LP and its feasible region are like in Definition 5.2.1, and that the said LP has n decision variables and m constraints, excluding the sign constraints. This means that c is an ndimensional column vector, A is an (n×m)-matrix, and b is an m-dimensional column vector. Theorem 5.2.3 below says that if you have two feasible solutions, and you draw a line segment between those two solutions, then every solution in that line segment is also feasible. 5.2.3 Theorem. The feasible region of an LP is convex: If x and y belong to the feasible region, then also αx + (1 − α)y belong to the feasible region for all α ∈ [0, 1]. 5.2.4 Remark.* Actually the feasible region has more structure than just convexity: It is a (closed) convex polytope. We shall not give a general definition of a convex polytope. If the (closed) convex polytope is bounded one can think it as the result of the following procedure: 1. Take some points p1 , p2 , . . . , pk to be corners of the convex polytope. These points belong to the convex polytope — they are its generators. 2. If some point, say q, is in a line segment connecting any two points of the convex polytope, then q must be included to the convex polytope also. This including procedure must be reiterated as new points are included into the set until the set there are no more new points to be included. For example, three points will generate a filled triangle as their convex polytope. The said three points will be the three corners of the polytope. Four points in a three-dimensional space will generate a filled (irregular) tetrahedron as their convex polytope with the said four points at its corners. Optima in Corners 5.2.5 Definition. Consider a feasible solution, or a decision, x of an LP. The constraint bi is active at the decision x if ai• x = bi . Constraint i being active at decision x means that the resource i is fully consumed, or utilized, with decision x. 5.2.6 Definition. A feasible solution, or decision, of an LP is Inner point if there are no active constraints at that decision, Boundary point if there is at least one active constraints at that decision,

Location of Linear Programs’ Optima

63

Corner point if there are at least n linearly independent active constraints at that decision. Corner points are also called Basic Feasible Solutions (BFS). Note that corner points are also boundary points, but not vice versa. 5.2.7 Remark. Linear independence means that the constraints are genuinely different. For example, the constraints 2x1 + 3x2 ≤ 2 x1 + x2 ≤ 4 are linearly independent, but the constraints 2x1 + 3x2 ≤ 2 6x1 + 9x2 ≤ 6 are not. The next picture illustrates Definition 5.2.6. In that picture: None of the constraints (1), (2), or (3) is active in the “Inner point”. In the “Boundary point” one of the constraints, viz. (1), is active. In the “Corner point” two (which is the number of the decision variables) of the constraints, viz. (1) and (3), are active.

Karush–Kuhn–Tucker Conditions*

64

x2 3 (2)

2

Boundary point

Corner point 1 Inner point (3) 0

0

1

2

(1) 3 x1

5.2.8 Theorem. Let x∗ be an optimal solution to an LP. Then x∗ is a boundary point of the feasible region. Theorem 5.2.8 can be refined considerably. Indeed, the next theorem tells us that in seeking the optimum we do not have to check the entire boundary — it is enough to check the corners! 5.2.9 Theorem. An optimal solution x∗ of an LP can be found — when it exists — in one of the corner points of the feasible region, i.e., an optimal solution is a BFS.

5.3

Karush–Kuhn–Tucker Conditions*

Sometimes one can make an educated guess about the optimal corner of an LP. In that case one asks if the guess is correct. The following Karush–Kuhn– Tucker theorem provides a way to check the correctness of one’s guess.

Proofs*

65

5.3.1 Theorem. Consider the LP max z = c0 x s.t. Ax ≤ b . x ≥ 0 Let x be a BFS of the LP. Suppose there are vectors s, u, v such that (i) (ii) (iii) (iv)

Ax + s = b, c = A0 v − u, u0 x + v0 s = 0, s, u, v ≥ 0.

Then x is an optimal solution to the LP. The vectors s, u, v in the Karush–Kuhn–Tucker theorem 5.3.1 have the following interpretation: s is the slack vector: si tells how much of the resource i is unused. If si = 0 then the constraint i is active, i.e., the resource i is completely used. This interpretation is obvious if you look condition (i) of Theorem 5.3.1. u is connected to the sign constraint x ≥ 0: if xi > 0 then the ith sign constraint is not active and ui = 0. v is connected to the resource constraints. If there is slack si > 0 in the resource i then vi = 0. 5.3.2 Remark. The KKT.PE, KKT.PB, KKT.DE, and KKT.DB in the glpsol’s report are related to the conditions (i), (ii), (iii), and (iv) of the Karush–Kuhn– Tucker theorem 5.3.1. If the Karush–Kuhn–Tucker conditions are satisfied the values in the glpsol’s Karush–Kuhn–Tucker section should all be zero.

5.4

Proofs*

Proof of Theorem 5.2.3. Let α ∈ [0, 1]. It is obvious that if x ≥ 0 and y ≥ 0, then also αx + (1 − α)y ≥ 0. So, it remains to show that if Ax ≥ b and Ay ≥ b then also A(αx + (1 − αy)) ≥ b. But this follows from basic matrix algebra: A(αx + (1 − α)y) = αAx + (1 − α)Ay ≥ αb + (1 − α)b = b.

Proofs*

66

Proof of Theorem 5.2.8. This is a proof by contradiction: Suppose there is an optimal point x∗ that is an inner point of the feasible region. Then, for a small enough r , all points that are not further away from x∗ than the distance r belong to the feasible region. In particular, the point r c w = x∗ + 2 kck will belong to the feasible region. Here kck denotes the Euclidean distance: v u n uX kck = t c2i , i=1

and thus c/kck is a unit-length vector pointing at the same direction as c. Now, at point w we get for the objective function f (x) = c0 x that (5.4.1) since

c0 w = c0 x∗ +

r c0 c 2 kck

r = c0 x∗ + kck > c0 x∗ , 2

c0 c = kck2 .

But inequality (5.4.1) is a contradiction, since x∗ was optimal. So the assumption that x∗ was an inner point must be wrong. Proof of Theorem 5.2.9. This proof requires rather deep knowledge of linear algebra, and of linear spaces, although the idea itself is not so complicated if you can visualize n-dimensional spaces. (Taking n = 3 should give you the idea.) Let x∗ be an optimal solution, and let z ∗ = c0 x∗ be the optimal value. We already know, by Theorem 5.2.8, that x∗ is in the boundary of the feasible region. So, at least one constraints is active. Let now V be the subspace of Rn spanned by the active constraints at point x∗ . Let k be the dimension of V . If k = n, then x∗ is a boundary point, and we are done. Suppose then that k < n. Then V is a proper subspace or Rn and any vector in Rn can be written as an orthogonal sum of a vector from the subspace V and a vector from the orthogonal complement V ⊥ . Let us write the vector c this way: c = cV + cV ⊥ . Next we show that c belongs to the subspace V , i.e., cV ⊥ = 0. Suppose the contrary: cV ⊥ 6= 0. This means that there is a small > 0 such that x+ = x∗ + cV ⊥ is a feasible solution. But now z + = c 0 x+ = c0 x∗ + c0 cV ⊥ = z ∗ + c0V cV ⊥ + c0V ⊥ cV ⊥

= z ∗ + kcV ⊥ k2 > z∗,

Proofs*

67

which is a contradiction, since z ∗ was the optimal value. ˜ = x∗ + αw Since k < n there is a non-zero point w in V ⊥ such that x is feasible when α > 0 is small enought, and not feasible when α > 0 is too ˜ large. Now, let α be just small enough for x to be feasible. Then at point x ˜ at least one more constraint will become active. So, the space V associated ˜ has at least the dimension k + 1. Moreover, the point x ˜ is at to the point x least as good as the point x∗ , since ˜ z˜ = c0 x = c0 (x∗ + αw) = z ∗ + αc0 w = z∗. (Here we used the fact that c belongs to V .) ˜ is “closer to a corner” than x∗ , since it has k + 1 active constraints. Now, x ˜ to be the new x∗ and repeating the procedure described above By taking x n − k − 1 times we will find an optimal solution in a corner. Proof of Theorem 5.3.1. Let y be some BFS of the LP. Theorem 5.3.1 is proved if we can show that c0 y ≤ c0 x. Now, since y is feasible there is t ≥ 0 such that Ay + t = b. Denote w = x − y . Then Aw = s − t, and cy = c0 (x + w) = c0 x + c0 w = c0 x + v0 Aw − u0 w = c 0 x + v 0 s − v 0 t − u0 w = c 0 x + v 0 s − v 0 t − u0 y + u0 x = c0 x − v0 t − u0 y ≤ c0 x. So, x was indeed optimal.

Chapter 6

Simplex Method

This chapter is the very hard core of this course! Here we learn how to solve LPs manually. You may think that it is useless to know such arcane things: We should use computers, you might say — especially since the practical LPs are so large that no-one really solves them manually. This criticism is valid. But, we do not study how to solve LPs manually in order to solve them manually in practical problems (although it is not a completely useless skill). We study how to solve them manually in order to understand them! This lecture is adapted from [2, Ch. 2] and [4, Ch. 4].

6.1

Towards Simplex Algorithm

Checking Corners Theorem 5.2.9 told us that the optimal solution of an LP is in one of the corners of the feasible region. So, it seems that we have a very simple algorithm for finding the optimum: Just check all the corners! And, indeed, this naïve approach works well with such petty examples that we have in this course. The problem with this naïve approach in practice is the so-called combinatorial curse, a.k.a. the curse of dimensionality: An LP with n decision variables and m constraints has n n! = m (n − m)!m! corners. Let us consider the curse of dimensionality more closely: Consider an LP with 30 decision variables and 15 constraints. This LP has 30 = 155,117,520 15 corners. Suppose you have a computer that checks 100 corners per second (this is pretty fast for today’s computers, and right-out impossible if you program

Towards Simplex Algorithm

69

with Java TM ). Then it would take almost three weeks for the computer to check all the 155,117,520 corners. You may think this is not a problem: Maybe three weeks is not such a long time, and a problem with 30 decision variables is way bigger than anything you would encounter in the real life anyway. Well, think again! Three weeks is a long time if you need to update your optimal solution in a changing environment, say, daily, and LPs with at 30 decision variables are actually rather small. Indeed, let us be a bit more realistic now: Consider a shop owner who has 200 different products in her stock (a rather small shop). Suppose the shop owner has 100 constraints (not unreasonable) and a supercomputer that checks 100 million corners per second (very optimistic, even if one does not program with Java TM ). Then checking all the corners to optimize the stock would take 6.89 × 1044 years. The author doubts that even the universe can wait that long! The bottom line is that checking all the corners will take too much time even with a fast computer and a good programmer. Simplex Idea The general idea of the Simplex algorithm is that you do not check all the corners. The following list explains the Simplex algorithm in a meta-level. We call the steps Meta-Steps since they are in such a general level that they are not immediately useful. In the same way the three Meta-Step algorithm could be called a Meta-Simplex algorithm. We shall see later how the Meta-Steps can be implemented in practice. Meta-Step 1 Start with some corner. Meta-Step 2 Check if the corner is optimal. If so, you have found the optimum, and the algorithm terminates. Otherwise go to the next Meta-Step. Meta-Step 3 Move to an adjacent corner. Of all the adjacent corners choose the best one. Go back to Meta-Step 2. One hopes that in moving around the corners one hits the optimal corner pretty soon so that one does not have to check all the corners. To use the meta-algorithm described above we have to: • identify the corners analytically, • know how to tell if a chosen corner is optimal, • know how to go to the best adjacent corner. Once the points raised above are solved we have a genuine algorithm. This algorithm is given in the next section. Before that we have to discuss how to prepare an LP before it can be used in the Simplex algorithm.

Towards Simplex Algorithm

70

Slack Forms Before we can use the Simplex algorithm we must transform the LP into a so-called canonical slack form. We start with the slack form. Here is an informal definition of the slack form: An LP is in slack form, if 1. It is a maximization problem. 2. The constraints are equalities, rather than inequalities. 3. The Right Hand Side (RHS) of each constraint is non-negative.

6.1.1 Example. max z = 4x1 + 8x2 s.t. x1 + 2x2 ≤ 500 x1 + x2 ≥ 100 x1 , x2 ≥ 0

(0) (1) (2) (3)

Let us transform the LP in Example 6.1.1 above into a slack form. This is a maximization problem already, so we do not have to touch the line (0). Line (1) is an inequality. We can transform it into an equality by adding an auxiliary non-negative slack (or surplus) variable s1 : We obtain the constraint x1 + 2x2 + s1 = 500

(10 )

and, since we assumed that the slack s1 was non-negative, we have the sign constraints x1 , x2 , s1 ≥ 0 (30 ) The interpretation of the slack variable s1 is that it tells how much of the resource (1) is unused. Let us then consider line (2). We see that the LP is not in standard form. We could change it into a standard form by multiplying the inequality (2) by −1. But that would make the RHS of (2) negative, which is not good. Instead we ad — or actually subtract — an auxiliary non-negative excess variable e2 to the inequality. We obtain the equality x1 + x2 − e2 = 100

(20 )

and the sign constraints x1 , x2 , s1 , e2 ≥ 0

(300 ).

Towards Simplex Algorithm

71

The interpretation of the excess is opposite to that of the slack: Excess e2 tells how much the minimal requirement (2) is, well, excessed. Now, the LP in 6.1.1 is transformed into a slack form: max z = 4x1 + 8x2 s.t. x1 + 2x2 + s1 = 500 x1 + x2 − e2 = 100 x1 , x2 , s1 , e2 ≥ 0

(0) (10 ) (20 ) (300 )

Solving this slack form with decisions x1 , x2 , s1 , e2 is equivalent to solving the original LP with decisions x1 , x2 . Here is the formal definition of the slack form: 6.1.2 Definition. An LP is in slack form if it is of the type max z = c0 x s.t. [A S]

x = b s x, s ≥ 0

where b ≥ 0. Here s is the vector of slacks/excesses and S is the diagonal matrix containing the coefficients of the slacks and the excesses: +1 for slack and −1 for excess. Here is an algorithm for transforming a standard form LP max z = c0 x s.t. Ax ≤ b x ≥ 0 into a slack form: Step 1: Add slacks If the bi in the constraint i is non-negative add a slack (or surplus): ai1 x1 + ai2 x2 + · · · + ain xn ≤ bi ai1 x1 + ai2 x2 + · · · + ain xn + si = bi . Step 2: Add excesses If the bi in the constraint i is negative change the direction of the inequality (thus making the RHS −bi non-negative), and add an excess: ai1 x1 + ai2 x2 + · · · + ain xn ≤ bi −ai1 x1 − ai2 x2 − · · · − ain xn − ei = −bi . Steps 1 and 2 must be done to each constraint.

Towards Simplex Algorithm

72

6.1.3 Remark. The index i of the slack/excess refers to resources. For example, if s3 = 2 it means that 2 units of the resource 3 is unused. 6.1.4 Remark. We have two kinds of auxiliary variables: slacks (or surpluses) and excesses. Mathematically there is no need to differentiate between them: Excess is just negative slack (or negative surplus). Indeed, in some textbooks slacks are used for both the surpluses and the excesses. However, making the sign difference makes the problem, and the solution, easier to interpret, especially since typically the all the coefficients of an LP are non-negative. Finally, let us give the definition of the canonical slack form. 6.1.5 Definition. A slack form LP is canonical slack form if each constraint equation has a unique variable with coefficient 1 that does not appear in any other constraint equation. 6.1.6 Remark. Note that the slack form we constructed in Example 6.1.1 is not a canonical one. This is basically due to “wrong sign” −1 of the excess variable. Indeed, if there were a slack instead of an excess in the constraint (2) we would have a canonical form: The slacks would be the unique variables with coefficient 1 that do not appear in any other constraint equation. We shall see in Chapter 7 how to transform the slack form of 6.1.1 into a canonical form by using the Big M method. In this lecture we shall have to confine ourselves to more simple problems. Basic Feasible Solutions, Basic Variables, and Non-Basic Variables There is still one more concept — or two, or three, depending on how you count — that we have to discuss before we can present the Simplex algorithm: That of Basic Variables (BV) and Non-Basic Variables (NBV). Basic variables (BV) and non-basic variables (NBV) are related to the corners, or the basic feasible solutions (BFS), of an underdetermined linear system. So, what we are discussing in this subsection is related to the MetaStep 1 of the Meta-Simplex algorithm. Before going into formal definitions let us consider the following problem: 6.1.7 Example. Leather Ltd. manufactures two types of belts: the regular model and the deluxe model. Each type requires 1 unit of leather. A regular belt requires 1 hour of skilled labor and deluxe belt requires 2 hours of of skilled labor. Each week 40 units of leather and 60 hours of skilled labor are available. Each regular belt contributes =C3 to profit, and each deluxe belt contributes =C4 to profit. Leather Ltd. wants to maximize its profit.

Towards Simplex Algorithm

73

Let us build the LP for Leather Ltd. First, we have to choose the decision variables. So, what is there for Leather Ltd. to decide? The number of products to produce! So, Leather Ltd. has the following decision variables: x1 = number of regular belts manufactured x2 = number of deluxe belts manufactured Second, we have to find the objective function. What is it that Leather Ltd. wants to optimize? The profit! What is the Leather Ltd.’s profit? Well, each regular belt contributes =C3 to the profit, and each deluxe belt contributes =C4 to the profit. Since we denoted the number of regular belts produced by x1 and the number of deluxe belts produced by x2 the profit to be maximized is z = 3x1 + 4x2 . Finally, we have to find the constraints. So, what are the restrictions Leather Ltd. has to satisfy in making the belts? There are two restrictions: available labor and available leather. Let us consider the leather restriction first. There are only 40 units of leather available, and producing one regular belt requires 1 unit of leather. So does producing one deluxe belt. So, the leather constraint is x1 + x2 ≤ 40. Let us then consider the labor constraint. There are only 60 hours of labor available. Each regular belt produced consumes 1 hour of labor and each deluxe belt produced consumes 2 hours of labor. So, the labor constraint is x1 + 2x2 ≤ 60. Putting what we have just obtained together Ltd. Here it is: max z = 3x1 + 4x2 s.t. x 1 + x2 x1 + 2x2 x1 , x2

we obtain the LP for Leather ≤ 40 ≤ 60 ≥ 0

Following the algorithm given after Definition 6.1.2 we can transform the LP above into a slack form. Here is what we get: max z = 3x1 + 4x2 s.t. x1 + x2 + s1 = 40 x1 + 2x2 + s2 = 60 x1 , x2 , s1 , s2 ≥ 0 Let us then solve this slack form by using the method of Brutus Forcius (108–44 BC), which corresponds to checking all the corners. The Brutus’s method is based on the following observation, listed here as Remark 6.1.8:

Towards Simplex Algorithm

74

6.1.8 Remark. Consider the constraints of an LP in slack form. This is a linear system with m equations and n + m unknowns: n actual decision variables and m slacks. Since n+m > m this linear system is underdetermined. In principle, to solve a system of m equations requires only m variables. The remaining n variables can be set to zero. So, according to Remark 6.1.8, we choose successively 2 = m of the 4 = n− m variables x1 , x2 , s1 , s2 to be our basic variables (BV) and set the remaining 2 = n variables to be zero (NBV) and solve the constraint system. If the solution turns out to be feasible (it may not be since we are omitting the nonnegativity constraints here) we check the value of the objective at this solution. Since we this way check all the BFSs of the system we must find the optimal value. The next table lists the results: BVs s1 , s2 x2 , s2 x2 , s1 x1 , s2 x1 , s1 x1 , x2

Linear system 0 + 0 + s1 = 40 0 + 0 + s2 = 60 0 + x2 + 0 = 40 0 + 2x2 + s2 = 60 0 + x2 + s 1 = 40 0 + 2x2 + 0 = 60 x1 + 0 + 0 = 40 x1 + 0 + s2 = 60 x1 + 0 + s1 = 40 x1 + 0 + 0 = 60 x1 + x2 + 0 = 40 x1 + 2x2 + 0 = 60

x1

x2

s1

s2 BFS

z

Pt

0

0

40

60

Yes

0

F

0

40

0

20

Yes

120

B

0

60

−20

0

No

–

D

40

0

0

−20

No

–

A

30

0

10

0

Yes

120

C

20

20

0

0

Yes

140

E

From this table we read that the decision x1 = 20,

x2 = 20,

s1 = 0,

s2 = 0

is optimal. The corresponding optimal value is z = =C140. So, we have solved Leather Ltd.’s problem. Note that both of the slacks, s1 and s2 , are NBV, i.e. zeros, at the optimal decision. This means that at the optimal solution the resources, leather and skilled labor, are fully utilized. This full utilization of the resources is not uncommon in LPs, but it is not always the case. Sometimes it may be optimal not to use all your resources. Next picture illustrates the situation. The centers of the red balls are the candidate BFSs (Pts in the previous table). Note that only the points B , C , E , and F are actual BFSs. The optimal BFS is the point E .

Simplex Algorithm

75

x2 60

50

40

30

A

C

E

20

10

0

6.2

F 0

10

20

30

B 40

50

D 60 x1

Simplex Algorithm

Simplex Steps Step 1: Transform the LP into canonical slack form Transforming LP into a slack form has been explained in the previous section. For now, let us just hope that the said slack form is also a canonical one. It will be if there are no excesses. If there are excesses then the slack form most likely will not be canonical — unless you are extremely lucky. From the canonical slack form we construct the first Simplex Tableau. The first Simplex tableau is the canonical slack form where • The 0th row represents the objective function as a 0th constraint as z − c0 x = 0. • The variables that have unique row with 1 as coefficient, and 0 as coefficient in all other rows, will be chosen to be the BVs. Typically, the slacks are chosen to be the BVs. In that case the decisions are

Simplex Algorithm

76

set to be zero, and thus the first Simplex tableau will be solved for the slacks. So, most typically, the slack form LP max z = s.t.

c1 x1 + · · · + a11 x1 + · · · + a21 x1 + · · · + .. .

cn xn a1n xn +s1 a2n xn +s2

am1 x1 + · · · + amn xn

= b1 = b2 .. .

· · · +sm = bm x1 , . . . , xn , s1 , . . . , sm ≥ 0

becomes max z s.t. z − c1 x1 − · · · − a11 x1 + · · · + a21 x1 + · · · + .. .

cn xn a1n xn +s1 a2n xn +s2

am1 x1 + · · · + amn xn

=0 = b1 = b2 .. .

· · · +sm = bm x1 , . . . , xn , s1 , . . . , sm ≥ 0

Since we have to keep track of the BVs, this form is then represented as the Simplex tableau Row 0 1 2 .. .

z 1 0 0 .. .

x1 −c1 a11 a21 .. .

··· ··· ··· ··· .. .

xn −cn a1n a2n .. .

s1 0 1 0 .. .

s2 0 0 1 .. .

··· ··· ··· ··· .. .

sm 0 0 0 .. .

BV z= s1 = s2 = .. .

RHS 0 b1 b2 .. .

m

0

am1

···

amn

0

0

···

1

sm =

bm

From this tableau one readily reads the BFS related to the BVs s1 , . . . , sm : [s1 · · · sm ]0 = [b1 · · · bm ]0 ; and [x1 · · · xn ]0 = [0 · · · 0]0 . Step 2: Check if the current BFS is optimal In the first Simplex tableau the BVs are s1 , . . . , sm , and a BFS related to this solution is x1 = 0, . . . , xn = 0, s1 = b1 , . . . , sm = bm . The value of the objective can be read from the 0th row: Row 0

z 1

x1 −c1

··· ···

xn −cn

s1 0

s2 0

··· ···

sm 0

BV z=

RHS 0

This solution is hardly optimal. Indeed, suppose that the coefficients ci are non-negative (as is usually the case). But now all the decisions xi related to the coefficients ci are zero, as they are NBVs. But then,

Simplex Algorithm

77

obviously increasing the value of any xi will increase the value of the objective z . Let us then consider the general case. Suppose that, after some steps, we have come up with a Simplex tableau with the 0th row Row 0

z 1

x1 d1

··· ···

xn dn

s1

dn+1

s2

dn+2

··· ···

sm

dn+m

BV z=

RHS z∗

where all the coefficients di are non-negative for all the NBVs. Then making any NBV a BV would decrease the value of the objective. So, the criterion for the optimality is: The Simplex tableau is optimal, if in the 0th row there are no negative coefficients in any NBVs. If the tableau is optimal the algorithm terminates, and the optimal value and decision can be read from the BV and RHS columns. Step 3: Determine the entering variable If the BFS is not optimal, we have to change the BVs. One of the NBVs will become a BV (entering), and one of the old BVs will become a NBV (leaving). The entering variable will be the one with smallest coefficient in the 0th row. Indeed, this way we increase the value of the objective z the most. Step 4: Determine the leaving variable In Step 3 we chose some variable to enter as a new BV. Now we have to make one of the old BVs to leave to be a NBV. Now each BV in a Simplex tableau is associated to some row. The leaving BV will the one associated to the row that wins the ratio test (the smallest value is the winner) RHS of row . Coefficient of entering varaible in row The idea of the ratio test is, that we shall increase the entering variable as much as possible. At some point the increasing of the entering variable will force one of the BVs to become zero. This BV will then leave. The ratio test picks up the row associated to the leaving variable. Step 5: Find a new BFS Now we have a new system of BVs. Next we have to solve the Simplex tableau in terms of the new BVs. This can be done by using the Gauss–Jordan method. Then we have a new Simplex tableau, and we go back to Step 2. 6.2.1 Remark. The Step 1 above corresponds to the Meta-Step 1. The Step 2 corresponds to the Meta-Step 2. The Steps 3–5 correspond to the Meta-Step 3.

Simplex Algorithm

78

Dakota Furniture’s Problem

6.2.2 Example. The Dakota Furniture Company manufactures desks, tables, and chairs. The manufacture of each type of furniture requires lumber and two types of skilled labor: finishing labor and carpentry labor. The amount of each resource needed to make each type of furniture is given in the table below: Resource Lumber Finishing hours Carpentry hours

Desk 8 units 4 hours 2 hours

Table 6 units 2 hours 1.5 hours

Chair 1 unit 1.5 hours 0.5 hours

At present, 48 units of lumber, 20 finishing hours, and 8 carpentry hours are available. A desk sells for =C60, a table for =C30, and a chair for =C20. Dakota believes that demand for desks and chairs is unlimited, but at most 5 tables can be sold. Since the available resources have already been purchased, Dakota wants to maximize total revenue.

As a modelling problem Dakota’s problem is very similar to Giapetto’s problem 3.3.1. After making some comparisons on how we modelled Giapetto’s problem we notice that we should define the decision variables as x1 = number of desks produced x2 = number of tables produced x3 = number of chairs produced and that Dakota should solve the following LP: max z = 60x1 s.t. 8x1 4x1 2x1

+ 30x2 + 6x2 + 2x2 + 1.5x2 x2

+ 20x3 + x3 ≤ 48 + 1.5x3 ≤ 20 + 0.5x3 ≤ 8 ≤ 5 x1 , x2 , x3 ≥ 0

Dakota Furniture’s Solution with Simplex

Simplex Algorithm

79

Step 1: We start by transforming the Dakota’s LP into a slack form. Since all the inequalities are of type ≤ we have no excesses, and consequently we obtain the canonical slack form max z = 60x1 s.t. 8x1 4x1 2x1

+ 30x2 + 6x2 + 2x2 + 1.5x2 x2

+ 20x3 + x3 + s1 + 1.5x3 + s2 + 0.5x3 + s3

+ s4 x1 , x2 , x3 , s1 , s2 , s3 , s4

= 48 = 20 = 8 = 5 ≥ 0

Taking s1 , s2 , s3 , s4 to be our first BVs our first Simplex tableau for Dakota is Row 0 1 2 3 4

z 1 0 0 0 0

x1 −60 8 4 2 0

x2 −30 6 2 1.5 1

x3 −20 1 1.5 0.5 0

s1 0 1 0 0 0

s2 0 0 1 0 0

s3 0 0 0 1 0

s4 0 0 0 0 1

BV z= s1 = s2 = s3 = s4 =

RHS 0 48 20 8 5

Step 2: We check if the current Simplex tableau is optimal. The 0th row is now Row 0

z 1

x1 −60

x2 −30

x3 −20

s1 0

s2 0

s3 0

s4 0

BV z=

RHS 0

We see that there is a NBV x1 with negative coefficient −60. So the first Simplex tableau is not optimal. (Well, one does not expect to get an optimal solution by slacking off!) Step 3: We determine the entering variable. Since x1 has the smallest coefficient in row 0, increasing x1 will allow the objective z to increase most. So, x1 will enter as a new BV. Step 4: We determine the leaving variable. The ratio test gives us Row Row Row Row

1 2 3 4

limit limit limit limit

in in in in

on on on on

x1 x1 x1 x1

= 48/8 = 6 = 20/4 = 5 = 8/2 = 4 = No limit, since xi ’s coefficient is non-positive

So, Row 3 wins the ratio test. Since s3 the the BV associated to row 3, s3 is no longer a BV. Step 5: Now we have new BVs: s1 , s2 , x1 , s4 (remember x1 replaced s3 ). This means we have the unsolved Simplex tableau

Simplex Algorithm

Row 0 1 2 3 4

z 1 0 0 0 0

x1 −60 8 4 2 0

80 x2 −30 6 2 1.5 1

x3 −20 1 1.5 0.5 0

s1 0 1 0 0 0

s2 0 0 1 0 0

s3 0 0 0 1 0

s4 0 0 0 0 1

BV z= s1 = s2 = x1 = s4 =

RHS 0 48 20 8 5

Now we have to solve this Simplex tableau in terms of the BVs. This means that each row must have coefficient 1 for its BV, and that BV must have coefficient 0 on the other rows. This can be done with EROs in the following way: ERO1: We create a coefficient of 1 for x1 in row 3 by multiplying row 3 by 0.5. Now we have the tableau Row 0 1 2 3 4

z 1 0 0 0 0

x1 −60 8 4 1 0

x2 −30 6 2 0.75 1

x3 −20 1 1.5 0.25 0

s1 0 1 0 0 0

s2 0 0 1 0 0

s3 0 0 0 0.5 0

s4 0 0 0 0 1

BV z= s1 = s2 = x1 = s4 =

RHS 0 48 20 4 5

ERO2: To create a 0 coefficient for x1 in row 0, we replace the row 0 with 60(row 3) + row 0. Now we have the tableau Row 0 1 2 3 4

z 1 0 0 0 0

x1 0 8 4 1 0

x2 15 6 2 0.75 1

x3 −5 1 1.5 0.25 0

s1 0 1 0 0 0

s2 0 0 1 0 0

s3 30 0 0 0.5 0

s4 0 0 0 0 1

BV z= s1 = s2 = x1 = s4 =

RHS 240 48 20 4 5

ERO2: To create a 0 coefficient for x1 in row 1, we replace row 1 with −8(row3) + row 1. Now we have the tableau Row 0 1 2 3 4

z 1 0 0 0 0

x1 0 0 4 1 0

x2 15 0 2 0.75 1

x3 −5 −1 1.5 0.25 0

s1 0 1 0 0 0

s2 0 0 1 0 0

s3 30 −4 0 0.5 0

s4 0 0 0 0 1

BV z= s1 = s2 = x1 = s4 =

RHS 240 16 20 4 5

ERO2: To create a 0 coefficient for x1 in row 2, we replace row 2 with −4(row 3) + row 2. Now we have the tableau

Simplex Algorithm

Row 0 1 2 3 4

z 1 0 0 0 0

x1 0 0 0 1 0

81 x2 15 0 −1 0.75 1

x3 −5 −1 0.5 0.25 0

s1 0 1 0 0 0

s2 0 0 1 0 0

s3 30 −4 −2 0.5 0

s4 0 0 0 0 1

BV z= s1 = s2 = x1 = s4 =

RHS 240 16 4 4 5

Now we see that this Simplex tableau is solved: Each of the BVs have coefficient 1 on their own rows and coefficient 0 in other rows. So, we go now back to Step 2. Step 2: We check if the Simplex tableau above is optimal. It is not, since the NBV x3 has negative coefficient on row 0. Step 3: We determine the entering variable. In this case it is obvious: x3 enters. Step 4: We determine the leaving variable. The ratio test gives us Row Row Row Row

1 2 3 4

limit limit limit limit

in in in in

on on on on

x3 x3 x3 x3

= No limit = 4/0.5 = 8 = 4/0.25 = 16 = No limit

So, row 2 wins the ratio test. Since s2 was the BV of row 2, s2 will leave and become a NBV. Step 5 : Now we have new BVs: s1 , x3 , x1 , s4 , since s2 was replaced with x3 in the previous step. So, we have the following unsolved Simplex tableau Row 0 1 2 3 4

z 1 0 0 0 0

x1 0 0 0 1 0

x2 15 0 −1 0.75 1

x3 −5 −1 0.5 0.25 0

s1 0 1 0 0 0

s2 0 0 1 0 0

s3 30 −4 −2 0.5 0

s4 0 0 0 0 1

BV z= s1 = x3 = x1 = s4 =

RHS 240 16 4 4 5

To solve this tableau we must invoke the Gauss–Jordan method again: ERO1: To create a coefficient of 1 for x3 in row 2, we multiply the row 2 by 2. Now we have the tableau Row 0 1 2 3 4

z 1 0 0 0 0

x1 0 0 0 1 0

x2 15 0 −2 0.75 1

x3 −5 −1 1 0.25 0

s1 0 1 0 0 0

s2 0 0 2 0 0

s3 30 −4 −4 0.5 0

s4 0 0 0 0 1

BV z= s1 = x3 = x1 = s4 =

RHS 240 16 8 4 5

Simplex Algorithm

82

ERO2: To create a coefficient 0 for x3 in row 0, we replace row 0 with 5(row 2) + row 0. Now we have the tableau Row 0 1 2 3 4

z 1 0 0 0 0

x1 0 0 0 1 0

x2 5 0 −2 0.75 1

x3 0 −1 1 0.25 0

s1 0 1 0 0 0

s2 10 0 2 0 0

s3 10 −4 −4 0.5 0

s4 0 0 0 0 1

BV z= s1 = x3 = x1 = s4 =

RHS 280 16 8 4 5

ERO2: To create a coefficient 0 for x3 in row 1, we replace row 1 with row 2 + row 1. Now we have the tableau Row 0 1 2 3 4

z 1 0 0 0 0

x1 0 0 0 1 0

x2 5 −2 −2 0.75 1

x3 0 0 1 0.25 0

s1 0 1 0 0 0

s2 10 2 2 0 0

s3 10 −8 −4 0.5 0

s4 0 0 0 0 1

BV z= s1 = x3 = x1 = s4 =

RHS 280 24 8 4 5

ERO2: To create a coefficient 0 for x3 in row 3, we replace row 3 with −0.25(row 3) + row 3. Now we have the tableau Row 0 1 2 3 4

z 1 0 0 0 0

x1 0 0 0 1 0

x2 5 −2 −2 1.25 1

x3 0 0 1 0 0

s1 0 1 0 0 0

s2 10 2 2 −0.5 0

s3 10 −8 −4 1.5 0

s4 0 0 0 0 1

BV z= s1 = x3 = x1 = s4 =

RHS 280 24 8 2 5

Now we see that this Simplex tableau is solved: Each of the BVs have coefficient 1 on their own rows and coefficient 0 in other rows. So, we go now back to Step 2. Step 2: We check if the Simplex tableau is optimal. We see that it is! Indeed, all the NBVs x2 , s2 , s3 have non-negative coefficients in row 0. Finally, let us interpret the result: The number of desks, tables, and chairs Dakota Furniture should manufacture is 2, 0, and 8. With this decision Dakota’s revenue is =C280. Of the resources: 24 units of lumber is left unused: s1 = 24. All the other actual resources are fully used: s2 = 0, s3 = 0, but the market demand for tables is not used at all s5 = 5, since no tables are manufactured.

Simplex Algorithm

83

Dakota Furniture’s Solution with glpsol We show briefly how to solve the Dakota’s problem of Example 6.2.2 with glpsol. Here are the contents of the file dakota.mod, where the Dakota’s problem is described in GNU MathProg modelling language. There are no slacks here in the code: GLPK will do the transformations for you internally. # # Dakota’s problem # # This finds the optimal solution for maximizing Dakota’s revenue # /* Decision variables */ var x1 >=0; /* desk */ var x2 >=0; /* table */ var x3 >=0; /* chair */ /* Objective function */ maximize z: 60*x1 + 30*x2 + 20*x3; /* Constraints */ s.t. Lumber : 8*x1 + 6*x2 + x3 <= 48; s.t. Finishing : 4*x1 + 2*x2 + 1.5*x3 <= 20; s.t. Carpentry : 2*x1 + 1.5*x2 + 0.5*x3 <= 8; s.t. Demand : x2 <= 40; end;

So, issue the command glpsol -m dakota.mod -o dakota.sol Now, you should get in your console something like the following: Reading model section from dakota.mod... 21 lines were read Generating z... Generating Lumber... Generating Finishing... Generating Carpentry... Generating Demand... Model has been successfully generated glp_simplex: original LP has 5 rows, 3 columns, 13 non-zeros glp_simplex: presolved LP has 3 rows, 3 columns, 9 non-zeros lpx_adv_basis: size of triangular part = 3 * 0: objval = 0.000000000e+00 infeas = 0.000000000e+00 (0) * 2: objval = 2.800000000e+02 infeas = 0.000000000e+00 (0) OPTIMAL SOLUTION FOUND Time used: 0.0 secs Memory used: 0.1 Mb (114563 bytes) lpx_print_sol: writing LP problem solution to ‘dakota.sol’...

Simplex Algorithm

84

The file dakota.sol should now contain the following report: Problem: Rows: Columns: Non-zeros: Status: Objective:

dakota 5 3 13 OPTIMAL z = 280 (MAXimum)

No. -----1 2 3 4 5

Row name -----------z Lumber Finishing Carpentry Demand

St Activity Lower bound Upper bound Marginal -- ------------- ------------- ------------- ------------B 280 B 24 48 NU 20 20 10 NU 8 8 10 B 0 40

No. -----1 2 3

Column name -----------x1 x2 x3

St Activity Lower bound Upper bound Marginal -- ------------- ------------- ------------- ------------B 2 0 NL 0 0 -5 B 8 0

Karush-Kuhn-Tucker optimality conditions: KKT.PE: max.abs.err. = 7.11e-15 on row 1 max.rel.err. = 7.11e-17 on row 2 High quality KKT.PB: max.abs.err. = 0.00e+00 on row 0 max.rel.err. = 0.00e+00 on row 0 High quality KKT.DE: max.abs.err. = 3.55e-15 on column 2 max.rel.err. = 9.87e-17 on column 2 High quality KKT.DB: max.abs.err. = 0.00e+00 on row 0 max.rel.err. = 0.00e+00 on row 0 High quality End of output

You should compare this output with the last, optimal, Simplex tableau. Indeed, recall that the optimal Simplex tableau was Row 0 1 2 3 4

z 1 0 0 0 0

x1 0 0 0 1 0

x2 5 −2 −2 1.25 1

x3 0 0 1 0 0

s1 0 1 0 0 0

s2 10 2 2 −0.5 0

s3 10 −8 −4 1.5 0

s4 0 0 0 0 1

BV z= s1 = x3 = x1 = s4 =

RHS 280 24 8 2 5

Simplex Algorithm

85

where x1 = Number of desks produced x2 = Number of tables produced x3 = Number of chairs produced s1 = Amount of lumber unused s2 = Number of finishing hours unused s3 = Number of carpentry hours unused s4 = Demand for tables unused Now, compare this to the following part of the glpsol’s report No. -----1 2 3 4 5

Row name -----------z Lumber Finishing Carpentry Demand

St Activity Lower bound Upper bound Marginal -- ------------- ------------- ------------- ------------B 280 B 24 48 NU 20 20 10 NU 8 8 10 B 0 40

No. -----1 2 3

Column name -----------x1 x2 x3

St Activity Lower bound Upper bound Marginal -- ------------- ------------- ------------- ------------B 2 0 NL 0 0 -5 B 8 0

We will talk more about the connection later in Chapter 8 where we talk about sensitivity analysis. Now we just make some remarks. First, note the connections x1 = x1 x2 = x2 x3 = x3 s1 = Lumber (unused) s2 = Finishing (unused) s3 = Carpentry (unused) s4 = Demand (unused) Whenever a variable is BV in the optimal Simplex tableau glpsol reports its status in the St column as bounded by the symbol B. So, you see that the status for x1, x3, z, Lumber, and Demand are bounded, since they are BVs. The NBVs glpsol will mark as NU or NL. The Marginal column in glpsol’s report is related to the 0th row in the glpsol’s report. The marginals, or the shadow prices, are very important in sensitivity analysis. They tell how much the value of the optimal solution will change if the corresponding constraint is relaxed

Simplex Algorithm

86

by one unit. (The negative marginal with x2 is a slightly different story, since x2 is not a constraint, but a decision variable.)

Chapter 7

More on Simplex Method

This chapter is adapted from [2, Ch. 2] and [4, Ch. 4].

7.1

Big M Algorithm

Problem with Canonical Slack Form Recall that the Simplex algorithm requires a starting BFS, i.e., the first Simplex tableau must be solved with some BVs. So far we have not had any excesses, and consequently the first Simplex tableau was solved with the slack variables as BVs. If an LP has inequalities of type ≥ with non-negative RHS, or equalities, then finding the starting BFS can be difficult, or even right out impossible. The following example illustrates this point.

7.1.1 Example. Bevco manufactures an orange-flavored soft drink called Oranj by combining orange soda and orange juice. Each unit of orange soda contains 0.5 units of sugar and 1 mg of vitamin C. Each unit of orange juice contains 0.25 units of sugar and 3 mg of vitamin C. It costs Bevco =C0.02 to produce a unit of orange soda and =C0.03 to produce a unit of orange juice. Bevco’s marketing department has decided that each each bottle of Oranj must be of size 10 units, and must contain at least 20 mg of vitamin C and at most 4 units of sugar. How can the marketing department’s requirements be met at minimum cost?

Let us construct the LP for Bevco. We have had many examples already that we have modelled as LPs. Remember, e.g., Giapetto 3.3.1, Leather Ltd. 6.1.7, and Dakota 6.2.2. In all those problems we had to decide how many of each products we should produce in

Big M Algorithm

88

order to maximize the profit. Bevco’s problem is different. For starters, there is only one product. So, what is the decision Bevco must make? The only thing that is actually not fixed in Example 7.1.1, and thus open for decision, is the actual mixture of orange soda and orange juice in a bottle of Oranj. So, the decision variables are: x1 = Number of units of orange soda in a bottle of Oranj x2 = Number of units of orange juice in a bottle of Oranj What about the objective then? What does Bevco want to maximize or minimize? We see that Bevco wants to minimize the cost of producing Oranj. So this is a minimization problem. So, what is the cost of producing one bottle of Oranj? Well, recall that it costs =C0.02 to produce one unit of orange soda, and =C0.03 to produce one unit of orange juice. So, the objective function to be minimized is z = 2x1 + 3x2 (here we measure in Cents rather than in Euros in order to avoid fractions). So far we have found out the decision variables and the objective. What about the constraints then? Remember the marketing department’s demands: Each bottle of Oranj must contain at least 20 mg of vitamin C and at most 4 units of sugar. Let us consider first the sugar demand. Each unit of orange soda has 0.5 units of sugar in it and each unit of orange juice has 0.25 units of sugar in it. Since a bottle of Oranj must not contain more than 4 units of sugar we obtain the constraint 0.5x1 + 0.25x2 ≤ 4 (here the fractions cannot be avoided — at least not easily). The vitamin C constraint is similar to the sugar constraint, but opposite. Indeed, each unit of orange soda contains 1 mg of vitamin C, and each unit of orange juice contains 3 mg of vitamin C. Since there must be at least 20 mg of vitamin C in a bottle of Oranj, we obtain x1 + 3x2 ≥ 20. There is still one more constraint. This constraint is not so obvious as the sugar and vitamin C constraints, but it must be included. Recall that each bottle of Oranj must be of size 10 units. So, as the bottle of Oranj only contains orange soda and orange juice, we must have x1 + x2 = 10.

Big M Algorithm

89

Finally, note the classical sign constraints x1 , x2 ≥ 0. Indeed, it would be pretty hard to put negative amount of either orange soda or orange juice in a bottle. We have found the LP for Bevco: min z = 2x1 s.t. 0.5x1 x1 x1

+ 3x2 + 0.25x2 + 3x2 + x2 x1 , x2

≤ 4 ≥ 20 = 10 ≥ 0

(sugar constraint) (vitamin C constraint) (10 units in a bottle of Oranj)

Let us then try to apply the Simplex method to this LP. First we turn the LP into a slack form (note that this is a minimization problem). We obtain max −z s.t. −z + (7.1.2)

2x1 0.5x1 x1 x1

+ 3x2 + 0.25x2 + s1 + 3x2 − e2 + x2 x1 , x2 , s1 , e2

= 0 = 4 = 20 = 10 ≥ 0

(0) (1) (2) (3) (4)

The problem now is that the slack form above is not in canonical form. We have three genuine constraint equations (1)–(3), but there are no three variables in the set x1 , x2 , s1 , e2 that could be taken as BVs under which the constraint system above could be solved while keeping the RHSs still nonnegative. 7.1.3 Remark. There are two reasons why Bevco’s slack form turned out to be non-canonical. One is the vitamin C constraint x1 + 3x2 ≥ 20. This is a lower-bound constraint that will give us an excess variable — and excess variables have “wrong” signs. The other is the equality constraint x1 + x2 = 10. This constraint does not give us any slacks or excesses. So, we fall short of variables. Solution with Artificial Variables The problem with the system (1)–(3) is that we do not have enough variables. So, the solution is obvious: We introduce new artificial variables where needed.

Big M Algorithm

90

Now, row (1) in (7.1.2) is fine: It has s1 . Rows (2) and (3) of (7.1.2), on the other hand, are not fine: They lack variables with coefficient 1 that do not appear in any other row. So, we introduce artificial variables: a2 will enter row (2) and a3 will enter row (3). We get the system max −z s.t. −z + (7.1.4)

2x1 0.5x1 x1 x1

+ 3x2 + 0.25x2 +s1 + 3x2 −e2 +a2 + x2 +a3 x1 , x2 , s1 , e2 , a2 , a3

= 0 (0) = 4 (1) = 20 (2) = 10 (3) ≥ 0 (4)

Now we have a BFS. Indeed, taking s1 , a2 , a3 to be the BVs we obtain z = 0,

s1 = 4,

a2 = 20,

a3 = 10.

But now we have a small problem. What would guarantee that an optimal solution to (7.1.4) is also an optimal solution to (7.1.2)? Well, actually, nothing! Indeed, we might find an optimal solution to (7.1.4) where some of the artificial variables are strictly positive. In such case it may turn out that the corresponding solution to (7.1.2) is not even feasible. For example, in (7.1.4) it can be shown that the solution z = 0,

s1 = 4,

a2 = 20,

a3 = 10,

x1 = 0,

x2 = 0

is optimal. But this solution is not feasible for the original problem. Indeed, this solution contains no vitamin C, and puts 0 units of soda and juice in a bottle. So, this solution cannot possibly be optimal in the original problem, as it is not even a solution. The critical point is: In the optimal solution all the artificial variables must be zero. How is this achieved? By changing the objective function! Recall that the original objective function for Bevco was (in max form) max −z = −2x1 − 3x2 . Now, let M be a very very very very very very very very very very very big number — if you approve that 0 × ∞ = 0, you may think that M = +∞. Consider then the objective function max −z = −2x1 − 3x2 − M a2 − M a3 . Now allowing a2 or a3 be strictly positive should penalize the value of −z so much that the solution could not possibly be optimal. This means that an

Big M Algorithm

91

optimal solution to the system, (7.1.5) max −z s.t. −z + 2x1 + 3x2 +M a2 +M a3 0.5x1 + 0.25x2 +s1 x1 + 3x2 −e2 +a2 x1 + x2 +a3 x1 , x2 , s1 , e2 , a2 , a3

= 0 (0) = 4 (1) = 20 (2) = 10 (3) ≥ 0 (4)

should have a2 = 0 and a3 = 0. But then an optimal solution of (7.1.5) is also an optimal solution to the original problem (7.1.2) of Example 7.1.1. The system (7.1.5) is not yet in canonical slack form. There is one more trick left. To solve (7.1.5) in terms of the prospective BVs s1 , a2 , a3 we must remove a2 and a3 from row 0. This is done by using the ERO2 (two times): Replace row 0 with row 0 − M (row 2) − M (row 3). This way we obtain the system (7.1.6) max −z s.t. −z + (2−2M )x1 + (3−4M )x2 +M e2 = −30M (0) 0.5x1 + 0.25x2 +s1 = 4 (1) x1 + 3x2 −e2 +a2 = 20 (2) x1 + x2 +a3 = 10 (3) x1 , x2 , s1 , e2 , a2 , a3 ≥ 0 (4) This is a canonical slack form. The BVs are s1 , a2 , a3 . Now we have a canonical slack form (7.1.6). So, we can carry out the Simplex algorithm. The next steps are the Simplex algorithm steps. Step 1: Our first Simplex — or Simplex/Big M — tableau is the system (7.1.6): Row 0 1 2 3

−z 1 0 0 0

x1 2 − 2M 0.5 1 1

x2 3 − 4M 0.25 3 1

s1 0 1 0 0

e2 M 0 −1 0

a2 0 0 1 0

a3 0 0 0 1

BV −z = s1 = a2 = a3 =

RHS −30M 4 20 10

Step 2: We check for optimality. Our Simplex tableau is not optimal since there are negative coefficients in row 0 for the NBVs x1 and x2 (remember that M is a very very very very big number). Step 3: We determine the entering variable. Now, when M is big enough — and M is always big enough — we have that 3 − 4M

≤ 2 − 2M.

Big M Algorithm

92

So, x2 will enter as a new BV. Step 4: We determine the leaving variable. The ratio tests give us Row 1 limit in on x2 = 4/0.25 = 16 Row 2 limit in on x2 = 20/3 = 6.667 Row 3 limit in on x2 = 10/1 = 10 So, Row 2 wins the ratio test. Since a2 the the BV associated to row 2, a2 is no longer a BV. Step 5: Now we have new BVs: s1 , x2 , a3 (remember x2 replaced a2 ). This means we have the unsolved Simplex tableau Row 0 1 2 3

−z 1 0 0 0

x1 2 − 2M 0.5 1 1

x2 3 − 4M 0.25 3 1

s1 0 1 0 0

e2 M 0 −1 0

a2 0 0 1 0

a3 0 0 0 1

BV −z = s1 = x2 = a3 =

RHS −30M 4 20 10

We have to solve this tableau in terms of the BVs s1 , x2 , a3 by using the Gauss– Jordan method. Applying the Gauss–Jordan method here is a bit tricky since we have the symbol M . So, we cannot just count with numbers — we have to do some algebra. Let us start. First we eliminate x2 from row 0. As a first step to that direction we use ERO1 and multiply row 2 by 1/3. We obtain the tableau Row 0 1 2 3

−z 1 0 0 0

x1 2 − 2M 0.5 1 3

1

x2 3 − 4M 0.25 1 1

s1 0 1 0 0

e2 M 0

a2 0 0

0

0

−1 3

1 3

a3 0 0 0 1

BV −z = s1 = x2 = a3 =

RHS −30M 4 20 3

10

Next, we eliminate x2 from row 0. This is done by adding (4M −3)row 2 to to row 0. Our new Simplex tableau is then Row 0 1 2 3

−z 1 0 0 0

x1 3−2M 3

0.5 1 3

1

x2 0 0.25 1 1

s1 0 1 0 0

e2 3−M 3

a2 4M −3 3

−1 3

1 3

0

0

0

0

a3 0 0 0 1

BV −z = s1 = x2 = a3 =

RHS −60−10M 3

4

20 3

10

Next two steps are to eliminate x1 from the rows 1 and 3. We omit the details, and just state the solved Simplex tableau: Row 0 1 2 3

−z 1 0 0 0

x1 3−2M 3 5 12 1 3 2 3

x2 0 0 1 0

s1 0 1 0 0

e2 3−M 3 1 12 −1 3 1 3

a2 4M −3 3 −1 12 1 3 −1 3

a3 0 0 0 1

BV −z = s1 = x2 = a3 =

RHS −60−10M 3 7 3 20 3 10 3

Big M Algorithm

93

Step 2: We check for optimality. The tableau above is not optimal: The NBV x1 has negative coefficient. Step 3: WE determine the entering variable. The NBV x1 has the smallest coefficient among all NBVs, so it enters. Step 4: We determine the leaving variable. The . 5 Row 1 limit in on x1 = 73 12 . 1 Row 2 limit in on x1 = 20 3 .3 2 Row 3 limit in on x1 = 10 3 3

ratio test gives us = 5.6 = 20 = 5

So, Row 3 wins the ratio test. Since a3 the BV associated to row 2, a3 is no longer a BV. Step 5: Now we have new BVs: s1 , x2 , x1 , and a new Simplex tableau we have to solve with the Gauss–Jordan method. We omit the cumbersome details. Here is the solved tableau: Row 0 1 2 3

−z 1 0 0 0

x1 0 0 0 1

x2 0 0 1 0

s1 0 1 0 0

e2

1 2 −1 8 −1 2 1 2

a2 2M −1 2 1 8 1 2 −1 2

a3 2M −3 2 −5 8 −1 2 3 2

BV −z = s1 = x2 = x1 =

RHS −25 1 4

5 5

Step 2: We check for optimality. We see that the tableau is optimal! Indeed, there are no strictly negative coefficients in the 0th row for NBVs. So, the solution for Bevco is to put equal amount — 5 units and 5 units — of orange soda and orange juice in a bottle of Oranj. Then the production cost of a bottle is minimized, and it is =C0.25 (remember we counted in Cents). At the optimal solution s1 = 0.25, which means that there is only 3.75 units sugar in a bottle of Oranj (remember that the maximal allowed sugar content was 4 units). The excess variable e2 = 0. This means that there is exactly 20 mg of vitamin C in the bottle of Oranj. Finally, note that the artificial variables a2 and a3 are both zero, as they should be. 7.1.7 Remark. It may turn out that in solving the Big M analog of an LP the optimal solution has non-zero artificial variables. If this happens, it means that the original LP does not have any feasible solutions.

Simplex Algorithm with Non-Unique Optima

94

Big M Steps Let us formalize, or algorihmize, the Big M method we explained by an example in the previous subsection. Step 1: Start with a slack form The slack form is constructed in the same way as in the plain Simplex case. Remember that the RHSs must be nonnegative, and that the problem must be a maximization problem. Step 2: Add artificial variables To each constraint row, say i, that does not have a slack variable si , add an artificial variable ai . For each artificial variable ai subtract the value M ai from the objective function z . This means, for each of the artificial variables ai , adding the value M ai in the 0th row in the column corresponding the BV ai . Step 3: Construct the first Simplex tableau Solve the system you got in Step 2 in terms of the slacks and the artificial variables. This is done by subtracting from the 0th row M times each row that has an artificial variable. Now, the slacks and the artificial variables will be the BVs, and the system is solved in terms of the BVs. Otherwise the first Simplex tableau is the same as in the plain Simplex case. Step 4: Carry out the Simplex algorithm In Step 3 we constructed the first Simplex tableau. Now this tableau is solved for the slacks and artificial variables, and we may start the Simplex algorithm. Find the optimal BFS of the system by using the Simplex algorithm. Step 5: Check feasibility Check that in the optimal solution obtained by the Simplex algorithm the artificial variables are zero. If this is the case we have a solution. If at least one of the artificial variables is strictly positive the original LP does not have feasible solutions.

7.2

Simplex Algorithm with Non-Unique Optima

All the LP examples we have had so far had a unique optimal point (probably, I haven’t really checked that properly). In this subsection we discuss what happens with the Simplex — or the Simplex/Big M — algorithm if there are many optima, unbounded optimum, or no optima at all. In this section we shall give all the examples simply as LPs, without any associated story.

Simplex Algorithm with Non-Unique Optima

95

Many Bounded Optima

7.2.1 Example. Consider the LP min z = −6x1 − 2x2 s.t. 2x1 + 4x2 ≤ 9 3x1 + x2 ≤ 6 x1 , x2 ≥ 0

(1) (2) (3)

This is a minimization problem. So, before we do anything else, we transform it into a maximization problem. Here is the transformed problem:

(7.2.2)

max z = 6x1 + 2x2 s.t. 2x1 + 4x2 ≤ 9 3x1 + x2 ≤ 6 x1 , x2 ≥ 0

(1) (2) (3)

Here is the graphical representation of the LP of Example 7.2.1 (or actually, of the transformed problem (7.2.2) where we have switched z to −z ). The isoprofit lines are dashed.

Simplex Algorithm with Non-Unique Optima

96

x2 4

(2)

3

D 2 Optimum

C 1

Feasible region 0 A0

1

(1) 2B

3

4 x1

From the picture it is clear that the are many optimal points. Indeed, all the points in the line segment from B to C are optimal — and only the points in the line segment from B to C are optimal. Let us see now what the Simplex algorithm says about this. First, we transform the LP of Example 7.2.1 — actually the LP (7.2.2) — into a slack form. We obtain max z s.t. z − 6x1 − 2x2 2x1 + 4x2 + s1 3x1 + x2 + s2 x1 , x2 , s1 , s2

= = = ≥

0 9 6 0

We see that this slack form is canonical, and we get the first Simplex tableau without resorting to the Big M method:

Simplex Algorithm with Non-Unique Optima

Row 0 1 2

z 1 0 0

x1 −6 2 3

x2 −2 4 1

s1 0 1 0

s2 0 0 1

BV z= s1 = s2 =

97

RHS 0 9 6

From the tableau we read that x1 should enter and s2 should leave. Solving, by using the Gauss–Jordan, the resulting tableau in terms of the BVs s1 , x2 gives us the tableau Row 0 1 2

z 1 0 0

x1 0 0 1

x2 0 3.333 0.333

s1 0 1 0

s2 2 −0.667 0.333

BV z= s1 = x1 =

RHS 12 5 2

We see that the tableau is optimal. The BFS found corresponds to the point B in the previous picture. So, we have found the optimal solution x1 = 2, x2 = 0, and z = 12. So, everything is fine. Except that we know — from the picture — that the found optimum is not unique. There are others. How does the Simplex tableau tell us this? Well, there is a NBV decision variable x2 with coefficient 0 in the 0th row. This means that making the decision variable x2 to enter as BV would not change the value of the objective. So let us — just for fun — make x2 to enter, and s1 leave. We get, after Gauss–Jordan, the following tableau Row 0 1 2

z 1 0 0

x1 0 0 1

x2 0 1 0

s1 0 0.3 −0.1

s2 2 −0.2 0.4

BV z= x2 = x1 =

RHS 12 1.5 1.5

We see that we have a new optimum. This time x1 = 1.5 and x2 = 1.5. This optimum corresponds to the point C in the previous picture. All the other optima are convex combinations of the optima we have just found, i.e. they are in the line segment from B to C . The bottom line is: Whenever there is a NBV with coefficient 0 on the 0th row of an optimal tableau, there are many optima.

Simplex Algorithm with Non-Unique Optima

98

Unbounded Optimum

7.2.3 Example. Consider the LP max z = x1 + s.t. x1 −

x2 x2 ≥ 1 6x2 ≥ 2 x1 , x2 ≥ 0

(1) (2) (3)

Here is the graphical representation of the LP of Example 7.2.1. The isoprofit lines are dashed.

x2

4 (1) 3

(2)

2

D

1 Feasible region 0

0

A 1

2

3

4

5 x1

−1 From the picture we see that this LP has unbounded optimum. Indeed, since the feasible region continues in the right up to infinity, one finds better and better solutions as one moves the isoprofit line further away from the origin (remember, no matter how far the isoprofit line is from the origin it will cross the x1 -axis somewhere). Let us then see what the Simplex algorithm have to say about unbounded optimum.

Simplex Algorithm with Non-Unique Optima

99

First, we transform the LP of Exercise 7.2.3 into a slack form. We obtain max z s.t. z − x1 − x1 −

(7.2.4)

x2 x2 − e 1 6x2 + s2 x1 , x2

= = = ≥

0 1 2 0

We see that we should use the Big M method here. That would indeed work. We do not go that way, however. Instead, we will be clever this time. We note that the system (7.2.4) is equivalent to the system max z s.t. z

− 2x2 − e1 x1 − x2 − e1 6x2 + s2 x1 , x2

= = = ≥

1 1 2 0

which is a canonical slack form if one chooses x1 and s2 as BVs. So, we obtain the first Simplex tableau Row 0 1 2

z 1 0 0

x1 0 1 0

x2 −2 −1 6

e1 −1 −1 0

s2 0 0 1

BV z= x1 = s2 =

RHS 1 1 2

Now it is obvious that x2 should enter as BV, and that s2 should leave. After solving the Tableau with x1 and x2 as BVs we obtain the Tableau Row 0 1 2

z 1 0 0

x1 0 1 0

x2 0 0 1

e1 −1 −1 0

s2 2 1 1

BV z= x1 = x2 =

RHS 5 3 2

Now it is obvious that e1 should enter as a new BV. So, let us try to determine the leaving variable. But the ratio test will give no limit. This means that no matter how much we increase the value of e1 the old BVs will remain positive, never reaching the boundary 0 and thus becoming NBVs. So, the conclusion is: Whenever the ratio test fails to identify a leaving variable the LP in question has an unbounded optimum. No Optima Let us agree that what does not exist is non-unique. The only way an LP may fail to have an optimal solution is that it has no feasible solutions at all.

Simplex Algorithm with Non-Unique Optima

100

7.2.5 Example. min z = 2x1 s.t. 0.5x1 x1 x1

+ 3x2 + 0.25x2 + 3x2 + x2 x1 , x2

≤ 4 ≥ 36 = 10 ≥ 0

(1) (2) (3) (4)

Here is the picture that illustrates the LP of Example 7.2.5. The isoprofit line is dashed.

x2 20

(1)

10

(2)

(3) 0

0

10

20 x1

Note that there are no arrows associated to the constraint line (3). This is because (3) is an equality, not an inequality. So, all the feasible solutions of the LP of Example 7.2.5 must be on the line (3). So, from the picture it is clear that the constraint (3) makes the feasible region empty, since lines (1)

Simplex Algorithm with Non-Unique Optima

101

and (2) don’t touch line (3) at the same point. Actually, line (2) does not touch the line (3) at all. Let us then see what the Simplex algorithm has to say about the LP of Example 7.2.5 To transform this LP into a canonical slack form we need artificial variables. This due to the ≥ constraint (2) and the = constraint (3). The first Simplex/Big M tableau for the LP of Example 7.2.5 is Row 0 1 2 3

−z 1 0 0 0

x1 2−2M 0.5 1 1

x2 3−4M 0.25 3 1

s1 0 1 0 0

e2 M 0 −1 0

a2 0 0 1 0

a3 0 0 0 1

BV z= s1 = a2 = a3 =

RHS −46M 4 36 10

Next we invoke the Simplex/Big M machinery in search for the optimal tableau. We omit the details here. Our final Simplex tableau will be Row 0 1 2 3

−z 1 0 0 0

x1 2M −1 0.25 −2 1

x2 0 0 0 1

s1 0 1 0 0

e2 M 0 −1 0

a2 0 0 1 0

a3 4M −3 −0.25 −3 1

BV z= s1 = a2 = x2 =

RHS −6M −30 1.5 6 10

We see that this is indeed the final Simplex tableau: All the NBVs have non-negative coefficients in the 0th row. But there is a problem: The artificial variable a2 is BV and a2 = 6 6= 0. But the artificial variables should be 0 in the final solution. Indeed, they were penalized heavily by the very very very very big number M in the objective function. Why did not the Simplex algorithm put them to be 0. The reason why the Simplex algorithm failed to put the artificial variables 0 was that otherwise there would not be any solutions. But in the original problem the artificial variables are 0 — or do not exist, which is the same thing. So, the conclusion is: Whenever there are non-zero artificial variables as BVs in the final Simplex tableau, the original LP does not have any feasible solutions.

Simplex/Big M Checklist

7.3

102

Simplex/Big M Checklist

The following list combines the algorithms presented in this and the previous lectures as checklist on how to solve LPs with Simplex/Big M method. It should be noted that in solving LPs in practice the may be many shortcuts — the checklist presented here is meant to be general rather than efficient. Step 1: Prepare the first canonical Simplex tableau Step 1-1: Transform the LP into a standard form. Step 1-2: Transform the standard form LP into a slack form. Step 1-3: Transform, if necessary, the slack form into a canonical slack form by adding artificial variables to each constraint row that lacks slacks. After this subtract M times the artificial variables from the objective function, and solve the system — by using the Gauss– Jordan method — in terms of the slacks and the artificial variables. Step 2: Carry out the Simplex algorithm Step 2-1: Transform the LP into canonical slack form. (Actually, this is what we just did in Step 1.) Step 2-2: Check if the current BFS is optimal. There are, omitting the possibility of multiple optimal solutions, three possibilities: (a) All the NBVs have non-negative coefficients, and all the artificial variables have zero coefficient: The algorithm terminates, and an optimum was found. The optimum can be read from the columns BV and RHS. (b) All the NBVs have non-negative coefficients, but some of the artificial variables have non-zero coefficients: The algorithm terminates, and the LP has no solutions. (c) Some of the NBVs have strictly negative coefficients: The algorithm moves to Step 2-3 (the next step). Step 2-3: Determine the entering variable. The NBV with smallest coefficient in 0th row will enter. Step 2-4: Determine the leaving variable. To do this perform the ratio test. Now there are two possibilities (a) Some BV wins the ratio test (gets the smallest number). That variable will leave. The algorithm then continues in Step 2-5. (b) All the ratios are either negative of ±∞. In this case the algorithm terminates, and the LP has an unbounded solution. Step 2-5: Find the new BFS for the new BVs by using the Gauss–Jordan method, and go back to Step 2-2.

Chapter 8

Sensitivity and Duality

The most important topic of linear programming is of course solving linear programs. We have just covered the topic in the previous lectures. The secondmost important topics in linear programming are sensitivity analysis and duality. (In [4, Ch 5] the author claims that sensitivity analysis and duality are the most important topics of linear programming. This author disagrees!) This lecture covers – at least the rudiments — of those. While sensitivity and duality are two distinct topics their connection is so close and profound that the Jane Austen type title “Sensitivity and Duality” is reasonable. This chapter is adapted from [2, Ch. 2] and [4, Ch. 5].

8.1

Sensitivity Analysis

What and Why is Sensitivity Analysis When one uses a mathematical model to describe reality one must make approximations. The world is more complicated than the kind of optimization problems that we are able to solve. Indeed, it may well be that the shortest model that explains the universe is the universe itself. Linearity assumptions usually are significant approximations. Another important approximation comes because one cannot be sure of the data one puts into the model. One’s knowledge of the relevant technology may be imprecise, forcing one to approximate the parameters A, b and c in the LP max z = c0 x s.t. Ax ≤ b x ≥ 0 Moreover, information may change. Sensitivity Analysis is a systematic study of how, well, sensitive, the solutions of the LP are to small changes in the data. The basic idea is to be able to give answers to questions of the form:

Sensitivity Analysis

104

1. If the objective function c changes in its parameter ci , how does the solution change? 2. If the resources available change, i.e., the constraint vector b change in its parameter bi , how does the solution change? 3. If a new constraint is added to the problem, how does the solution change? We shall give answers to the questions 1 and 2. Question 1 is related to the concept of reduced cost, a.k.a. the opportunity cost. Question 2 is related to the concept of shadow price, a.k.a. the marginal price. The question 3 will be completely ignored in these lectures. One approach to these questions is to solve lots and lots of LPs: One LP to each change in the parameters. For example, in Giapetto’s problem 3.3.1 there might be uncertainty in what is the actual market demand for soldiers. It was assumed to be 40, but it could be anything between 30 and 50. We could then solve the Giapetto’s LP separately for market demands 30, 31, . . . , 49, 50. So, we would solve 20 different LPs (21, actually, but who’s counting). If it is also assumed that the the profit for soldiers might not be exactly =C3 but could be anything between =C2.5 and =C3.5, then we could also solve the LP separately for profits =C2.5, =C2.6, . . . , =C3.4, =C3.5. Combining this with the different LPs we got from the uncertainty in the market demand we would then have 20 × 10 = 200 different LPs to solve (well, 21 × 11 = 231 if you count correctly). This “checking the scenarios” method would work, and it is indeed widely used in practice. This method has only two problems: (1) It is inelegant, and (2) it would involve a large amount of calculations. These problems are, however, not critical. Indeed, solving hundreds of LPs is not that time-consuming with modern computers and efficient algorithms like the Simplex. As for the inelegance of the scenario-based method: Who cares about elegance these days? Nevertheless, we shall try to be at least a little bit elegant in this chapter. Shadow Prices There are two central concepts in sensitivity analysis. They are so important that LP solvers will typically print their values in their standard reports. These are the shadow prices for constraints and reduced costs for decision variables. In this subsection we consider the shadow prices, and show where they are represented in the glpsol reports. 8.1.1 Definition. The Shadow Price of a constraint is the amount that the objective function value would change if the said constraint is changed by one unit — given that the optimal BVs don’t change. The shadow prices are typically denoted as the vector π = [π1 · · · πm ]0 . The following remarks of Definition 8.1.1 should help you to understand the concept.

Sensitivity Analysis

105

8.1.2 Remark. Note the clause “given that the optimal BVs don’t change”. This means that the shadow price is valid for small changes in the constraints. If the optimal corner changes when a constraint is changed, then the interpretation of the shadow price is no longer valid. It is valid, however, for all changes that are small enough, i.e., below some critical threshold. 8.1.3 Remark. Shadow prices are sometimes called Marginal Prices. E.g., GLPK calls them marginals. This is actually a much more informative name than the nebulous shadow price. Indeed, suppose you have a constraint that limits the amount of labor available to 40 hours per week. Then the shadow price will tell you how much you would be willing to pay for an additional hour of labor. If your shadow price is =C10 for the labor constraint, for instance, you should pay no more than =C10 an hour for additional labor. Labor costs of less than =C10 per hour will increase the objective value; labor costs of more than =C10 per hour will decrease the objective value. Labor costs of exactly =C10 will cause the objective function value to remain the same. If you like mathematical formulas — and even if you don’t — the shadow prices can be defined as follows: Consider the LP max z = c0 x s.t. Ax ≤ b x ≥ 0 Now, the optimal solution z ∗ of this LP is a function of the objective vector c, the technology matrix A, and the constraint vector b. So, z ∗ = z ∗ (c, A, b). Then the shadow price πi associated to the constraint bi is the partial derivative ∂z ∗ πi = , ∂bi or, in vector form, π = where

π1 π = ... πm

∂z ∗ , ∂b

and

∂z ∗ = ∂b

∂z ∗ ∂b1

.. .

∂z ∗ ∂bm

.

Suppose now that = [1 · · · m ]0 is a small vector, and z∗ = z ∗ (c, A, b + )

Sensitivity Analysis

106

is the optimal value, when the constraints b are changed by . Then the first order “Taylor approximation” for the new optimal value is ∂z ∗ ∂b = z ∗ + 0 π, .

z∗ = z ∗ + 0 (8.1.4)

The equality (8.1.4) is valid as long as the elements i of are small enough (in absolute value). If some of the elements of are too big, then the equality (8.1.4) may fail to be true. Let us see now how to use formula (8.1.4) in sensitivity analysis.

8.1.5 Example. Consider the LP max z = s.t.

4x1 + 3x2 2x1 + 3x2 −3x1 + 2x2 2x2 2x1 + x2 x1 , x2

≤ ≤ ≤ ≤ ≥

6 3 5 4 0

(0) (1) (2) (3) (4)

Here is the picture representing the LP. You have seen this picture before.

Sensitivity Analysis

107

x2 4

(2)

(4)

Redundant

3

(3)

2

D

E 1

Optimum Feasible region

C

Isoprofit lines (1) 0 A0

B 2

1

3

4 x1

From the picture we read — by moving the isoprofit line away from the origin — that the optimal point for the decision x = [x1 x2 ]0 is C

= (1.5, 1).

Therefore, the optimal value is of the objective is z = 4×1.5 + 3×1 = 9. We also see that the constraints (1) and (4) are active at the optimum. So, changing them should change the optimal value. Indeed, they should have positive shadow prices. In contrast, the constraints (2) and (3) are not active at the optimum. So, changing them — slightly — should have no effect on the optimum. So, both of them should have 0 as their shadow price. Let us then calculate the shadow prices. We could read the shadow prices from the final Simplex tableau. This would require much work, so we use GLPK instead. Here is the code that solves Example 8.1.5:

Sensitivity Analysis

108

# Sensitivity analysis for Example 8.1.7 # Part 1 - The original problem # Decision variables var x1 >=0; var x2 >=0; # Objective maximize z: 4*x1 + 3*x2; # Constraints s.t. r1: 2*x1 + 3*x2 <= 6; s.t. r2: -3*x1 + 2*x2 <= 3; s.t. r3: 2*x2 <= 5; s.t. r4: 2*x1 + x2 <= 4; end;

And here is the relevant part of the solution: No. -----1 2 3 4 5

Row name -----------z r1 r2 r3 r4

St Activity Lower bound Upper bound Marginal -- ------------- ------------- ------------- ------------B 9 NU 6 6 0.5 B -2.5 3 B 2 5 NU 4 4 1.5

No. -----1 2

Column name -----------x1 x2

St Activity Lower bound Upper bound Marginal -- ------------- ------------- ------------- ------------B 1.5 0 B 1 0

From the report we read that the shadow prices (called Marginal in the glpsol report) for the constraints (1) (r1 in the code) and (4) (r4 in the code) are 0.5 and 1.5, respectively. All other shadow prices are 0 (number omitted in the glpsol report). So, the shadow price vector is 0.5 0 π = 0 . 1.5 Let us then try to use formula (8.1.4) to see what happens if the constraints (1)–(4) change. Suppose each constraint is relaxed by 0.5. That means that in formula (8.1.4) we have = [0.5 0.5 0.5 0.5]0 . So, the new optimum should be z∗ = z ∗ + 0 π = 9 + 0.5×0.5 + 0.5×0 + 0.5×0 + 0.5×1.5 = 10.

Sensitivity Analysis

109

Is this correct? Let us see. Let us solve the new LP where the constraints are relaxed. So, we have to solve max z = s.t.

4x1 + 3x2 2x1 + 3x2 −3x1 + 2x2 2x2 2x1 + x2 x1 , x2

≤ ≤ ≤ ≤ ≥

6.5 3.5 5.5 4.5 0

(0) (1) (2) (3) (4)

The GNU MathProg code for this problem is # Sensitivity analysis for Example 8.1.7 # Part 2 - r1,r2,r3,r4 relaxed by 0.5 # Decision variables var x1 >=0; var x2 >=0; # Objective maximize z: 4*x1 + 3*x2; # Constraints s.t. r1: 2*x1 + 3*x2 <= 6.5; s.t. r2: -3*x1 + 2*x2 <= 3.5; s.t. r3: 2*x2 <= 5.5; s.t. r4: 2*x1 + x2 <= 4.5; end;

The relevant part of the glpsol report reads: No. -----1 2 3 4 5

Row name -----------z r1 r2 r3 r4

St Activity Lower bound Upper bound Marginal -- ------------- ------------- ------------- ------------B 10 NU 6.5 6.5 0.5 B -3.25 3.5 B 2 5.5 NU 4.5 4.5 1.5

No. -----1 2

Column name -----------x1 x2

St Activity Lower bound Upper bound Marginal -- ------------- ------------- ------------- ------------B 1.75 0 B 1 0

So, we see that formula (8.1.4) is correct, when ≤ [0.5 0.5 0.5 0.5]0 . Let us then consider a big change. Let = [10 10 10 0]0 . Then formula (8.1.4) would give us z∗ = z ∗ + 0 π = 9 + 10×0.5 + 10×0 + 10×0 + 0×1.5 = 14.

Sensitivity Analysis

110

Let’s see what really happens. The MathProg code for this relaxed LP is # Sensitivity analysis for Example 8.1.7 # Part 3 - r1,r2,r3 relaxed by 10; r4 not relaxed # Decision variables var x1 >=0; var x2 >=0; # Objective maximize z: 4*x1 + 3*x2; # Constraints s.t. r1: 2*x1 + 3*x2 <= 16; s.t. r2: -3*x1 + 2*x2 <= 13; s.t. r3: 2*x2 <= 15; s.t. r4: 2*x1 + x2 <= 4; end;

And the relevant part of the glpsol report tells us that No. -----1 2 3 4 5

Row name -----------z r1 r2 r3 r4

St Activity Lower bound Upper bound Marginal -- ------------- ------------- ------------- ------------B 12 B 12 16 B 8 13 B 8 15 NU 4 4 3

No. -----1 2

Column name -----------x1 x2

St Activity Lower bound Upper bound Marginal -- ------------- ------------- ------------- ------------NL 0 0 -2 B 4 0

We see that glpsol reports the optimum to be 12. The formula (8.1.4) gave us the optimum 14. So, the change [10 10 10 0]0 is not small enough for the formula (8.1.4) to be valid. What actually happened here was that the optimal point jumped corners. Reduced Costs Let us then consider the reduced costs. Remember that the shadow prices were associated to the constraints, or — if you like Simplex language — to the slacks. The reduced costs are associated to the decision variables. 8.1.6 Definition. The Reduced Cost ui for an NBV decision variable xi is the amount the objective value would decrease if xi would be forced to be 1, and thus a BV — given that the change from xi = 0 to xi = 1 is small. 8.1.7 Remark. Here are some interpretations and remarks of reduced costs that should help you to understand the concept:

Sensitivity Analysis

111

• The clause “given that the change from xi = 0 to xi = 1 is small” is a similar clause that the clause “given that the optimal BVs don’t change” was in Definition 8.1.1 of shadow price. Indeed, it may be, e.g., that forcing xi ≥ 1 will make the LP infeasible. Remember: In sensitivity analysis we are talking about small changes — whatever that means. The analysis may, and most often will, fail for big changes. • Decision variables that are BVs do not have reduced costs, or, if you like, their reduced costs are zero. • The reduced cost is also known as Opportunity Cost. Indeed, suppose we are given the forced opportunity (there are no problems — only opportunities) to produce one unit of xi that we would not otherwise manufacture at all. This opportunity would cost us, since our optimized objective would decrease to a suboptimal value. Indeed, we have now one more constraint — the forced opportunity — in our optimization problem. So, the optimal solution can only get worse. The decrease of the objective value is the opportunity cost. • The reduced cost ui of xi is the amount by which the objective coefficient ci for xi needs to change before xi will become non-zero. • As an example of the point above, consider that you are producing x1 , . . . , xn that will give you profits c1 , . . . , cn . You have some constraints, but the actual form of them does not matter here. Now, you form the LP to optimize your profit, and you solve it. You get optimal solution for productions: x∗1 , x∗2 , . . . , x∗n , and you get your optimal profit z ∗ . You notice that, say, x∗2 = 0. So, obviously the profit c2 for x2 is not big enough. Then you ask yourself: How big should the profit c2 for x2 be so that it becomes more profitable to produce x2 , at least a little, rather than not to produce x2 at all? The answer is c2 + u2 . This means that the profit must increase at least by the reduced cost before it becomes more profitable to produce a product you would not produce otherwise. Let us now consider the reduced cost with GLPK with the example: 8.1.8 Example. max z = s.t.

4x1 + 3x2 2x1 + 3x2 −3x1 + 2x2 2x2 2x1 + x2 x1 , x2

≤ 16 ≤ 13 ≤ 15 ≤ 4 ≥ 0

The GNU MathProg code for Example 8.1.8 is

(0) (1) (2) (3) (4)

Sensitivity Analysis

112

# Sensitivity analysis for Example 8.1.10 # Part 1 - The original problem # Decision variables var x1 >=0; var x2 >=0; # Objective maximize z: 4*x1 + 3*x2; # Constraints s.t. r1: 2*x1 + 3*x2 <= 16; s.t. r2: -3*x1 + 2*x2 <= 13; s.t. r3: 2*x2 <= 15; s.t. r4: 2*x1 + x2 <= 4; end;

The relevant part of the glpsol report says No. -----1 2 3 4 5

Row name -----------z r1 r2 r3 r4

St Activity Lower bound Upper bound Marginal -- ------------- ------------- ------------- ------------B 12 B 12 16 B 8 13 B 8 15 NU 4 4 3

No. -----1 2

Column name -----------x1 x2

St Activity Lower bound Upper bound Marginal -- ------------- ------------- ------------- ------------NL 0 0 -2 B 4 0

So, where are the reduced costs for x1 and x2? They are in the Marginal column. So, glpsol calls reduced costs and shadow prices with the same name “marginal”. How strange! Well, it is actually not at all so strange once we have learnt about duality. For now, let us just accept this as a mystery to be solved later. Back to the glpsol report: For x1 there is the reduced cost of 2 (glpsol uses a non-standard sign). For the decision variable x2 there is no reduced cost, since x2 is a BV. Let us then test the interpretation “reduced cost is the decrease in the value of the objective if we are forced to produce one unit where we otherwise would produce none”.

Sensitivity Analysis

113

We test the interpretation with the following LP: max z = s.t.

4x1 + 3x2 2x1 + 3x2 −3x1 + 2x2 2x2 2x1 + x2 x1 x1 , x2

≤ 16 ≤ 13 ≤ 15 ≤ 4 ≥ 1 ≥ 0

(0) (1) (2) (3) (4) (5)

So, we have added to the LP of Example 8.1.8 the requirement that we must have at least one x1 in the solution. This is the constraint (5). Remember that without this requirement we would have zero x1 ’s in the solution. So, here is the GNU MathProg code for this problem: # Sensitivity analysis for Example 8.1.10 # Part 2 - x1 forced to be at least one # Decision variables var x1 >=0; var x2 >=0; # Objective maximize z: 4*x1 + 3*x2; # Constraints s.t. r1: 2*x1 + 3*x2 <= 16; s.t. r2: -3*x1 + 2*x2 <= 13; s.t. r3: 2*x2 <= 15; s.t. r4: 2*x1 + x2 <= 4; s.t. r5: x1 >= 1; end;

Here is the glpsol report: No. -----1 2 3 4 5 6

Row name -----------z r1 r2 r3 r4 r5

St Activity Lower bound Upper bound Marginal -- ------------- ------------- ------------- ------------B 10 B 8 16 B 1 13 B 4 15 NU 4 4 3 NL 1 1 -2

No. -----1 2

Column name -----------x1 x2

St Activity Lower bound Upper bound Marginal -- ------------- ------------- ------------- ------------B 1 0 B 2 0

We see that the interpretation is indeed correct: The previous optimal value 12 dropped by 2 into 10.

Sensitivity Analysis

114

Sensitivity Analysis with Simplex with Matrices We have learned how to perform sensitivity analysis with GLPK. Let us then consider how sensitivity analysis works with the Simplex tableaux. To be more precise we explain how the shadow prices and the reduced costs can be read from the final optimal Simplex tableau. It would actually be very easy to just to give you the answer. Instead, we will try to explain why the answer is what it is. So we shall revisit the Simplex algorithm, now in matrix form. The answer to the critical question “how to read shadow prices and reduced costs from the optimal Simplex tableau” is deferred to the end of the subsection. Before we start let us introduce some matrix notation for the Simplex method. In doing this we also explain a little why the Simplex method works. It is best, at least here in the beginning, to consider an example:

8.1.9 Example. Consider the Giapetto’s LP 3.3.1 in slack form: max z = 3x1 + 2x2 s.t. 2x1 + x2 + s1 x1 + x2 + s2 x1 + s3 x1 , x2 , s1 , s2 , s3

Define 3 2 c= 0 0 0

,

x=

x1 x2 s1 s2 s3

= 100 = 80 = 40 ≥ 0

,

2 1 1 0 0 A = 1 1 0 1 0 , 1 0 0 0 1

100 b = 80 . 40

Then the Giapetto’s slack form LP in Example 8.1.9 can be written neatly in the matrix form as (8.1.10)

max z = c0 x s.t. Ax = b x ≥ 0

Now, the first Simplex tableau for Example 8.1.9 is

Sensitivity Analysis

Row 0 1 2 3

z 1 0 0 0

x1 −3 2 1 1

115 x2 −2 1 1 0

s1 0 1 0 0

s2 0 0 1 0

s3 0 0 0 1

BV z= s1 = s2 = s3 =

RHS 0 100 80 40

This tableau corresponds to the augmented matrix 1 −c0 0 . 0 A b The last (optimal) Simplex tableau is Row 0 1 2 3

z 1 0 0 0

x1 0 0 0 1

x2 0 1 0 0

s1 1 −1 −1 0

s2 1 2 1 −1

s3 0 0 1 0

BV z= x2 = s3 = x1 =

RHS 180 60 20 20

This tableau corresponds to the augmented matrix 1 −c∗BV 0 −c∗NBV 0 z∗ (8.1.11) , 0 B∗ N∗ x∗BV

Sensitivity Analysis

116

where we have used the denotations x2 60 x∗BV = s3 = 20 x1 20 = the BV component of the optimal BFS x∗ , s1 0 ∗ xNBV = = s2 0 = the NBV component of the optimal BFS x∗ , 1 0 0 B∗ = I = 0 1 0 0 0 1 = the BV columns of

N∗

the “augmented constraint matrix at optimum”, −1 2 1 = −1 0 −1 = the NBV columns of

c∗BV = = c∗NBV = =

the “augmented constraint matrix at optimum”, 0 0 0 the BV coefficients at optimum, 1 1 the NBV coefficients at optimum.

8.1.12 Remark. The columns in (8.1.11) are in different order than in the corresponding optimal Simplex tableau. This is unfortunate, but cannot be avoided, since we want to list first the BVs and then the NBVs. This is not a problem, however. Indeed, it is obvious that the original optimal tableau Row 0 1 2 3

z 1 0 0 0

x1 0 0 0 1

x2 0 1 0 0

s1 1 −1 −1 0

s2 1 2 1 −1

s3 0 0 1 0

BV z= x2 = s3 = x1 =

RHS 180 60 20 20

s2 1 2 1 −1

BV z= x2 = s3 = x1 =

RHS 180 60 20 20

and the column-changed tableau Row 0 1 2 3

z 1 0 0 0

x2 0 1 0 0

s3 0 0 1 0

x1 0 0 0 1

s1 1 −1 −1 0

Sensitivity Analysis

117

are the same. And this latter tableau should look very similar to the augmented matrix (8.1.11), indeed. Now, using the obvious non-starred analog of the notation we have just introduced the matrix form LP (8.1.10) can be stated in matrix form as (8.1.13)

max z = c0BV xBV + c0NBV xNBV s.t. BxBV + NxNBV = b xBV , xNBV ≥ 0

In augmented matrix form (8.1.13) reads 1 −c0BV −c0NBV 0 (8.1.14) . 0 B N b (Note that we are now back in the initial problem, or the first tableau.) Now, we show how the last simplex tableau is found analytically from the LP (8.1.13), or from the first simplex tableau, which is nothing but the augmented matrix form (8.1.14). First, we multiply the constraint equation from the left by the inverse B−1 . At this point you may wonder how does one find the inverse B−1 . Well, actually we found it in the Simplex Step 5, where we solved the tableau. So, the Step 5 was actually finding the inverse of the BV part B of the matrix A. So, the constraint BxBV + NxNBV = b becomes xBV + B−1 NxNBV = B−1 b. Now, recall that the NBVs are all zeros. So, B−1 NxNBV = 0, and we actually have the constraint xBV = B−1 b. What have we found out? We have found out that the optimization problem (8.1.13), that was equivalent to the problem (8.1.10), is actually equivalent to the problem (8.1.15)

max z = c0BV xBV + c0NBV xNBV s.t. xBV = B−1 b xBV , xNBV ≥ 0

Now, look at the objective row “constraint” z = c0BV xBV + c0NBV xNBV . With a little bit of matrix algebra we see that this is equivalent to z = c0BV B−1 b.

Sensitivity Analysis

118

Indeed, the c0NBV xNBV part vanishes, since xNBV = 0, and we know already that xBV = B−1 b. So, we have found that the LP (8.1.10), or the LP (8.1.13), or the LP (8.1.15), is actually equivalent to the LP (8.1.16)

max z = c0BV B−1 b s.t. xBV = B−1 b xBV , xNBV ≥ 0

It may not be obvious at this point, but we have actually seen the vector c0BV B−1 — or actually the vector −c0BV B−1 — already. They are the shadow prices: π 0 = c0BV B−1 . Actually, defined this way, the shadow prices are associated to all BVs, not only to BVs that are slacks, and thus related to the constraints only. Let us now collect what we have found out — at least to some extend — of the Simplex algorithm as a theorem. 8.1.17 Theorem. Let xBV be a BFS of the LP (8.1.10). Then xBV

= B−1 b,

and the corresponding value of the objective is z

= c0BV B−1 b = π 0 b.

Moreover, the coefficients in the 0th row of the NBV part of the Simplex tableau are π 0 N − c0NBV . The point x is feasible if B−1 b ≥ 0, and the point x is optimal if it is feasible and the reduced cost vector satisfies π 0 N − c0NBV ≥ 00 . Now, finally, comes the answer you have been waiting for. Well, not yet! We still need some notation. We shall split the coefficients in the final Simplex tableau in two parts. Here we also use different notation for decision variables and slacks: previously the vector x contained both the decision variables and the slack. Now, the vector x will denote only the decision variables, and the slacks (and excesses) will be denoted by the vector s. The matrix A will still denote the extended matrix that includes the slacks. Let u = coefficients of the decision variables at the 0th row, v = coefficients of the slacks at the 0th row. Now, we have just learned that if the decision variable xj is BV, then uj = 0, if the decision variable xj is NBV, then uj = π 0 a•j − cj .

Sensitivity Analysis

119

We have also just learned that if the slack si is BV, then vi = 0, if the slack si is NBV, then vi = π 0 a•(n+i) = πi . With the notation introduced above, the 0th row of the optimal Simplex tableau represents the equation z+

(8.1.18)

n X

uj xj +

m X

j=1

vi si = z ∗ ,

i=1

or — if you like matrix notation — the equation z + u0 x + v 0 s = z ∗ . Suppose then that we change the situation so that a NBV xk becomes a BV with value xk = 1. This of course requires changing the values of BVs, but the bottom line is that the equation (8.1.18) will now have an additional term uk xk = uk on its LHS (Left Hand Side). This is the only change in the 0th row since the BVs had coefficients 0 in the optimal 0th row, as they were eliminated from the 0th row by the Simplex algorithm. Consequently the RHS has decreased by uk . The conclusion is that uk is the reduced cost of the NBV xk . Suppose then that some resource constraint bk is increased to bk + 1. Since cBV and B−1 remain unchanged, remains π = c0BV B−1 also unchanged. So, the new optimum must be z new∗ = π 0 bnew n X = πi bnew i = =

i=1 n X

πi bi + πk (bk + 1)

i=1,i6=k n X

πi bi + πk

i=1 old∗

= z = z

old∗

+ πk + vk .

The conclusion is that vk is the shadow price of the resource bk .

Sensitivity Analysis

120

(Until now the association π 0 = c0BV B−1 was just notation. Now we have just justified that if π is defined this way, then it actually is the vector of shadow prices in the sense of Definition 8.1.1.) So, back to the Giapetto’s problem 8.1.9 with the final optimal Simplex tableau: Row 0 1 2 3

z 1 0 0 0

x1 0 0 0 1

x2 0 1 0 0

s1 1 −1 −1 0

s2 1 2 1 −1

s3 0 0 1 0

BV z= x2 = s3 = x1 =

RHS 180 60 20 20

For the 0th row we read that u =

0 0

1 v = 1 . 0

and

So, the shadow prices for the NBV slacks s1 and s2 are 1 for both. So, additional carpentry and finishing hours are worth 1 Euro per hour both for Giapetto. Since s3 is a non-zero BV additional unit of market demand for soldiers is worthless to Giapetto. Indeed, in the optimal solution Giapetto is not meeting the market demand. Finally, we see that the reduced costs are zero, since all the decision variables are BVs. Let us end this section with yet another example that illustrates how shadow prices and reduced costs can be read from the optimal Simplex tableau. The following Example is actually the Dakota Furniture’s problem 6.2.2 without the market demand constraint x2 ≤ 5 (that turned out to be redundant anyway).

8.1.19 Example. We want to perform sensitivity analysis to the LP max z = 60x1 s.t. 8x1 4x1 2x1

+ 30x2 + 20x3 + 6x2 + x3 + 2x2 + 1.5x3 + 3x2 + 0.5x3 x1 , x2 , s1 , s2 , s3

≤ 48 ≤ 20 ≤ 8 ≥ 0

by using the Simplex method.

We see that the LP in Example 8.1.19 Has only inequalities of type ≤, and all the RHSs are non-negative. So, we will only have surplus type slacks, and

Dual Problem

121

we obtain immediately a canonical slack form. So, our first simplex tableau is (was) Row 0 1 2 3

z 1 0 0 0

x1 −60 8 4 2

x2 −30 6 2 3

x3 −20 1 1.5 0.5

s1 0 1 0 0

s2 0 0 1 0

s3 0 0 0 1

BV z= s1 = s2 = s3 =

RHS 0 48 20 8

Next we have to perform the Simplex algorithm to find the optimal tableau. After long and tedious tableau-dancing in Chapter 6 we obtained the optimal tableau: Row 0 1 2 3

z 1 0 0 0

x1 0 0 0 1

x2 5 −2 −2 1.25

x3 0 0 1 0

s1 0 1 0 0

s2 10 2 2 −0.5

s3 10 −8 −4 1.5

BV z= s1 = x3 = x1 =

RHS 280 24 8 2

Now we can read sensitivity information from the 0th row: 0 0 u = 5 and v = 10 . 0 10 We see that the reduced cost for the not-produced product x2 (tables) is 5. This means, e.g., that the profit for making tables should increase at least =C5 before it makes sense to produce them. Or, if you like, producing one table would decrease the profit =C280 by =C5. The reduced costs for x1 and x3 Âăare zero, since they are BVs. The shadow prices are: 0 for the slack s1 (lumber), since it is not active, and 10 for both the active constraints s2 and s3 (finishing and carpentry). So, additional carpentry and finishing hours are both worth =C10 for Dakota and additional lumber is worthless.

8.2

Dual Problem

Finding Dual Associated with any LP there is another LP, called the dual — and then the original LP is called the primal. The relationship between the primal and dual is important because it gives interesting economic insights. Also, it is important because it gives a connection between the shadow prices and the reduced costs. In general, if the primal LP is a maximization problem, the dual is a minimization problem — and vice versa. Also, the constraints of the primal LP are the coefficients of the objective of the dual problem — and vice versa. If

Dual Problem

122

the constraints of the primal LP are of type ≤ then the constraints of the dual LP are of type ≥ — and vice versa. Let us now give the formal definition of the dual. We assume that the primal LP is in standard form of Definition 5.1.9. Since all LPs can be transformed into a standard form this assumption does not restrict the generality of the duality. The assumption is made only for the sake of convenience. 8.2.1 Definition. The dual of the standard form LP max z = s.t. (8.2.2)

c1 x1 a11 x1 a21 x1 .. .

+ + + .. .

c2 x2 a12 x2 a22 x2 .. .

am1 x1 + am2 x2

+ · · · + cn xn + · · · + a1n xn + · · · + a2n xn .. . . . .. . .. . . + · · · + amn xn x1 , x2 , . . . , xn

≤ ≤ .. .

b1 , b2 , .. .

≤ bm , ≥ 0.

is

(8.2.3)

min w = b1 y1 s.t. a11 y1 a11 y1 .. . a1n y1

+ b2 y2 + a21 y2 + a22 y2 .. .. . . + a2n y2

+ · · · + bm ym + · · · + am1 ym + · · · + am2 xm .. . . . .. . .. . . + · · · + amn ym y1 , y2 , . . . , ym

≥ c1 , ≥ c2 , .. .. . . ≥ cn , ≥ 0.

In matrix form the duality can be written as: The dual of the LP max z = c0 x s.t. Ax ≤ b x ≥ 0 is

min w = b0 y s.t. A0 y ≥ c y ≥ 0

Dual Problem

123

8.2.4 Example. Consider the LP x1 max z = [1 2 3] x2 x3

s.t. The dual LP is

4 5 6 7 8 9

x1 10 x2 ≤ . 11 x3

y1 min w = [10 11] y2 4 7 1 y 1 s.t. 5 8 ≥ 2 . y2 6 9 3

8.2.5 Remark. Let us discuss briefly about concept of duality in general and the duality of Definition 8.2.1 in particular. • In general, dual is a transformation with the following property: Transforming twice you get back. This is the abstract definition of duality. In mathematics a function is f is called involution if it is its own inverse, i.e., f −1 = f . So, duality is a meta-mathematical involution. • Looking at Definition 8.2.1 one sees the dual is LP itself. So, it can be transformed into a standard form, and the one can construct the dual of the dual. When one does so one gets back to the original primal LP, i.e., the dual of the dual is primal. So, the dual of Definition 8.2.1 deserves its name. • We have already seen one duality between LPs before: A minimization problem is in duality with a maximization problem with the transform where the objective function is multiplied by −1. The usefulness of this simple duality was that we only need to consider maximization problems, and the solution of the minimization problem is −1 times the solution of the corresponding maximization problem in this simple duality. Also, the optimal decisions in the maximization and minimization problems are the same. • The duality of Definition 8.2.1 is more complicated than the simple “multiply by −1 duality” of the previous point. This makes the duality of Definition 8.2.1 in some sense more useful than the simple “multiply by −1 duality”. Indeed, since the transformation is more complicated, our

Dual Problem

124

change of perspective is more radical, and thus this transformation gives us better intuition of the original problem. • The duality of Definition 8.2.1 is very useful because of the following theorems: The weak duality theorem states that the objective function value w of the dual at any feasible solution y is always greater than or equal to the objective function value z of the primal at any feasible solution x: w = b0 y ≥ c0 x = z. The weak duality theorem can be used to get upper bounds to the primal LP. The strong duality theorem states that if the primal has an optimal solution, x∗ , then the dual also has an optimal solution, y∗ , such that z ∗ = c0 x∗ = b0 y∗ = w∗ . The strong duality theorem can be used to solve the primal LP. We shall prove the weak and strong duality theorems later in this lecture. Let us find a dual of an LP that is not in standard form. 8.2.6 Example. Consider the LP min z = 50x1 s.t. 2x1 12x1 x1

+ 20x2 + 3x2 + 13x2 + x2

+ 30x3 + 4x3 + 14x3 + x3 x1 , x2 , x3

≥ 11 ≤ 111 = 1 ≥ 0

The LP of Example 8.2.6 is not in standard form. So, before constructing its dual, we transform it into standard form. This is not necessary. Sometimes we can be clever, and find the dual without first transforming the primal into standard form. But we don’t feel clever now. So, here is the standard form: max −z = −50x1 s.t. −2x1 12x1 x1 −x1

− 20x2 − 3x2 + 13x2 + x2 − x2

− 30x3 − 4x3 + 14x3 + x3 − x3 x1 , x2 , x3

≤ −11 ≤ 111 ≤ 1 ≤ −1 ≥ 0

Now we are ready to present the dual:

(8.2.7)

min −w = −11y1 s.t. −2y1 −3y1 −4y1

+ 111y2 + 12y2 + 13y2 + 14y2

+ + + +

y3 − y4 y3 − y4 y3 − y4 y3 − y4 y1 , y2 , y3 , y4

≥ −50 ≥ −20 ≥ −30 ≥ 0

Dual Problem

125

(we used variable −w in the dual because there was variable −z in the standard form primal). Note now that th dual LP (8.2.7) in in “dual standard form”: It is a minimization problem with only inequalities of type ≥. The original primal LP was a minimization problem. So, it is natural to express the dual LP as a maximization problem. Also, inequalities of type ≤ are more natural to maximization problems than the opposite type inequalities ≥. So, let us transform the dual LP (8.2.7) into a maximization problem with ≤ type inequalities. In fact, let us transform the dual LP (8.2.7) into a standard form. We obtain max w = 11y1 s.t. 2y1 3y1 4y1

− 111y2 − 12y2 − 13y2 − 14y2

− − − −

y3 + y4 y3 + y4 y3 + y4 y3 + y4 y1 , y2 , y3 , y4

≤ 50 ≤ 20 ≤ 30 ≥ 0

Economic Interpretation of Dual Let us recall — again — the Dakota Furniture’s problem 6.2.2 (without the market demand constraint that turned out to be irrelevant anyway): max z = 60x1 s.t. 8x1 4x1 2x1

+ 30x2 + 6x2 + 2x2 + 1.5x2

+ 20x3 + x3 + 1.5x3 + 0.5x3 x1 , x2 , x3

≤ 48 ≤ 20 ≤ 8 ≥ 0

(lumber) (finishing) (carpentry)

where x1 = number of desks manufactured x2 = number of tables manufactured x3 = number of chairs manufactured Now, the dual of this problem is

(8.2.8)

min w = 48y1 s.t. 8y1 6y1 x1

+ 20y2 + 4y2 + 2y2 + 1.5y2

+ 8y3 + 2y3 + 1.5y3 + 0.5y3 y1 , y2 , y3

≥ 60 ≥ 30 ≥ 20 ≥ 0

(desk) (table) (chair)

We have given the constraints the names (desk), (table), and (chair). Those were the decision variables x1 , x2 and x3 in the primal LP. By symmetry, or duality, we could say that y1 is associated with lumber, y2 with finishing, and y3 with carpentry. What is going on here? It is instructive to represent the

Dual Problem

126

data of the Dakota’s problem in a table where we try to avoid taking Dakota’s point of view: Lumber Finishing Carpentry Price

Desk 8 units 4 hours 2 hours =C60

Table 6 units 2 hours 1.5 hours =C30

Chair 1 unit 1.5 hours 0.5 hours =C20

Availability 48 units 20 hours 8 hours

Now, the table above can be read either horizontally of vertically. You should already know how the read the table above horizontally. That is the Dakota’s point of view. But what does it mean to read the table vertically? Here is the explanation, that is also the economic interpretation of the dual LP: Suppose you are an entrepreneur who wants to purchase all of Dakota’s resources — maybe you are a competing furniture manufacturer, or maybe you need the resources to produce soldiers and trains like Giapetto. Then you must determine the price you are willing to pay for a unit of each of Dakota’s resources. But what are the Dakota’s resources? Well they are lumber, finishing hours, and carpentry hours, that Dakota uses to make its products. So, the decision variables for the entrepreneur who wants to buy Dakota’s resources are: y1 = price to pay for one unit of lumber y2 = price to pay for one hour of finishing labor y3 = price to pay for one hour of carpentry labor Now we argue that the resource prices y1 , y2 , y3 should be determined by solving the Dakota dual (8.2.8). First note that you are buying all of Dakota’s resources. Also, note that this is a minimization problem: You want to pay as little as possible. So, the objective function is min w = 48y1 + 20y2 + 8y3 . Indeed, Dakota has 48 units of lumber, 20 hours of finishing labor, and 8 hours of carpentry labor. Now we have the decision variables and the objective. How about constraints? In setting the resource prices y1 , y2 , and y3 , what kind of constraints do you face? You must set the resource prices high enough so that Dakota would sell them to you. Now Dakota can either use the resources itself, or sell them to you. How is Dakota using its resources? Dakota manufactures desks, tables, and chair. Take desks first. With 8 units of lumber, 4 hours of finishing labor, and 2 hours of carpentry labor Dakota can make a desk that will sell for =C60. So, you have to offer more than =C60 for this particular combination of resources. So, you have the constraint 8y1 + 4y2 + 2y3 ≥ 60.

Dual Problem

127

But this is just the first constraint in the Dakota dual, denoted by (desk). Similar reasoning shows that you must pay at least =C30 for the resources Dakota uses to produce one table. So, you get the second constraint, denoted by (table), of the Dakota dual: 6y1 + 2y2 + 1.5y3 ≥ 30. Similarly, you must offer more than =C20 for the resources the Dakota can use itself to produce one chair. That way you get the last constraint, labeled as (chair), of the Dakota dual: y1 + 1.5y2 + 0.5y3 ≥ 20. We have just interpreted economically the dual of a maximization problem. Let us then change our point of view to the opposite and interpret economically the dual of a minimization problem.

8.2.9 Example. My diet requires that all the food I eat come from the four “basic food groups”: chocolate cake, ice cream, soda, and cheese cake. At present four foods are available: brownies, chocolate ice cream, cola, and pineapple cheesecake. The costs of the foods (in Cents) and my daily nutritional requirements together with their calorie, chocolate, sugar, and fat contents are listed in the table below this box. I want to minimize the cost of my diet. How should I eat?

Brownie Chocolate ice cream Cola Pineapple cheesecake Requirement

Calories 400 200 150 500 500

Chocolate 3 2 0 0 6

Sugar 2 2 4 4 10

Fat 2 4 1 5 8

Price 50 20 30 80

Let us then find the LP for the Diet problem of Example 8.2.9. As always, we first determine the decision variables. The decision to be made is: how much each type of food should be eaten daily. So, we have the decision variables x1 = number of brownies eaten daily, x2 = number (of scoops) of chocolate ice creams eaten daily, x3 = number (of bottles) of cola drunk daily, x4 = number (of pieces) of pineapple cheesecake eaten daily.

Dual Problem

128

Next we define the objective. We want to minimize the cost of the diet. So, the objective is min z = 50x1 + 20x2 + 30x3 + 80x4 . Finally, we define the constraints. The daily calorie intake requirement gives 400x1 + 200x2 + 150x3 + 500x4 ≥ 500. The daily chocolate requirement gives 3x1 + 2x2 ≥ 6. The daily sugar requirement gives 2x1 + 2x2 + 4x3 + 4x4 ≥ 10, and the daily fat requirement gives 2x1 + 4x2 + x3 + 5x4 ≥ 8. So, we see that the Diet problem of Example 8.2.9 is the LP (8.2.10) min z = 50x1 + 20x2 + 30x3 + 80x4 s.t. 400x1 + 200x2 + 150x3 + 500x4 ≥ 400 3x1 + 2x2 ≥ 6 2x1 + 2x2 + 4x3 + 4x4 ≥ 10 2x1 + 4x2 + x3 + 5x4 ≥ 8 x1 , x2 , x3 , x4 ≥ 0

(calorie) (chocolate) (sugar) (fat)

What about the dual of (8.2.10) then. Now, the LP (8.2.10) is not in standard form. So, in principle we should first transform it into standard form, and then construct the dual. We shall not do that, however. Instead, we remember that the dual of the dual is primal. So, we read the Definition 8.2.1 backwards, and obtain immediately the dual of (8.2.10): (8.2.11) max w = 500y1 + 6y2 + 10y3 + 8y4 s.t. 400y1 + 3y2 + 2y3 + 2y4 ≤ 50 (brownie) 200y1 + 2y2 + 2y3 + 4y4 ≤ 20 (ice cream) 150y1 + 4y3 + y4 ≤ 30 (soda) 500y1 + 4y3 + 5y4 ≤ 80 (cheesecake) y1 , y2 , y3 , y4 ≥ 0 What is then the economic interpretation of this dual? Well, reading the table

Dual Problem

129

Brownie Chocolate ice cream Cola Pineapple cheesecake Requirement

Calories 400 200 150 500 500

Chocolate 3 2 0 0 6

Sugar 2 2 4 4 10

Fat 2 4 1 5 8

Price 50 20 30 80

vertically, instead of horizontally, we see that we can consider a “nutrient” salesperson who sells calories, chocolate, sugar, and fat. The salesperson wishes to ensure that a dieter will meet all of his daily requirements by purchasing calories, sugar, fat, and chocolate from the the salesperson. So, the salesperson must determine the prices of her products: y1 = price of a calorie, y2 = price of a unit of chocolate, y3 = price of a unit of sugar, y4 = price of a unit of fat. The salesperson wants to maximize her profit. So, what is the salesperson selling? She is selling daily diets. So, the objective is max w = 500y1 + 6y2 + 10y3 + 8y4 . What are the constraints for the salesperson? In setting the nutrient prices she must set the prices low enough so that it will be in the dieter’s economic interest to purchase as his nutrients from her. For example, by purchasing a brownie for =C0.50, the dieter can obtain 400 calories, 3 units of chocolate, 2 units of sugar, and 2 units of fat. So, the salesperson cannot charge more than =C0.50 for this combination of nutrients. This gives her the brownie constraint 400y1 + 3y2 + 2y3 + 2y4 ≤ 50 (remember, we counted in Cents). In the similar way the salesperson will have the ice cream, soda, and cheesecake constraints listed in the dual LP (8.2.11). Duality Theorem In this subsection we discuss one of the most important results in linear programming: the Duality Theorem. In essence, the Duality Theorem states that the primal and the dual have equal optimal objective function values — given that the problems have optimal solutions. While this result is interesting in its own right, we will see that in proving it we gain many important insights into linear programming. As before, we assume — for the sake of convenience — that the primal is in standard form. So, the primal will be a maximization problem, and the dual

Dual Problem

130

will be a minimization problem. For the sake of reference you may think that we are dealing with the Dakota’s problem 6.2.2 (without the irrelevant market demand constraint) and its dual (8.2.8). The next theorem is the Weak Duality Theorem. 8.2.12 Theorem. Let x be any BFS of the primal LP and let y be any BFS of the dual LP. Then z = c0 x ≤ b0 y = w. Let us actually prove the Weak Duality Theorem 8.2.12: Proof. Consider any of the dual decision variable yi , i = 1, . . . , m. Since yi ≥ 0 we can multiply the ith primal constraint by yi without changing the direction of the constraint number i. (Moreover, the system remains equivalent, but that’s not important here). We obtain (8.2.13)

yi ai1 x1 + · · · + yi ain xn ≤ bi yi

for all i = 1, . . . , m.

Adding up all the m inequalities (8.2.13), we find that m X n X

(8.2.14)

yi aij xj

≤

i=1 j=1

m X

bi yi .

i=1

Similarly, if we consider any of the primal decision variables xj , j = 1, . . . , n, we have that xj ≥ 0. So, we can multiply the j th dual constraint by the decision xj without changing the direction of the constraint. We obtain xj a1j y1 + · · · + xj amj ym ≥ cj xj .

(8.2.15)

Adding up all the n inequalities (8.2.15), we find that m X n X

(8.2.16)

yi aij xj

i=1 j=1

≥

n X

cj xj .

j=1

Combining (8.2.14) and (8.2.16), we obtain double-inequality n X j=1

cj xj

≤

m X n X i=1 j=1

yi aij xj

≤

m X

bi yi .

i=1

But, that’s it! Let us then consider the consequences — or corollaries — of the Weak Duality Theorem 8.2.12. 8.2.17 Corollary. If the primal LP and dual LP both are feasible, then the their optimal solutions are bounded.

Dual Problem

131

Proof. Let y be a BFS for the dual problem. Then the Weak Duality Theorem 8.2.12 shows that b0 y is an upper bound for any objective value c0 x associated with any BFS x of the primal LP: c0 x ≤ b0 y. Since this is true for any primal decision x, it is true for an optimal primal decision x∗ also. So, z ∗ = c 0 x∗ = max c0 x ; Ax ≤ b, x ≥ 0 ≤ max b0 y ; Ax ≤ b, x ≥ 0 ≤ b0 y < ∞ is bounded. Now, change the rôles of the primal and the dual, and you see that the claim of Corollary 8.2.17 is true. 8.2.18 Corollary. Suppose x∗ is a BFS for the primal and y∗ is a BFS for the dual. Suppose further that c0 x∗ = b0 y∗ . Then both x∗ and y∗ are optimal for their respective problems. Proof. If x is any BFS for the primal, then the Weak Duality Theorem 8.2.12 tells us that c0 x ≤ b0 y∗ = c0 x∗ . But this means that x∗ is primal optimal. Now, change the rôles of the primal and dual, and you see that the claim of the Corollary 8.2.18 is true. Here is the Duality Theorem, or the Strong Duality Theorem. To understand the notation, recall Theorem 8.1.17 earlier in this Chapter. 8.2.19 Theorem. Let xBV = B−1 b be the optimal BFS to the primal with the corresponding optimal objective value z ∗ = c0BV B−1 b = π 0 b. Then π is the optimal BFS for the dual. Also, the values of the objectives at the optimum are the same: w∗

= π 0 b = c0BV xBV .

We shall not prove Theorem 8.2.19 in these notes. Instead, we leave it as an exercise. The author is well aware that this is a very demanding exercise, but not all of the exercises have to be easy! Also, there will be no proofs in the final exam, so it is justified that there is at least one proof in the exercises.

Dual Problem

132

Complementary Slackness It is possible to obtain an optimal solution to the dual when only an optimal solution to the primal is known using the Theorem of Complementary Slackness. To state this theorem, we assume that the primal is in standard form with non-negative RHSs and objective coefficients. The primal decision variables are x = [x1 · · · xn ]0 and the primal slacks are s = [s1 · · · sm ]0 . The dual is then a minimization problem with decision variables y = [y1 · · · ym ]0 , and with n constraints of type ≥ with non-negative RHSs. Let e = [e1 · · · en ]0 be the excesses of the dual problem associated with the constraints. So, in slack form the primal LP is max z = s.t.

c1 x1 + · · · + cn xn a11 x1 + · · · + a1n xn +s1 a21 x1 + · · · + a2n xn +s2 .. .. .. .. . . . . am1 x1 + · · · + amn xn +sm x1 , . . . , xn , s1 , . . . , sm

= b1 = b2 .. . = bm ≥0

Similarly, the dual LP in slack — or rather excess — form is min w = b1 y1 + · · · + bm ym s.t. a11 y1 + · · · + am1 ym −e1 a12 y1 + · · · + am2 ym −e2 .. .. .. .. . . . . a1n y1 + · · · + amn ym −en y1 , . . . , ym , e1 , . . . , en

= c1 = c2 .. . = cn ≥0

Here is the Theorem of Complementary Slackness. 8.2.20 Theorem. Let x be a primal BFS, and let y be a dual BFS. Then x is primal optimal and y is dual optimal if and only if s i yi = 0

for all i = 1, . . . , m,

e j xj

for all j = 1, . . . , n.

= 0

Before going into the proof of the Complementary Slackness Theorem 8.2.20 let us note that: 8.2.21 Remark. Theorem 8.2.20 says that if a constraint in either the primal or the dual is non-active, then the corresponding variable in the other — complementary — problem must be zero. Hence the name complementary slackness.

Dual Problem

133

Proof. The theorem 8.2.20 is of the type “if and only if”. So, there are two parts in the proof: the “if part” and the “only if part”. Before going to those parts let us note that (8.2.22)

si = 0

m X

if and only if

aij x∗j = bi ,

j=1

(8.2.23)

ej = 0

n X

if and only if

aij yi∗ = cj .

i=1

If part: By (8.2.22) we see that m X

bi yi∗

=

i=1

m X i=1

=

yi∗

n X

aij x∗j

j=1

m X n X

yi∗ aij x∗j

i=1 j=1

In the same way, by using (8.2.23) we see that n X

cj x∗j =

j=1

m X n X

yi∗ aij x∗j .

i=1 j=1

So, the conclusion is that n X

cj x∗j =

j=1

m X

bi yi∗ ,

i=1

and the “if part” follows from the Weak Duality Theorem 8.2.12. Only if part: Like in the proof of the Weak Duality Theorem 8.2.12 we obtain (8.2.24)

n X

cj x∗j ≤

j=1

m X n X i=1 j=1

yi∗ aij x∗j ≤

m X

bi yi∗ .

i=1

Now, by the Strong Duality Theorem 8.2.19, if x∗ and y∗ are optimal, then the LHS of (8.2.24) is equal to the RHS of (8.2.24). But this means that ! n m X X (8.2.25) cj − yi∗ aij x∗j = 0. j=1

i=1

Now, both x∗ and y∗ are feasible. This means that the terms in (8.2.25) are all non-negative. This implies that the terms are all zeros. But this means

Dual Problem

134

that that ej xj = 0. The validity of si yi = 0 can be seen in the same way by considering the equality m m X X bi − aij x∗j yi∗ = 0. i=1

j=1

This finishes the proof of the Complementary Slackness Theorem 8.2.20. As an example of the use of the Complementary Slackness Theorem 8.2.20, let us consider solving the following LP: 8.2.26 Example. You want to solve the LP min w = 4y1 + 12y2 + y3 s.t. y1 + 4y2 − y3 ≥ 1 2y1 + 2y2 + y3 ≥ 1 y1 , y2 , y3 ≥ 0

Now, suppose you have already solved, e.g. graphically — which is challenging for the LP in 8.2.26 — the much easier LP max z = s.t.

x1 x1 4x1 −x1

+ x2 + 2x2 + 2x2 + x2 x1 , x2

≤ 4 ≤ 12 ≤ 1 ≥ 0

The solution to this dual is x∗1 = 8/3 x∗2 = 2/3 with the optimal value z ∗ = x∗1 + x∗2 = 10/3. This means that you have already solved the dual — or primal, if you take the opposite point of view — of the Example 8.2.26. Now, how can the solution above, combined with the Complementary Slackness Theorem 8.2.20, help you to solve the LP of Example 8.2.26? Here is how: First note that x∗1 > 0 and x∗1 > 0. So, the Complementary Slackness Theorem 8.2.20 tells us that the optimal solution y∗ = [y1∗ y2∗ y3∗ ]0 of

Dual Problem

135

the LP in Example 8.2.26 must have zero excesses. So, the inequalities in 8.2.26 are actually equalities at the optimum. Also, if we check the optimum x∗ in the first three constraints of the maximum problem, we find equalities in the first two of them, and a strict inequality in the third one. So, the Complementary Slackness Theorem 8.2.20 tells us that y3∗ = 0. So, in the optimum y∗ of the LP in Example 8.2.26 we must have y1∗ + 4y2∗ 2y1∗ + 2y2∗

y3∗

= 1 = 1 = 0

But this is a very easy system to solve. We obtain y1∗ = 1/3, y2∗ = 1/6, y3∗ = 0 with the optimal value w∗ = 4y1∗ + 12y2∗ + y3 = 10/3.

Primal and Dual Sensitivity

8.3

136

Primal and Dual Sensitivity

By now it should be clear — although we have not stated it explicitly — what is the connection between sensitivity and duality. Let us be explicit. To put it shortly, it is:

x∗primal πprimal s∗primal uprimal ∗ zprimal

= = = = =

πdual, ∗ ydual , udual, e∗dual, ∗ wdual .

In the above uprimal and udual denote the vectors of reduced costs in the primal and dual, respectively. Also, s∗primal and e∗dual denote the slacks and excesses of the primal and dual at optimum, respectively. All the other notations should be clear. 8.3.1 Remark. The equality ∗ πprimal = ydual

explains the name “shadow prices”. Indeed, the dual is a “shadow problem”. So, the shadow prices of the constraints at the primal optimum are the prices of the dual variables (that are related to the constraints) at the dual optimum. Sometimes the shadow prices are called the dual prices.

Part III

Applications of Linear Programming

Chapter 9

Data Envelopment Analysis

In this chapter we discuss how to apply LP in the problem of evaluating the relative efficiency of different units, relative only to themselves. This is a nice application because of three reasons: 1. it is not at all obvious that LP can be used in this problem, 2. the application gives valuable insight to the LP duality, 3. the application itself is extremely useful. This chapter is adapted from [2, Ch. 2] and from J.E. Beasley’s web notes.

9.1

Graphical Introduction to Data Envelopment Analysis

Data Envelopment Analysis and Decision-Making Units Data envelopment analysis (DEA), occasionally called frontier analysis, was introduced by Charnes, Cooper and Rhodes in 1978 (CCR). DEA is a performance measurement technique which can be used for evaluating the relative efficiency of decision-making units (DMUs). Here DMU is an abstract term for an entity that transforms inputs into outputs. The term is abstract on purpose: Typically one thinks of DMUs as manufacturers of some goods (outputs) who use some resources (inputs). This way of thinking, while correct, is very narrow-minded: DMU is a much more general concept, and DEA can be applied in very diverse situations. Indeed, basically the DMUs can be pretty much anything. The main restrictions are: 1. the DMUs have the same inputs and outputs, 2. the DMUs’ inputs and outputs can be measured numerically. Indeed, if either one of the points 1. or 2. above fails, it would not make any sense to compare the DMUs quantitatively, i.e., with numbers. Examples of DMUs to which DEA has been applied are: • banks,

Graphical Introduction to Data Envelopment Analysis

• • • • • • •

139

police stations, hospitals, tax offices, prisons, military defence bases, schools, university departments.

9.1.1 Remark. There are two points of DEA that must be emphasized: 1. DEA is a data oriented. This means that it will only use the data related to the inputs and outputs of the DMUs under consideration. It does not use any extra theoretical — or practical, or philosophical — knowledge. In this respect, it differs from classical comparison methods where DMUs are compared either to a “representative” DMU or to some “theoretically best” DMU. 2. DEA is an extreme point method. It does not compare DMUs to any “representative” or “average” DMU. No, DEA compares the different DMUs to the “best” DMU. Relative Efficiency DEA is about comparing the relative efficiency of DMUs. Efficiency is defined by the following meta-mathematical formula: efficiency =

outputs . inputs

There is nothing relative in the definition of efficiency above. Well, not as such. The relativity comes in later. In the next subsection we will illustrate DEA by means of a small example of Kaupþing Bank branches. Note here that much of what you will see below is a graphical approach to DEA. This is very useful if you are attempting to explain DEA to those less technically qualified — such as many you might meet in the management world. There is a mathematical approach to DEA that can be adopted however — this is illustrated later in the following sections.

Graphical Introduction to Data Envelopment Analysis

140

One Input — One Output

9.1.2 Example. Consider a number of Kaupþing Bank’s branches. For each branch we have a single output measure: Number of personal transactions completed per week. Also, we have a single input measure: Number of staff. The data we have is as follows: Branch Reykjavík Akureyri Kópavogur Hafnarfjörður

Personal transactions 125 44 80 23

Number of staff 18 16 17 11

How then can we compare these branches — or DMUs — and measure their performance using this data? A commonly used method is ratios, which means that we will compare efficiencies. For our Kaupþing Bank branch example 9.1.2 we have a single input measure, the number of staff, and a single output measure, the number of personal transactions. Hence the meta-mathematical formula efficiency =

outputs inputs

=

output input

is a well-defined mathematical formula — no metas involved. We have: Branch Reykjavík Akureyri Kópavogur Hafnarfjörður

Personal transactions per staff member 6.94 2.75 4.71 2.09

Here we can see that Reykjavík has the highest ratio of personal transactions per staff member, whereas Hafnarfjörður has the lowest ratio of personal transactions per staff member. So, relative to each others, Reykjavík branch is the best (most efficient), and the Hafnarfjörður branch is the worst (least efficient). As Reykjavík branch is the most efficient branch with the highest ratio of 6.94, it makes sense to compare all the other branches to it. To do this we calculate their relative efficiency with respect to Reykjavík branch: We divide

Graphical Introduction to Data Envelopment Analysis

141

the ratio for any branch by the Reykjavík’s efficiency 6.94, and multiply by 100% (which is one) to convert to a percentage. This gives: Branch Reykjavík Akureyri Kópavogur Hafnarfjörður

Relative Efficiency 100% × (6.94/6.94) = 100% 100% × (2.75/6.94) = 40% 100% × (4.71/6.94) = 68% 100% × (2.09/6.94) = 30%

The other branches do not compare well with Reykjavík. That is, they are relatively less efficient at using their given input resource (staff members) to produce output (number of personal transactions). 9.1.3 Remark. We could, if we wish, use the comparison with Reykjavík to set targets for the other branches: 1. We could set a target for Hafnarfjörður of continuing to process the same level of output but with one less member of staff. This is an example of an input target as it deals with an input measure. 2. An example of an output target would be for Hafnarfjörður to increase the number of personal transactions by 10% — e.g. by obtaining new accounts. We could, of course, also set Hafnarfjörður a mix of input and output targets which we want it to achieve. One Input — Two Outputs Typically we have more than one input and one output. In this subsection we consider the case of one input and two outputs. While the case of one input and one output was almost trivial, the case of two outputs and one input is still simple enough to allow for graphical analysis. The analog with LPs would be: LPs with one decision variable are trivial, and LPs with two decision variables are still simple enough to allow for graphical analysis. Let us extend the Kaupþing Bank branch example 9.1.2:

Graphical Introduction to Data Envelopment Analysis

142

9.1.4 Example. Consider a number of Kaupþing Bank’s branches. For each branch we have a two output measures: Number of personal transactions completed per week, and number of business transaction completed per week. We have a single input measure: Number of staff. The data we have is as follows: Branch

Reykjavík Akureyri Kópavogur Hafnarfjörður

Personal transactions

Business transactions

Number of staff

125 44 80 23

50 20 55 12

18 16 17 11

How now can we compare these branches and measure their performance using this data? As before, a commonly used method is ratios, just as in the case considered before of a single output and a single input. Typically we take one of the output measures and divide it by one of the input measures. For our bank branch example 9.1.4 the input measure is plainly the number of staff (as before) and the two output measures are number of personal transactions and number of business transactions. Hence we have the two ratios: Branch

Reykjavík Akureyri Kópavogur Hafnarfjörður

Personal transactions per staff member

Business transactions per staff member

6.94 2.75 4.71 2.09

2.78 1.25 3.24 1.09

Here we can see that Reykjavík has the highest ratio of personal transactions per staff member, whereas Kópavogur has the highest ratio of business transactions per staff member. So, it seems that Reykjavík and Kópavogur are the best performers. Akureyri and Hafnarfjörður do not compare so well with Reykjavík and Kópavogur. That is, they are relatively less efficient at using their given input resource (staff members) to produce outputs (personal and business transactions). One problem with comparison via ratios is that different ratios can give a different picture and it is difficult to combine the entire set of ratios into a single numeric judgement. For example, consider Akureyri and Hafnarfjörður: • Akureyri is 2.75/2.09 = 1.32 times as efficient as Hafnarfjörður at personal transactions, • Akureyri is 1.25/1.09 = 1.15 times as efficient as Hafnarfjöður at business transactions.

Graphical Introduction to Data Envelopment Analysis

143

How would you combine these figures — 1.32 and 1.15 — into a single judgement? This problem of different ratios giving different pictures would be especially true if we were to increase the number of branches or increase the number of inputs or outputs. 9.1.5 Example. We ad five extra branches, Sellfoss, Hveragerði, Akranes, Borgarnes, and Keflavík, to Example 9.1.4. The data is now: Branch

Personal transactions per staff member

Business transactions per staff member

6.94 2.75 4.71 2.09 1.23 4.43 3.32 3.70 3.34

2.78 1.25 3.24 1.09 2.92 2.23 2.81 2.68 2.96

Reykjavík Akureyri Kópavogur Hafnarfjörður Sellfoss Hveragerði Akranes Borgarnes Keflavík

What can be now said about the efficiencies of the branches?

One way around the problem of interpreting different ratios, at least for problems involving just two outputs and a single input, is a simple graphical analysis. Suppose we plot the two ratios for each branch as shown below. In the picture we have no ticks to express the scale. The ticks are left out on purpose: DEA is about relative efficiency. So, the scales do not matter. BT/S

Kopavogur Sellfoss

Keflavik Akranes

Borgarnes

Reykjavik

Hveragerdi Akureyri Hafnarfjordur

PT/S

The positions of Reykjavík and Kópavogur in the graph demonstrate that they are superior to all other branches: They are the extreme points, other DMUs

Graphical Introduction to Data Envelopment Analysis

144

are inferior to them. The line drawn in the picture is called the efficient frontier. It was drawn by taking the extreme points, and then connecting them to each others and to the axes. That was a very vague drawing algorithm, but I hope you got the picture. 9.1.6 Remark. The name Data Envelopment Analysis arises from the efficient frontier that envelopes, or encloses, all the data we have. 9.1.7 Definition. We say that any DMU on the efficient frontier is 100% efficient. In our Kaupþing Bank branch examples 9.1.4 and 9.1.5, Reykjavík and Kópavogur branches are 100% efficient. This is not to say that the performance of Reykjavík or Kópavogur could not be improved. It may, or may not, be possible to do that. However, we can say that, based on the evidence provided by the different branches, we have no idea of the extent to which their performance can be improved. 9.1.8 Remark. It is important to note here that: • DEA only gives relative efficiencies, i.e., efficiencies relative to the data considered. It does not — and cannot — give absolute efficiencies. • No extra information or theory was used in determining the relative efficiencies of the DMUs. What happened was that we merely took data on inputs and outputs of the DMUs we considered, and presented the data in a particular way. • The statement that a DMU is 100% efficient simply means that we have no other DMU that can be said to be better than it. Now we know when a DMU is 100% efficient: A DMU is 100% efficient if it is on the efficient frontier. How about the non-efficient DMUs? Can we associate the DMUs that are not in the efficient frontier with a number representing their efficiency? We can. How to do this, is explained below. So, we will now discuss about quantifying efficiency scores for inefficient DMUs. Let us take Hafnarfjörður as an example of a non-efficient branch. We can see that, with respect to both of the ratios Reykjavík — and Kópavogur, too — dominates Hafnarfjörður. Plainly, Hafnarfjörður is less than 100% efficient. But how much? Now, Hafnarfjörður has • • • • •

number of staff 11, personal transactions 23, personal transactions per staff member 23/11 = 2.09, business transactions 12, business transactions per staff member 12/11 = 1.09.

Graphical Introduction to Data Envelopment Analysis

145

For Hafnarfjörður we have the ratio personal transactions business transactions

=

23 12

= 1.92.

This means that there are 1.92 personal transactions for every business transaction. This figure of 1.92 is also the ratio personal transactions per staff member . business transactions per staff member Indeed, personal transactions business transactions personal transactions number of staff members = × business transactions number of staff members personal transactions × number of staff members = business transactions × number of staff members personal transactions / number of staff members = business transactions / number of staff members personal transactions per staff member . = business transactions per staff member

BT/S

This number, 1.92, is the business mix of the Hafnarfjörður branch. It can be also be interpreted that Hafnarfjörður branch weighs its outputs, personal transactions and business transactions, so that personal transactions get weight 1.92 and business transactions get weight 1. Consider now the diagram below. In the diagram we have removed all the inefficient branches, except Hafnarfjörður. The line with the percentage ruler attached drawn through Hafnarfjörður represent all the possible — or virtual, if you like — branches having the same business mix, 1.92, as Hafnarfjörður. Kopavogur

Best Reykjavik

Hafnarfjordur

36% PT/S

Graphical Introduction to Data Envelopment Analysis

146

Note the virtual branch Best in the picture above. Best represents a branch that, were it to exist, would have the same business mix as Hafnarfjörður and would have an efficiency of 100%. Now, since Best and Hafnarfjörður have the same business mix, it makes sense to compare them numerically. Here is how to do it: Hafnarfjörður’s relative position in the ruler line from the worst branch with the same business mix (the origin) to the best branch with the same business mix (Best) is 36%. In other words, 36% of the ruler line is before Hafnarfjörður, and 64% of the ruler line is after Hafnarfjörður. So, it makes sense to say that Hafnarfjörður is, relative to the best branches, 36% efficient — or 64% inefficient, if you like. So, given the graphical consideration above we have the following definition for the (relative) efficiency of a DMU with two outputs and one input: 9.1.9 Definition. Draw a line segment from the origin through the DMU in question until you hit the efficient frontier. The efficiency of the DMU is length of the line segment from the origin to the DMU × 100%. total length of the line segment

9.1.10 Remark. The picture — and the definition — above is relative: You can change the scale of either the PT/S or the BT/S axis, or even switch the axes, but the relative efficiency of the Hafnarfjörður branch — or any other branch — won’t change. In the next picture we have written the relative efficiencies of the DMUs (Kaupþing Bank’s branches). BT/S

Kopavogur 100% Sellfoss 91%

Keflavik 92% Akranes 87%

Borgarnes 82%

Reykjavik 100%

Hveragerdi 74% Akureyri 43% Hafnarfjordur 36%

PT/S

The data of the picture above is written in tabular form below. Now it is up to you to decide which one of these two ways of presenting the data you

Graphical Introduction to Data Envelopment Analysis

147

prefer. The author prefers the picture, and therefore has huge respect for the “tabular-minded”. Branch

Personal transactions per staff member

Business transactions transactions per staff member

Relative efficiency

6.94 2.75 4.71 2.09 1.23 4.43 3.32 3.70 3.34

2.78 1.25 3.24 1.09 2.92 2.23 2.81 2.68 2.96

100% 43% 100% 36% 91% 74% 87% 82% 92%

Reykjavík Akureyri Kópavogur Hafnarfjörður Sellfoss Hveragerði Akranes Borgarnes Keflavik

9.1.11 Remark. Consider the picture with the ruler going from the origin through Hafnarfjörður to the efficient frontier. The point labelled Best on the efficient frontier is considered to represent the best possible performance that Hafnarfjörður can reasonably be expected to achieve. There are a number of ways by which Hafnarfjörður can move towards that point. 1. It can reduce its input (number of staff) while keeping its outputs (personal and business transaction) constant. This is an input target. 2. It can increase both its outputs, retaining the current business mix ratio of 1.92 while keeping its input (number of staff). This is an output target. 3. It can do some combination of the above. 9.1.12 Remark. It is important to be clear about the appropriate use of the (relative) efficiencies we have calculated. Here we have, e.g., • Reykjavík 100%, • Kópavogur 100%, • Hafnarfjörður 36%. This does not automatically mean that Hafnarfjörður is only approximately one-third as efficient as the best branches. Rather the efficiencies here would usually be taken as indicative of the fact that other branches are adopting practices and procedures which, if Hafnarfjörður were to adopt them, would enable it to improve its performance. This naturally invokes issues of highlighting and disseminating examples of best practice. Equally there are issues relating to the identification of poor practice. We end this subsection by further illustrating the relative nature of the DEA efficiencies. We shall ad two extra branches to Example 9.1.5 — Surtsey and Flatey — and see what happens.

Graphical Introduction to Data Envelopment Analysis

148

Let us start with Surtsey:

9.1.13 Example. Suppose we have an extra branch, Surtsey, added to the branches of Example 9.1.5. Assume that Surtsey has 1. 1 personal transactions per staff member, and 2. 6 business transactions per staff member. What changes in the efficiency analysis as a result of including the extra branch Surtsey?

BT/S

(There are actually no Kaupþing Bank branches in Surtsey. There are no people in Surtsey: People are not allowed in the Fire-Demon’s island. There are only puffins is Surtsey.) The effect of including Surtsey to the graphical Data Envelopment Analysis can be seen in the next picture: Surtsey

Kopavogur Sellfoss

Keflavik Akranes

Borgarnes

Reykjavik

Hveragerdi Akureyri Hafnarfjordur

PT/S

Graphical Introduction to Data Envelopment Analysis

149

Note that the efficient frontier now excludes Kópavogur. We do not draw that efficient frontier from Reykjavík to Kópavogur and from Kópavogur to Surtsey for two reasons: 1. Mathematically the efficient frontier must be convex, 2. although we have not seen any branches on the line from Reykjavík to Surtsey it is assumed in DEA that we could construct virtual branches, which would be a linear combination of Reykjavík and Surtsey, and which would lie on the straight line from Reykjavík to Surtsey. Also, note that the relative efficiencies of all the inefficient branches have changed. We have not calculated the new relative efficiencies. In the above it is clear why Reykjavík and Surtsey have a relative efficiency of 100% (i.e. are efficient): Both are the top performers with respect to one of the two ratios we are considering. The example below, where we have added the Flatey branch, illustrates that a branch can be efficient even if it is not a top performer. In the diagram below Flatey is efficient since under DEA it is judged to have “strength with respect to both ratios”, even though it is not the top performer in either.

9.1.14 Example. Suppose we have an extra branch, Flatey, added to the branches of Example 9.1.5. Assume that Flatey has 1. 5 personal transactions per staff member, and 2. 5 business transactions per staff member. What changes as a result of this extra branch being included in the analysis?

Here is the new picture with Flatey added. Note that Flatey is on the efficient frontier, i.e., 100% efficient, but it is not at top performer in either of the criteria “personal transactions per staff” (PT/S) or “business transactions per staff” (BT/S).

BT/S

Graphical Introduction to Data Envelopment Analysis

150

Surtsey

Flatey

Kopavogur Sellfoss

Keflavik Akranes

Borgarnes

Reykjavik

Hveragerdi Akureyri Hafnarfjordur

PT/S

Multiple Input — Multiple Output In our simple examples 9.1.4, 9.1.5, 9.1.13, and 9.1.14 of the Kaupþing Bank branches we had just one input and two outputs. This is ideal for a simple graphical analysis. If we have more inputs or outputs then drawing simple pictures is not possible without sculptures. However, it is still possible to carry out exactly the same analysis as before, but using mathematics rather than pictures. In words DEA, in evaluating any number of DMUs, with any number of inputs and outputs: 1. requires the inputs and outputs for each DMU to be specified, 2. defines efficiency for each DMU as a weighted sum of outputs divided by a weighted sum of inputs, where 3. all efficiencies are restricted to lie between zero and one (i.e. between 0% and 100%), 4. in calculating the numerical value for the efficiency of a particular DMU weights are chosen so as to maximize its efficiency, thereby presenting the DMU in the best possible light.

Graphical Introduction to Data Envelopment Analysis

151

How to carry out the vague four-point list presented above is the topic of the next section 9.2.

Charnes–Cooper–Rhodes Model

9.2

152

Charnes–Cooper–Rhodes Model

Now we consider mathematically what we have considered graphically in Section 9.1. We consider n Decision Making Units (DMUs). We call them unimaginatively as DMU1 , DMU2 , DMU3 , and so on upto DMUn . We are interested in assigning a measure of relative efficiency for each DMU without resorting to any other data than the one provided by the inputs and output of the DMUs themselves. Data Envelopment Analysis with Matrices We assume that each DMU has m inputs and s outputs. So, the m inputs of the DMUk are x1k x•k = ... . xmk In the same way the s outputs of the DMUk y1k .. y•k = . ysk

are .

If we collect the inputs and the outputs into single matrices input matrix x11 · · · x1n .. .. .. X = [xjk ] = [x•1 · · · x•n ] = . . . xm1 · · ·

xmn ,

we have the

and the output matrix y11 · · · .. .. · · · y•n ] = . . ys1 · · ·

Y = [yik ] = [y•1

So, xjk = the input j of the DMUk , yik = the output i of the DMUk .

y1n .. . . ysn

Charnes–Cooper–Rhodes Model

153

Charnes–Cooper–Rhodes Fractional Program Given what we have learnt it seems reasonable to measure the (relative) efficiency of the DMUs as weighted sums. So, let u1 u = ... us

be the weights associated with the s outputs of the DMUs. Similarly, let v1 v = ... vm

be the weights associated with the inputs of the DMUs. Then the weighted efficiency, with weights u and v , of any DMU, say DMUo (o for DMU under Observation) is

(9.2.1)

ho (u, v) = the (u, v) weighted efficiency of DMUo u weighted outputs of DMUo = v weighted inputs of DMUo Ps j=1 uj yjo = Pm i=1 vi xio 0 u y•o = . v0 x•o

9.2.2 Example. Consider the Kaupþing Bank’s branches of Example 9.1.4 of the previous section: Branch

Reykjavík Akureyri Kópavogur Hafnarfjörður

Personal transactions

Business transactions

Number of staff

125 44 80 23

50 20 55 12

18 16 17 11

Denote the data of Example 9.2.2 by x1• = number of staff, y1• = number of personal transactions, y2• = number of business transactions.

Charnes–Cooper–Rhodes Model

154

So, e.g., x1• is the 4-dimensional row vector consisting of the number of staff data for the DMUs Reykjavík, Akureyri, Kópavogur, and Hafnarfjörður. Similarly, y1• and y2• are the 4-dimensional row vectors indicating the number of personal and business transactions for each of the four DMUs: Reykjavík (1), Akureyri (2), Kópavogur (3), and Hafnarfjörður (4). The output matrix for this example is: y1• Y = y2• y11 y12 y13 y14 = y21 y22 y23 y24 125 44 80 23 = . 50 20 55 12 The input matrix is X = x1• = [x11 x12 x13 x14 ] = [18 16 17 11] . Let us then take the Hafnarfjörður branch under consideration. So, DMUo = DMU4 is Hafnarfjörður. With our vector notation Hafnarfjörður would have the (weighted) efficiency ho (u, v) =

u1 y1o + u2 y2o v1 x1o

=

u1 ×23 + u2 ×12 . v1 ×11

Now there is the problem of fixing the weight u and v of the outputs and the inputs. Each DMU would — of course — want to fix the weights u and v in such a way that they would look best in comparison with the other DMUs. So, it is in the interests of each and every one of the DMUs to maximize the weighted efficiency ho (u, v). In particular, this means that Hafnarfjörður faces an optimization problem (9.2.3)

max ho (u, v) = u,v

max

u1 ,u2 ,v1

u1 ×23 + u2 ×12 . v1 ×11

Obviously there must be constraints to the decision variables u and v . Indeed, otherwise the optimization problem (9.2.3) would yield an unbounded optimum. So, what are the constraints? Well, obviously we have the sign constraints u, v ≥ 0. This does not help too much yet, though. The optimum of (9.2.3) is still unbounded. Now, we remember that we are dealing with efficiencies. But, an efficiency is a number between 0% and 100%. So, we have the constraint ho (u, v) ≤ 1.

Charnes–Cooper–Rhodes Model

155

This does not help too much either. Indeed, now the optimum for (9.2.3) would be 1, or 100%. But we are close now. Remember that the efficiency is always a number between 0% and 100%. So, the efficiencies of the other DMUs must also be between 0% and 100%. So, we let Hafnarfjörður set the weights u and v , and the other DMUs are then measured in the same way. So, the constraints are hk (u, v) ≤ 1 for all DMUk , k = 1, . . . , n. Collecting what we have found above we have found the fractional form of the Charnes–Cooper–Rhodes (CCR) model 9.2.4 Definition. The CCR Fractional Program for DMUo relative to DMU1 , . . . , DMUn is (9.2.5)

u0 y•o v0 x•o u0 y•k ≤1 v0 x•k u, v ≥ 0.

max θ = s.t.

for all k = 1, . . . , n

The figure θ is the DMU0 ’s DEA Efficiency. Charnes–Cooper–Rhodes Linear Program Consider the optimization problem in Definition 9.2.4. This is not an LP. But the name of this part of the Chapters is “Applications of Linear Programming”. So, it seems that we have a misnomer! Also, we do not know how to solve fractional programs like the (9.2.5) in Definition 9.2.4. Fortunately there is a way of transforming the fractional program (9.2.5) into an LP. Before going to the LP let us note that while the efficiency score θ of the CCR fractional program (9.2.5) is unique, there are many different weights u, v that give the same efficiency. Indeed, if the weights u, v give the optimal efficiency, then so do the weights αu, αv for any α > 0. This is due to the fact that we are dealing with ratios. Indeed, for any α > 0 u0 y•o v0 x•o αu0 y•o = αv0 x•o = ho (αu, αv).

ho (u, v) =

There is an easy way out, however. We just normalize the denominator in the ratio, i.e., we insist that v0 x•o = 1. Now we are ready to give the LP formulation of the fractional program 9.2.5:

Charnes–Cooper–Rhodes Model

156

9.2.6 Definition. The CCR LP for DMUo relative to DMU1 , . . . , DMUn is (9.2.7)

max θ = u0 y•o s.t.

v0 x•o = 1, u0 Y ≤ v0 X u, v ≥ 0.

The figure θ is the DMUo ’s DEA Efficiency. It may not be immediately clear why the LP (9.2.7) is the same optimization problem as the fractional program (9.2.5). So, we explain a little why this is so. Consider the fractional program (9.2.5). First, note that the extra assumption v0 x•o = 1 does not change the optimal value of θ in the fractional program. Indeed, we have already seen that this restriction merely chooses one optimal choice among many. Next note that in the LP (9.2.7) we have θ = u0 y•o , while in the fractional program (9.2.5) we have θ =

u0 y•o . v0 x•o

Remember, that we have now the normalizing assumption v0 x0 = 1. So, we see that the fractional and linear objectives are actually the same. Finally, let us look the constraints u0 y•k ≤ 1 v0 x•k of the fractional program (9.2.5) and compare them to the constraints u0 y•k

≤ v0 x•k

of the linear program (9.2.7) (written in the matrix form there). If you multiply both sides of the fractional programs constraints by v0 x•k you notice that these constraints are actually the same. So, we see that the fractional program (9.2.5) and the linear program (9.2.7) are the same. Efficiency for Hafnarfjörður and Reykjavík Let us calculate mathematically, as opposed to graphically, Hafnarfjörður’s efficiency and Reykjavík’s efficiency in Example 9.2.2 by using the CCR LP (9.2.7). Recall the data

Charnes–Cooper–Rhodes Model

Branch

157

Personal transactions

Business transactions

Number of staff

125 44 80 23

50 20 55 12

18 16 17 11

Reykjavík Akureyri Kópavogur Hafnarfjörður and the notation

x1• = number of staff, y1• = number of personal transactions, y2• = number of business transactions. 9.2.8 Remark. Note that x1• , y1• , y2• are not the decision variables. They are the data. The decision variables are the weights v1 , u1 , u2 associated with the data x1• , y1• , y2• . Here is the LP for Hafnarfjörður max θ = s.t.

23u1 11v1 125u1 44u1 80u1 23u1

+ 12u2 + 50u2 + 20u2 + 55u2 + 12u2 u1 , u2 , v1

= ≤ ≤ ≤ ≤ ≥

1 18v1 16v1 17v1 11v1 0

(DEA Efficiency) (Normalization) (DMU Reykjavik) (DMU Akureyri) (DMU Kopavogur) (DMU Hafnarfjordur)

Now, this LP is certainly not in standard form. We shall solve it with GLPK, however. So, there is no reason to transform it into a standard form.

Charnes–Cooper–Rhodes Model

158

Here is the GNU MathProg code for the Hafnarfjörður’s LP: # DEA efficiency for Hafnarfjordur # Decision variables var u1 >=0; # weight for personal transactions var u2 >=0; # weight for business transactions var v1 >=0; # weight for staff members # Hafnarfjordur’s DEA efficiency maximize theta: 23*u1 + 12*u2; # normalization constraint s.t. Norm: 11*v1 = 1; # constraints from the set of DMUs s.t. Reykjavik: 125*u1+50*u2 <= s.t. Akureyri: 44*u1+20*u2 <= s.t. Kopavogur: 80*u1+55*u2 <= s.t. Hafnarfjordur: 23*u1+12*u2 <=

18*v1; 16*v1; 17*v1; 11*v1;

end;

Here is the “Simplex section” of the glpsol report: No. -----1 2 3 4 5 6

Row name St Activity Lower bound Upper bound Marginal ------------ -- ------------- ------------- ------------- ------------theta B 0.361739 Norm NS 1 1 = 0.361739 Reykjavik NU 0 -0 0.106087 Akureyri B -0.826561 -0 Kopavogur NU 0 -0 0.121739 Hafnarfjordur B -0.638261 -0

No. -----1 2 3

Column name -----------u1 u2 v1

St Activity Lower bound Upper bound Marginal -- ------------- ------------- ------------- ------------B 0.00442688 0 B 0.0216601 0 B 0.0909091 0

We see that the DEA efficiency of Hafnarfjörður is 36%. This is no news to us. We learnt this in the previous section with graphical analysis. But it is nice to see that the graphical and the mathematical approach agree of the efficiency.

Charnes–Cooper–Rhodes Model

159

Let us then consider the Reykjavík branch in Example 9.2.2. Here is the GNU MathProg code for the Reykjavík branch: # DEA efficiency for Reykjavik # Decision variables var u1 >=0; # weight for personal transactions var u2 >=0; # weight for business transactions var v1 >=0; # weight for staff members # Reykjavik’s DEA efficiency maximize theta: 125*u1 + 50*u2; # normalization constraint s.t. Norm: 18*v1 = 1; # constraints from the set of DMUs s.t. Reykjavik: 125*u1+50*u2 <= s.t. Akureyri: 44*u1+20*u2 <= s.t. Kopavogur: 80*u1+55*u2 <= s.t. Hafnarfjordur: 23*u1+12*u2 <=

18*v1; 16*v1; 17*v1; 11*v1;

end;

And here is the “Simplex part” of the glpsol report: No. -----1 2 3 4 5 6

Row name St Activity Lower bound Upper bound Marginal ------------ -- ------------- ------------- ------------- ------------theta B 1 Norm NS 1 1 = 1 Reykjavik NU 0 -0 1 Akureyri B -0.536889 -0 Kopavogur B -0.304444 -0 Hafnarfjordur B -0.427111 -0

No. -----1 2 3

Column name -----------u1 u2 v1

St Activity Lower bound Upper bound Marginal -- ------------- ------------- ------------- ------------B 0.008 0 NL 0 0 < eps B 0.0555556 0

We see that the mathematical approach agrees with the graphical approach: Reykjavík is 100% efficient.

Charnes–Cooper–Rhodes Model’s Dual

9.3

160

Charnes–Cooper–Rhodes Model’s Dual

Finding the DEA efficiency of a DMUo in the CCR model is an LP (9.2.7). So, it must have a dual LP associated with it. In this section we explain how to construct the dual, and how to interpret it. Also, in the last subsection we illustrate how to find the DEA efficiencies of the Hafnarfjörður and Reykjavík branch of Example 9.2.2 by using the the dual of the CCR model. Finding Dual In this subsection we will find the dual of the CCR LP (9.2.7) by using brute force matrix calculus. We throw all intuition to the wind, and simply follow mathematical algorithms. In the next subsection we shall interpret the dual we find in this subsection. Recall the LP form of the CCR model: (9.3.1)

max θ = u0 y•o s.t.

v0 x•o = 1, u0 Y ≤ v0 X u, v ≥ 0.

To find the dual, we write the LP (9.3.1) in standard form. We could use the block matrix notation, but the derivation is probably easier to understand if we do not use the concise matrix notation. So, we abandon matrices in the derivation. Without matrices the CCR LP (9.3.1) can be written as

(9.3.2)

max θ = u1 y1o + · · · + us yso s.t. v1 x1o + · · · + vm xmo u1 y11 + · · · + us ys1 .. .

= 1 ≤ v1 x11 + · · · + vm xmo .. .. . .

u1 y1n + · · · + us ysn ≤ v1 x1n + · · · + vm xmn u1 , . . . , us , v1 , . . . , vm ≥ 0. The LP (9.3.2) is certainly not in standard form. Actually, it is pretty far from it. As a first step in transforming (9.3.2) into a standard form let us write it so that the decision variables u = [u1 · · · us ]0 and v = [v1 · · · vm ]0 are in the right places, i.e., coefficients are in front of the decision variables, all the decision variables are represented everywhere, and there are no decision

Charnes–Cooper–Rhodes Model’s Dual

variables in the RHSs. (9.3.3) max θ = y1o u1 s.t. 0u1 y11 u1 .. .

161

We obtain: + · · · + yso us +···+ 0us + · · · + ys1 us .. .

y1n u1 + · · · + ysn us u1 ... us

+0v1 + · · · + 0vm +x1o v1 + · · · + xmo vm −x11 v1 − · · · − xm1 vm .. .. .. . . . −x1n v1 − · · · − xmn vm v1 ... vm

= 1 ≤ 0 .. .. . . ≤ 0 ≥ 0

Next, we split the equality constraint in (9.3.3) into two ≤ inequalities. We obtain: (9.3.4) max θ = y1o u1 + · · · + yso us +0v1 + · · · + 0vm s.t. 0u1 + · · · + 0us +x1o v1 + · · · + xmo vm ≤ 1 −0u1 − · · · − −0us −x1o v1 − · · · − −xmo vm ≤ −1 y11 u1 + · · · + ys1 us −x11 v1 − · · · − xm1 vm ≤ 0 .. .. .. .. .. .. .. . . . . . . . y1n u1 + · · · + ysn us −x1n v1 − · · · − u1 ... us v1 ...

xmn vm ≤ vm ≥

0 0

Now it is pretty straightforward to transform the LP (9.3.3) into the dual. Let ϑ be the objective, and let µ = [µ1 µ2 µ3 · · · µn+2 ]0 be the decision variables. We obtain: min ϑ = s.t.

(9.3.5)

µ 1 − µ2 0µ1 − 0µ2 .. .

+y11 µ3 + · · · + .. .

0µ1 − 0µ2 x1o µ1 − x1o µ2 .. .

+ys1 µ3 + · · · + −x11 µ3 − · · · − .. .

y1n µn+2 ≥ y1o .. .. . . ysn µn+2 ≥ yso x1n µn+1 ≥ 0 .. .. . .

xmo µ1 − xmo µ2 −xm1 µ3 − · · · − xmn µn+1 ≥ µ1 , . . . , µn+2 ≥

0 0

We have found the dual (9.3.5) of the CCR LP (9.2.7). Unfortunately, this dual is not easy to interpret. So, we have to transform it slightly in order to understand what is going on. This is what we do in the next subsection. Interpreting Dual Let us substitute the objective ϑ = µ1 − µ 2

Charnes–Cooper–Rhodes Model’s Dual

162

into the constraints of (9.3.5). In doing so, we actually eliminate all the occurrences on µ1 and µ2 in the system. We obtain: min ϑ s.t.

y11 µ3 + · · · + .. .

(9.3.6)

x1o ϑ .. .

ys1 µ3 + · · · + −x11 µ3 − · · · − .. .

y1n µn+2 ≥ y1o .. .. . . ysn µn+2 ≥ yso x1n µn+1 ≥ 0 .. .. . .

xmo ϑ −xm1 µ3 − · · · − xmn µn+1 ≥ µ1 , . . . , µn+2 ≥

0 0

Next, we shall renumber the remaining decision variables. The new decision variables will be λ = [λ1 · · · λn ]0 , where λ1 = µ3 , λ2 = µ4 , . . . , λn = µn+2 . So, the LP (9.3.6) becomes min ϑ s.t.

+y11 λ1 + · · · + .. .

(9.3.7)

x1o ϑ .. .

+ys1 λ1 + · · · + −x11 λ1 − · · · − .. .

y1n λn ≥ y1o .. .. . . ysn λn ≥ yso x1n λn ≥ 0 .. .. . .

xmo ϑ −xm1 λ1 − · · · − xmn λn ≥ λ1 , . . . , λ n ≥

0 0

Finally, we reorganize the ≥ inequalities, for a reason that will become apparent later when we get to the interpretation. We obtain: min ϑ s.t.

(9.3.8)

y11 λ1 + · · · + .. .

y1n λn ≥ .. .

y1o .. .

ys1 λ1 + · · · + x11 λ1 + · · · + .. .

ysn λn ≥ x1n λn ≤ .. .

yso x1o ϑ .. .

xm1 λ1 + · · · + xmn λn ≤ xmo ϑ λ1 , . . . , λ n ≥ 0 We have found out a formulation of the dual of the CCR LP (9.2.7) that we can understand in a meaningful way: The dual variables λ1 λ = ... λn

Charnes–Cooper–Rhodes Model’s Dual

163

are the weights for a virtual DMU — denoted by DMUvirtual — that will be the reference point of the DMUo — the DMU under observation. The virtual DMU, DMUvirtual , is constructed from the actual DMUs — DMU1 , . . . , DMUn — by weighting the actual DMUk with weight λk : DMUvirtual =

n X

λk DMUk .

k=1

So, the the restrictions y11 λ1 + · · · + y1n λn ≥ y1o .. .. .. . . . ys1 λ1 + · · · + ysn λn ≥ yso say All the outputs of the virtual DMU are at least as great as the corresponding outputs of the DMU under observation. The restrictions x11 λ1 + · · · + .. .

x1n λn ≤ .. .

x1o ϑ .. .

xm1 λ1 + · · · + xmn λn ≤ xmo ϑ say If the inputs of the DMU under observation are scaled down by ϑ, then all the inputs are at least as great as the corresponding inputs of the virtual DMU. Finally, here is the CCR dual LP (9.3.8) in matrix form: 9.3.9 Definition. The CCR Dual LP for DMUo relative to DMU1 , . . . , DMUn is (9.3.10)

min ϑ s.t.

ϑx•o ≥ Xλ, Yλ ≥ y•o λ ≥ 0.

The figure ϑ is the DMUo ’s DEA Efficiency.

Charnes–Cooper–Rhodes Model’s Dual

164

Dual Efficiency for Hafnarfjörður and Reykjavík Let us see what are the (dual) DEA efficiencies for the Hafnarfjörður and Reykjavík in Example 9.2.2. We have already found out the solutions in previous sections by using the graphical approach and the primal CCR approach: Hafnarfjörður is 36% DEA efficient and Reykjavík is 100% DEA efficient. So, we shall now check if this third approach — the dual CCR approach — will give the same results, as it should. Here is the GNU MathProg code for the dual CCR LP (9.3.10) for Hafnarfjörður. Note that we have the objective ϑ as a variable. This is due to the fact that ϑ does not have an equation in the dual CCR LP (9.3.10). The solution of declaring ϑ as a decision variable and then equating it with the objective is not an elegant one, but it works. # Dual DEA efficiency for Hafnarfjordur # Decision variables var var var var var

lambda1 >=0; # weight for lambda2 >=0; # weight for lambda3 >=0; # weight for lambda4 >=0; # weight for theta; # objective has no

Reykjavik Akureyri Kopavogur Hafnarfjordur equation, so it is a variable

# Hafnarfjordur’s dual DEA efficiency minimize obj: theta; # Input constraints s.t. input1: 18*lambda1 + 16*lambda2 + 17*lambda3 + 11*lambda4 <= theta*11; # number of staff # Output constraints s.t. output1: 125*lambda1 + 44*lambda2 + 80*lambda3 + 23*lambda4 >= 23; # personal transactions s.t. output2: 50*lambda1 + 20*lambda2 + 55*lambda3 + 12*lambda4 >= 12; # business transactions end;

Charnes–Cooper–Rhodes Model’s Dual

165

Here is the relevant part of the glpsol report: No. -----1 2 3 4

Row name -----------obj input1 output1 output2

St Activity Lower bound Upper bound Marginal -- ------------- ------------- ------------- ------------B 0.361739 NU 0 -0 -0.0909091 NL 23 23 0.00442688 NL 12 12 0.0216601

No. -----1 2 3 4 5

Column name -----------lambda1 lambda2 lambda3 lambda4 theta

St Activity Lower bound Upper bound Marginal -- ------------- ------------- ------------- ------------B 0.106087 0 NL 0 0 0.826561 B 0.121739 0 NL 0 0 0.638261 B 0.361739

We see that the (dual) DEA efficiency for Hafnarfjörður is 36%, as it should. We can also read the composition of the virtual DMU associated with Hafnarfjörður: DMUvirtual = λ1 DMU1 + λ2 DMU2 + λ3 DMU3 + λ4 DMU4 = 10.6%×DMUReykjavik + 12.2%×DMUKopavogur . So, one way to interpret the result is: Consider the virtual DMU that is composed of 10.6% of Reykjavík and 12.2% of Kópavogur. Then the outputs of this virtual DMU are the same as those of Hafnarfjörður, but the virtual DMU uses only 36% of the inputs Hafnarfjörður uses.

Charnes–Cooper–Rhodes Model’s Dual

166

Here is the GNU MathProg code for Reykjavík: # Dual DEA efficiency for Reykjavik # Decision variables var var var var var

lambda1 >=0; # weight for lambda2 >=0; # weight for lambda3 >=0; # weight for lambda4 >=0; # weight for theta; # objective has no

Reykjavik Akureyri Kopavogur Hafnarfjordur equation, so it is a variable

# Reykjavik’s dual DEA efficiency minimize obj: theta; # Input constraints s.t. input1: 18*lambda1 + 16*lambda2 + 17*lambda3 + 11*lambda4 <= theta*18; # number of staff # Output constraints s.t. output1: 125*lambda1 + 44*lambda2 + 80*lambda3 + 23*lambda4 >= 125; # personal transactions s.t. output2: 50*lambda1 + 20*lambda2 + 55*lambda3 + 12*lambda4 >= 50; # business transactions end;

Here is the relevant part of the glpsol report: No. -----1 2 3 4

Row name -----------obj input1 output1 output2

St Activity Lower bound Upper bound Marginal -- ------------- ------------- ------------- ------------B 1 NU 0 -0 -0.0555556 NL 125 125 0.008 B 50 50

No. -----1 2 3 4 5

Column name -----------lambda1 lambda2 lambda3 lambda4 theta

St Activity Lower bound Upper bound Marginal -- ------------- ------------- ------------- ------------B 1 0 NL 0 0 0.536889 NL 0 0 0.304444 NL 0 0 0.427111 B 1

We see that Reykjavík is 100% (dual) DEA efficient. We also see that the virtual DMU associated with Reykjavík is Reykjavík itself. This is not surprising: Remember that the virtual DMU was a DMU that “is the same in outputs” but “uses less or equal inputs”. Since Reykjavík is 100% efficient, there should not be a virtual DMU that uses less inputs and produces the same outputs.

Strengths and Weaknesses of Data Envelopment Analysis

9.4

167

Strengths and Weaknesses of Data Envelopment Analysis

Data Envelopment Analysis is a very general framework that draws conclusions from the data available and makes very few — is any — assumptions of the context where the data came from. This is its main strength and this its main weakness. Strengths 1. DEA is simple enough to be modelled with LPs. 2. DEA can handle multiple input and multiple outputs. 3. DEA does not require an assumption of a functional form relating inputs to outputs. In particular, one does not have to think that the outputs are consequences of the inputs. 4. DMUs are directly compared against a peer or (virtual) combination of peers. 5. Inputs and outputs can have very different units. For example, output y1 could be in units of lives saved and input x1 could be in units of Euros without requiring an a priori trade-off between the two. Weaknesses 1. Since a standard formulation of DEA creates a separate LP for each DMU, large problems can be computationally intensive. 2. Since DEA is an extreme point technique, noise (even symmetrical noise with zero mean) such as measurement error can cause significant problems. 3. DEA is good at estimating relative efficiency of a DMU but it converges very slowly to “absolute” efficiency. In other words, it can tell you how well you are doing compared to your peers but not compared to a “theoretical maximum”. 4. Since DEA is a nonparametric technique, statistical hypothesis tests are difficult to apply in the DEA context. 5. DEA is very generous: If a DMU excels in just one output (or input) it is likely to get 100% efficiency, even if it performs very badly in all the other outputs (or inputs). So, if there are many outputs (or inputs) one is likely to get 100% efficiency for all the DMUs, which means that DEA cannot differentiate between the DMUs (it does not mean that all the DMUs are doing well in any absolute sense).

Chapter 10

Transportation Problems

In this chapter we shall briefly consider a set of so-called transportation problems that can be modelled as LPs, and thus solved with, say, the Simplex/Big M algorithm. There is, however, a specific structure in the models that allows us to use specialized algorithms that are much faster — and more pen-andpaper-friendly — than the Simplex/Big M algorithm. This chapter is adapted from [2, Ch. 3], [4, Ch. 6], and T. S. Ferguson’s web notes.

10.1

Transportation Algorithm

Transportation Problems as Linear Programs

10.1.1 Example. Frábært ehf. produces skyr. It has two production plants: P1 and P2 . Plant P1 produces 15 tons of skyr per day, and plant P2 produces 20 tons of skyr per day. All the produced skyr is shipped to the markets M1 , M2 , and M3 . The market M1 demands 17 tons of skyr, the market M2 demands 8 tons of skyr, and the market M3 demands 10 tons of skyr. To ship one ton of skyr from plant P1 to markets M1 , M2 , and M3 costs =C3, =C4, and =C6, respectively. To ship one ton of skyr from plant P2 to markets M1 , M2 , and M3 costs =C5, =C7, and =C5, respectively. Frábært ehf. wants to deliver its skyr to the markets with the minimal possible shipping cost while still meeting the market demands. How should Frábært ship its skyr?

The next picture illustrates the Frábært ehf.’s situation:

Transportation Algorithm

169

M3

5

P2

10

6

20

7 M2

4

8

5

P1

15 3 M1

17 The Frábært’s problem is a typical transportation problem, and it can be modelled as an LP. Let xij

= tons of skyr shipped from plant i to market j.

These are obviously the decision variables for Frábært. Everything else is fixed, and the only thing that is left open is the actual amounts transported from ports to markets. The objective of any transportation problem is to minimize the total shipping cost (while meeting the market demand). So, Frábært’s objective is (10.1.2)

min z =

2 X 3 X

cij xij

i=1 j=1

where cij

= the cost of shipping one ton of skyr from plant i to market j.

Transportation Algorithm

170

The objective (10.1.2) is a linear objective: Sums are linear, and double-sums doubly so. What about the constraints then? There are of course the sign constraints xij

≥ 0

for all plants i and markets j.

Indeed, it would be pretty hard to transport negative amount of skyr. There are also the supply and demand constraints. Each plant Pi has only so many tons of skyr it can supply. So, if si is the supply limit for plant Pi then we have the supply constraints 3 X

xij

≤ si

for all plants i.

j=1

Each market demands so many tons of skyr, and according to the problem we are committed to meet the market demands. So, if dj is the demand for skyr in the market Mj then we have the demand constraints 2 X

xij

≥ dj

for all markets j.

i=1

Here is the LP for Frábært ehf.: (10.1.3) min z = 3x11 +4x12 +6x13 s.t. x11 +x12 +x13 x11

x12

x13

+5x21 +7x22 +5x23 x21 +x21

+x22 +x22

≤ +x23 ≤ ≥ ≥ +x23 ≥ xij ≥

15 20 17 8 10 0

The LP (10.1.3) can be solved by using, e.g., the Simplex/Big M method. There is, however, a much more efficient specialized algorithm for solving LPs of type (10.1.3). We shall learn this method later in this Chapter. For now, let us end this subsection by noting the form of a general transportation problem as an LP and note its dual LP: 10.1.4 Definition. The transportation problem with ports Pi , i = 1, . . . , m, markets Mj , j = 1, . . . , n, shipping costs cij , i = 1, . . . , m, j = 1, . . . , n, supplies sj , j = 1, . . . , m, and

Transportation Algorithm

171

demands di , i = 1, . . . , n is min z =

m X n X

cij xij

(total shipping cost)

i=1 j=1

subject to n X

xij

≤ si

for all i = 1, . . . , m,

(supply)

xij

≥ dj

for all j = 1, . . . , n,

(demand)

xij

≥ 0

for all i = 1, . . . , m, j = 1, . . . , n.

j=1 m X i=1

The dual LP of the transportation problem of Definition 10.1.4 is max w = (10.1.5)

n P

dj vj +

j=1

s.t.

m P

si ui

i=1

vj − ui ≤ cij ui , vj ≥ 0

for all i = 1, . . . , m and j = 1, . . . , n for all i = 1, . . . , m and j = 1, . . . , n

Balancing Transportation Problems Example 10.1.1 is balanced : Total supply equals total demand, i.e., m X

si =

i=1

n X

dj .

j=1

This is essential in transportation problems. Also, note that with balanced problems it follows that all the products will be shipped, and that the market demands will be met exactly. So, the supply and demand constraints will realize as equalities rather than inequalities. So, we can, for balanced problems, restate Definition 10.1.4 as 10.1.6 Definition. The transportation problem with ports Pi , i = 1, . . . , m, markets Mj , j = 1, . . . , n, shipping costs cij , i = 1, . . . , m, j = 1, . . . , n, supplies sj , j = 1, . . . , m, and demands di , i = 1, . . . , n is min z =

m X n X i=1 j=1

cij xij

(total shipping cost)

Transportation Algorithm

172

subject to n X

xij

= si

for all i = 1, . . . , m,

(supply)

xij

= dj

for all j = 1, . . . , n,

(demand)

xij

≥ 0

for all i = 1, . . . , m, j = 1, . . . , n.

j=1 m X i=1

What about non-balanced transportation problems then? There are two possible cases: More supply than demand In this case we can introduce an imaginary market called the dump. The dump has just enough demand so that you can — dump — you excess supply there. Shipping to dump is costless. This solution may seem silly to you, but remember that the objective was to minimize the transportation cost while meeting the market demand. So, you are allowed to lose your shipments, if it helps. More demand than supply In this case the problem is infeasible: You cannot meet the market demands if you do not have enough to supply. One can of course generalize the transportation problem so that if the supply does not meet the demand there is a (linear) penalty associated with each market whose demand is not satisfied. E.g., one could subtract value ! m X M dj − xij i=1

from the objective for each market j whose demand dj is not met. We shall not consider penalized transportation problems here, however. Specialized Algorithm for Transportation Problems The transportation algorithm is a tableau-dance with transportation tableaux. An empty transportation tableau is a tabular representation of the transport problem’s 10.1.4 data: P1 P2

M1 c11

M2 c12

c21

c22

···

Mn c1n c2n

s2 .. .

.. . Pm

s1

cm1

cm2 d1

cmn d2

···

sm dn

Transportation Algorithm

173

The Frábært’s transportation problem’s 10.1.1 tabular representation is

P1 P2

3

M1

4

5

M2

6

7 17

M3 15

5 8

20 10

The transportation algorithm fills the empty transportation tableau with the shipping schedule xij :

P1 P2

M1 c11 c21

x11 x21

M2 c12 c22

··· x12

Mn c1n c2n

x22

x1n x2n

.. . Pm

s1 s2 .. .

cm1

xm1

d1

cm2

xm2

d2

cmn ···

xmn

dn

sm

Transportation Algorithm

174

The general idea of the specialized Transportation Algorithm is the same as in the Simplex algorithm, viz. Meta-Step 1: Find a BFS. Meta-Step 2: Check for optimality. If solution is optimal, the algorithm terminates. Otherwise move to step 3. Meta-Step 3: Find a new, improved, BFS. Go back to Meta-Step 2. Next we explain the metas away from the Meta-Steps 1–3 above. Finding a BFS The first BFS can be found, e.g., by using the so-called NW Corner Method. The method is called such since the traditional way of choosing available squares goes from NW to SE. The method goes as follows: 1. Choose any available square, say (i0 , j0 ). Specify the shipment xi0 ,j0 as large as possible subject to the supply and demand constraints, and mark this variable. 2. Delete from consideration whichever row or column has its constraint satisfied, but not both. If there is a choice, do not delete a row (column) if it is the last row (resp. column) undeleted. 3. Repeat 1. and 2. until the last available square is filled with a marked variable, and then delete from consideration both row and column. Now we construct the first BFS for Example 10.1.1. The marked variables will be in parentheses, and the deleted squares will have 0. We start with the NW corner. So (i0 , j0 ) = (1, 1). We put there as a big number as possible. This means that we ship all the plant P1 ’s supply to market M1 , i.e., x11 = 15. Now there is nothing left in P1 . So, we must have x12 = x13 = 0. So, squares (1, 2) and (1, 3) get deleted — or, if you like — row 1 gets deleted. P1 P2

3 5

M1 (15) 17

4

M2 0

7 8

6

M3 0

5

15 20

10

Next we move south, since east is deleted. The biggest number we can now put to square (2, 1) is x21 = 2, since there is already 15 tons of skyr shipped to the market M1 that demands 17 tons. No rows or columns will get deleted because of this operation.

Transportation Algorithm

P1 P2

3 5

M1 (15) (2) 17

4

175

M2 0

7

6

M3 0

5

8

15 20

10

Next we move east to square (2, 2). There the biggest number we can put is x22 = 8 since that is the market demand for M2 . No rows or columns will be deleted because of this operation.

P1 P2

3 5

M1 (15) (2) 17

4

M2 0

7 8

(8)

6

M3 0

5

15 20

10

Finally, we move to the last free square (2, 3). It is obvious that x23 = 10. We get the first BFS:

P1 P2

3 5

M1 (15) (2) 17

4

M2 0

7 8

(8)

6 5

M3 0 (10) 10

15 20

Checking Optimality Given a BFS, i.e., a feasible shipping schedule xij , we shall use the Complementary Slackness Theorem 8.2.20 to check whether the BFS is optimal. This means finding dual variables ui and vj that satisfy xij > 0

implies that

vj − ui = cij .

One method of finding the dual variables ui and vj is to solve the equations vj − ui = cij for all (i, j)-squares containing marked variables. There are m + n − 1 marked variables, and so we have m + n − 1 equations with m + n unknowns. This means that one of the variables ui , vj can be fixed to 0, say. Some of the ui or vj may turn out to be negative, which as such is not allowed in the dual problem, but this is not a problem. Indeed, one can always add a big enough constant to all the ui s and vj s without changing the values of vj − ui . Once the dual variables ui and vj are found we can check the optimality of the BFS by using the following algorithm:

Transportation Algorithm

176

1. Set one of the vj or ui to zero, and use the condition vj − ui = cij for squares containing marked variables to find all the vj and ui . 2. Check feasibility, vj − ui ≤ cij , for the remaining squares. If the BFS is feasible, it is optimal for the problem and its dual, due to the Complementary Slackness Theorem 8.2.20 Let us then find the dual variables ui and vj for the BFS Frábært’s problem 10.1.1. The next tableau is the BFS we found with the, yet unknown, dual variables in their appropriate places. 3

u1

5

u2

v1

4

(15)

v2

7

(2) 17

6

0 8

5

(8)

v3 0 (10) 10

15 20

To solve ui and vj for the marked variables we put u2 = 0. Remember we can choose any one of the ui or vj be zero. We chose u2 because then we see immediately thatv1 = 5, v2 = 7, and v3 = 5. As for the last unknown u1 , we have u1 = v1 − c11 = 5 − 3 = 2. So, we have the following BFS with the duals

2 0

3 5

5 (15) (2) 17

4

7 0

7 8

(8)

6 5

5 0 (10) 10

15 20

Now we check the remaining squares. For the BFS to be optimal we must have vj − ui ≤ cij . We see that this is not the case. The culprit is the square (1, 2): v2 − u1 = 7 − 2 = 5 > 4 = c12 . This means that we must improve our BFS. Improvement Routine Now we have a BFS that is not optimal. So, we must have a square (i0 , j0 ), say, for which vj0 − ui0

> ci0 j0 .

Transportation Algorithm

177

We would like to ship some amount of skyr, say, from the port Pi0 to the market Mj0 . The current amount xi0 j0 = 0 will be changed to a new amount denoted by ∆. But if we change xi0 j0 to ∆ we must subtract and add ∆ to other squares containing marked variables. This means that we are looking forward to a new BFS P1 P2

M1 M2 3 4 −∆ (15) +∆ 0 5 7 +∆ (2) −∆ (8) 17 8

6 5

M3 15

0 (10) 10

20

Now we choose the change ∆ to be as big as possible bearing in mind that the shipments cannot be negative. This means that ∆ will be the minimum of the xij s in the squares we are subtracting ∆. We see that the biggest possible change is ∆ = 8, which makes the the shipment x22 zero. 10.1.7 Remark. It may turn out that that we have ∆ = 0. This means that the value of the objective won’t change. However, the shipping schedule and the marked variables will change. While the new shipping schedule is no better than the old one, one can hope that from this new shipping schedule one can improve to a better one. Now we have the new BFS 3

P1

5

P2

M1

4

(7)

M2 (8)

7

(10) 17

6 5

0

8

M3 15

0

20

(10) 10

We have to go back to the previous step and check optimality for this BFS. So, we have to solve the dual variables ui and vj . We set now u2 = 0 which gives us immediately that v1 = 5 and v3 = 5. So, we find out that u1 = v1 − c11 = 2, and that v2 = u1 + c12 = 6. So, we have the BFS with dual variables 2 0

3 5

5 (7) (10) 17

4

6 (8)

7 8

0

6 5

5 0 (10) 10

15 20

Transportation Algorithm

178

This solutions passes the optimality test: vj − ui ≤ cij for all i and j . So, we have found an optimal shipping schedule. The cost associated with this shipping schedule can now be easily read from the tableau above: z =

2 X 3 X

cij x∗ij

i=1 j=1

= 3×7 + 4×8 + 5 ×10 + 5 ×10 = 153. The improvement algorithm can now be stated as 1. Choose any square (i, j) with vj − ui > cij . Set xij = ∆, but keep the constraints satisfied by subtracting and adding ∆ to appropriate marked variables. 2. Choose ∆ to be the minimum of the variables in the squares in which ∆ is subtracted. 3. Mark the new variable xij and remove from the marked variables one of the variables from which ∆ was subtracted that is now zero.

Assignment Problem

10.2

179

Assignment Problem

In this section we consider assignment problems that are — although it may not seem so at first sight — special cases of transportation problems. Assignment Problems as Linear Programs

10.2.1 Example. Machineco has four machines and four jobs to be completed. Each machine must be assigned to complete one job. The times required to set up each machine for completing each job are: Machine Machine Machine Machine

1 2 3 4

Job 1 14 2 7 2

Job 2 5 12 8 4

Job 3 8 6 3 6

Job 4 7 5 9 10

Machineco wants to minimize the total setup time needed to complete the four jobs. How can LP be used to solve Machineco’s problem?

The key point in modelling the Machineco’s problem 10.2.1 is to find out the decision variables — everything else is easy after that. So what are the decisions Machineco must make? Machineco must choose which machine is assigned to which job. Now, how could we write this analytically with variables? A common trick here is to use binary variables, i.e., variables that can take only two possible values: 0 or 1. So, we set binary variables xij , i = 1, . . . , 4, j = 1, . . . , 4, for each machine and each job to be xij

= 1 if machine i is assigned to meet the demands of job j,

xij

= 0 if machine i is not assigned to meet the demands of job j.

In other words, the variable xij is an indicator of the claim “Machine i is assigned to job j ”. Now it is fairly easy to formulate a program, i.e., an optimization problem, for Machineco’s problem 10.2.1. Indeed, the objective is to minimize the total

Assignment Problem

180

setup time. With our binary variables we can write the total setup time as z =

14x11 + 5x12 + 8x13 + 7x14 +2x21 + 12x22 + 6x23 + 4x24 +7x31 + 8x32 + 3x33 + 9x34 +2x41 + 4x42 + 6x43 + 10x44

Note that there will be a lot of zeros in the objective function above. What about the constraints for Machineco? First, we have to ensure that each machine is assigned to a job. This will give us the supply constraints x11 x21 x31 x41

+ + + +

x12 x22 x32 x42

+ + + +

x13 x23 x33 x43

+ + + +

x14 x24 x34 x44

= = = =

1 1 1 1

Second, we have to ensure that each job is completed, i.e., each job has a machine assigned to it. This will give us the demand constraints x11 x12 x13 x14

+ + + +

x21 x22 x23 x24

+ + + +

x31 x32 x33 x34

+ + + +

x41 x42 x43 x44

= = = =

1 1 1 1

So, putting the objective and the constraints we have just found together, and not forgetting the binary nature of the decisions, we have obtained the following program for Machineco’s problem 10.2.1: min z =

s.t. (10.2.2)

14x11 + 5x12 + 8x13 + 7x14 +2x21 + 12x22 + 6x23 + 4x24 +7x31 + 8x32 + 3x33 + 9x34 +2x41 + 4x42 + 6x43 + 10x44 x11 + x12 + x13 + x14 x21 + x22 + x23 + x24 x31 + x32 + x33 + x34 x41 + x42 + x43 + x44 x11 + x21 + x31 + x41 x11 + x21 + x31 + x41 x12 + x22 + x32 + x42 x13 + x23 + x33 + x43 x14 + x24 + x34 + x44 xij = 0 or xij = 1

= = = = = = = = =

1 1 1 1 1 1 1 1 1

(Machine)

(Job)

In (10.2.2) we have binary constraints xij = 0 or xij = 1 for the decision variables. So, at first sight it seems that the program (10.2.2) is not a linear

Assignment Problem

181

one. However, the structure of the assignment problems is such that if one omits the assumption xij = 0 or xij = 1, and simply assumes that xij ≥ 0, one will get on optimal solution where the decisions x∗ij are either 0 or 1. Hence, the program (10.2.2) is a linear one, i.e., it is an LP. Or, to be more precise, the program (10.2.2) and its linear relaxation min z =

s.t. (10.2.3)

14x11 + 5x12 + 8x13 + 7x14 +2x21 + 12x22 + 6x23 + 4x24 +7x31 + 8x32 + 3x33 + 9x34 +2x41 + 4x42 + 6x43 + 10x44 x11 + x12 + x13 + x14 x21 + x22 + x23 + x24 x31 + x32 + x33 + x34 x41 + x42 + x43 + x44 x11 + x21 + x31 + x41 x11 + x21 + x31 + x41 x12 + x22 + x32 + x42 x13 + x23 + x33 + x43 x14 + x24 + x34 + x44 xij

= = = = = = = = = ≥

1 1 1 1 1 1 1 1 1 0

(Machine)

(Job)

are equivalent. Equivalence of programs means that they have the same optimal decision and the same optimal objective value. Now, notice that the assignment LP (10.2.3) can be considered as a transportation problem. Indeed, consider the four machines as ports and the four jobs as markets. Then the supplies and demands are all ones. This interpretation gives us the following definition of an assignment problem: 10.2.4 Definition. An assignment problem is a transportation problem with equal amount of ports and markets, where the demands and supplies for each port and market are equal to one. Since assignment problems are LPs they can be solved as any LP with the Simplex/Big M method, say. Also, since they are transportation problems, they can be solved by using the specialized transportation algorithm. There is, however, an even more specialized algorithm for transportation problems: The so-called Hungarian method. Hungarian Method In the Hungarian method, the data of the assignment problem is presented in an n×n-table, where n is the number of ports, or markets, which is the same. For Example 10.2.1 the “Hungarian tableau” is

Assignment Problem

182

14 5

8

7

2

12 6

5

7

8

3

9

2

4

6

10

The Hungarian method is based on the following two observation: 1. The problem is solved if we can choose n squares from the “Hungarian tableau” so that: (a) Exactly one square is chosen from each row. (b) Exactly one square is chosen from each column. (c) The sum of the costs in the chosen n squares is the smallest possible. 2. If a same number is subtracted from all squares in a row, then the optimal selection of squares does not change. The same is true, if a same number is subtracted from all squares in a column. Here is the Hungarian Algorithm: Step 1 For each row, subtract the row minimum from each element in the row. Step 2 For each column, subtract the column minimum from each element in the column. Step 3 Draw the minimum number of lines — horizontal, vertical, or both — that are needed to cover all the zeros in the tableau. If n lines were required then the optimal assignment can be found among the covered zeros in the tableau, and the optimal cost can be read from the first tableau by summing up the number in the squares corresponding to the lined-out zeros. Step 4 Find the smallest non-zero element that is uncovered by lines. Subtract this element from each uncovered element, and add this element to each square that is covered with two lines. Return to Step 3. Here are steps 1 and 2 for Example 10.2.1:

14 5

8

7

9

0

3

2

9

0

3

0

2

12 6

5

0

10 4

3

0

10 4

1

7

8

3

9

4

5

0

6

4

5

0

4

2

4

6

10

0

2

4

8

0

2

4

6

Assignment Problem

183

Here are the remaining steps 3 and 4 for Example 10.2.1:

9

0

3

0

10 0

4

0

0

10 4

1

0

9

4

0

4

5

0

4

4

4

0

3

0

2

4

6

0

1

4

5

Now, we go back to Step 3, and line-out zeros:

10 0

4

0

0

9

4

0

4

4

0

3

0

1

4

5

Now we can read the optimal assignment from the covered zeros. First we note that the only covered zero in column 3 is in square (3, 3). So, we must have assignment x33 = 1. Also, in column 2 the only available zero is in square (1, 2). Consequently, x12 = 1. Now, as we can no longer use row 1 the only available zero in column 4 is in square (2, 4). So, x24 = 1. Finally, we choose x41 = 1. Next we read the corresponding setup cost from the first Hungarian tableau. We obtain setup cost = 5 + 5 + 3 + 2 = 15.

Transshipment Problem

10.3

184

Transshipment Problem

In this section we consider transportation problems. At first sight the transportation problems are network problems. However, one can think of them as generalizations of transportation problems. In that sense we are considering here problems that are opposite to assignment problems in the sense that assignment problems are special cases of transportation problems and transportation problems are special cases of transshipment problems. It turns out, however, that we can express transshipment problems as transportation problems. So the generalization from transportation problems to transshipment problems is mathematically no generalization at all. Transshipment Problems as Transportation Problems

10.3.1 Example. The Generic Company produces generic products and ships them to the markets M1 and M2 from the ports P1 and P2 . The products can be either shipped directly to the markets, or the Generic Company can use a transshipment point T1 . The ports P1 and P2 supply 150 and 200 units, respectively. The markets M1 and M2 demand both 130 units. The shipment costs from P1 to M1 is 25, from P1 to M2 it is 28, from P2 to M1 it is 26, and from P2 to M2 it is 22. The shipping cost from P1 to the transshipment point T1 is 8 and the shipping cost from P2 to the transshipment point T1 is 15. The shipping costs from the transshipment point T1 to the markets M1 and M2 are 16 and 17, respectively. The Generic Company wants to minimize its shipping costs while meeting the market demands. How should the Generic Company ship its products, if (a) the transshipment point T1 has infinite capacity, (b) only 100 products can be transshipped through T1 ?

Here is the data of Example 10.3.1 in a graph form.

Transshipment Problem

185

22

P2

200

M2

130

17 28 15 T1

∞/100 8 26 16 P1

150

25

M1

130

The general idea in modelling transshipment problems is to model them as transportation problems where the transshipment points are both ports and markets. To be more precise: • A transshipment problem is a transportation problem, where each transshipment point is both a port and a market. The shipping cost from a transshipment point to itself is, of course, zero. • If a transshipment point has capacity N , then N will be both demand and supply for that transshipment point. If the capacity is unlimited set the demand and supply to be the total supply of the system for that transshipment point. • Cost for shipping from transshipment points to the balancing dump is a very very very very big number M . This ensures that the excess supply to be dumped is dumped immediately.

Transshipment Problem

186

Solving Transshipment Problems Solution to Variant (a) of Example 10.3.1 In Example 10.3.1 the total supply is 350 and the total demand is 260. So, in order to balance the problem we need the dump market with demand 90. Let us denote the dump by D . The supply and demand for the transshipment point will be 350, so that everything can be, if necessary, transported via T1 . We also do not want to transship anything to the dump via the transshipment point T1 . So, we make the cost of shipping from T1 to D a very very very very very big number M . The direct dumping from P1 or P2 to D of costless, of course. So, we have the following initial transportation tableau

P1 P2 T1

8

T1

25

M1

28

M2

0

15

26

22

0

0

16

17

M

350

130

130

D 150 200 350 90

Now we can apply the transportation algorithm to this tableau. Here is a (greedy minimal cost) initial BFS

P1 P2 T1

8 15 0

T1 (150) (110) (90) 350

25

M1 0

26 16

0 (130) 130

28

M2 0

22 17

0 (130) 130

0 0 M

D 0 (90) 90

0

150 200 350

We leave it as an exercise to carry out the transportation algorithm for this BFS. We note the optimal transportation tableau:

P1 P2 T1

8 15 0

T1 (130) 0 (220) 350

25 26 16

M1 0 0 (130) 130

28 22 17

M2 0 (130) 130

0

0 0 M

D (20) (70) 90

0

150 200 350

Here is a graphical representation of the optimal transportation tableau above (the costs are in parentheses):

Transshipment Problem

187

130 (22)

P2

M2

130

200

T1

∞/100 130 (8)

130 (16)

P1

M1

150

130

From the picture we see that the total shipping cost is 130 × 8 + 130 × 16 + 130 × 22 = 5980. We cannot read the transshipment and dump data of the last Transportation Tableau from the picture above, but that data was auxiliary anyway, i.e. we might well not be interested in it. Solution to Variant (b) of Example 10.3.1 In Example 10.3.1 the total supply is 350 and the total demand is 260. So, in order to balance the problem we need the dump market with demand 90. Let us denote the dump by D . The supply and demand for the transshipment point will be 100, so that at most 100 units can be transported via T1 . We also do not want to transship anything to the dump via the transshipment point T1 . So, we make the cost of shipping from T1 to D a very very very very very big number M . The direct dumping from P1 or P2 to D of costless, of course. So, we have the following initial transportation tableau

Transshipment Problem

P1 P2 T1

8

T1

25

188

M1

28

M2

0

15

26

22

0

0

16

17

M

100

130

130

D 150 200 100 90

We leave it for the students to carry out the transportation algorithm, and simply note the solution:

P1 P2 T1

8 15

T1 (100) 0

0 100

0

25 26 16

M1 (30) 0 (100) 130

28 22 17

M2 0 (130) 130

0

0 0 M

D (20) (70) 90

0

150 200 100

Here is a graphical representation of the optimal transportation tableau above (the costs are in parentheses):

Transshipment Problem

189

130 (22)

P2

M2

130

200

T1

∞/100 100 (8)

100 (16)

P1

30 (25)

M1

130

150 From the picture we see that the total shipping cost is 100 × 8 + 100 × 16 + 30 × 25 + 130 × 22 = 6010.

We cannot read the transshipment and dump data of the last Transportation Tableau from the picture above, but that data was auxiliary anyway, i.e. we might well not be interested in it.

Part IV

Non-Linear Programming

Chapter 11

Integer Programming

In this chapter we lift the divisibility assumption of LPs. This means that we are looking at LPs where some (or all) of the decision variables are required to be integers. No longer can Tela Inc. from Example 2.1.1 in Chapter 2 manufacture 566.667 number of product #1. This chapter is adapted from [4, Ch. 8].

11.1

Integer Programming Terminology

Integer Programs’ Linear Relaxations Although the name Integer Program (IP) does not state it explicitly, it is assumed that IP is an LP with the additional requirement that some of the decision variables are integers. If the additional requirement that some of the decision variables are integers is lifted, then the resulting LP is called the LP relaxation of the IP in question. Pure Integer Programs An IP in which all the decision variables are required to be integers is called a Pure Integer Program, or simply an Integer Program (IP). For example,

(11.1.1)

max z = 3x1 + 2x2 s.t. x 1 + x2 ≤ 6 x1 , x2 ≥ 0, integer

is a pure integer program.

Branch-And-Bound Method

192

Mixed Integer Programs An IP in which some but not all of the decision variables are required to be integers is called a Mixed Integer Program (MIP). For example,

(11.1.2)

max z = 3x1 + 2x2 s.t. x 1 + x2 ≤ 6 x1 , x2 ≥ 0, x1 integer

is a pure integer program. Indeed, x1 is required to be an integer, but x2 is not.

11.2

Branch-And-Bound Method

In this section we provide a relatively fast algorithm for solving IPs and MIPs. The general idea of the algorithm is to solve LP relaxations of the IP and to look for an integer solution by branching and bounding on the decision variables provided by the the LP relaxations. Branch-And-Bound by Example Let us solve the following pure IP:

11.2.1 Example. The Integrity Furniture Corporation manufactures tables and chairs. A table requires 1 hour of labor and 9 units of wood. A chair requires 1 hour of labor and 5 units of wood. Currently 6 hours of labor and 45 units of wood are available. Each table contributes =C8 to profit, and each chair contributes =C5 to profit. Formulate and solve an IP to maximize Integrity Furniture’s profit.

Let x1 = number of tables manufactured, x2 = number of chairs manufactured. Since x1 and x2 must be integers, Integrity Furniture wishes to solve the following (pure) IP:

(11.2.2)

max z = 8x1 + 5x2 s.t. x 1 + x2 ≤ 6 9x1 + 5x2 ≤ 45 x1 , x2 ≥ 0, integer

(Labor) (Wood)

Branch-And-Bound Method

193

The first step in the branch-and-bound method is to solve the LP relaxation of the IP (11.2.2):

(11.2.3)

max z = 8x1 + 5x2 s.t. x 1 + x2 ≤ 6 9x1 + 5x2 ≤ 45 x1 , x2 ≥ 0

(Labor) (Wood)

If all the decision variables (x1 for tables and x2 for chairs in the example we are considering) in the LP relaxation’s optimum happen to be integers, then the optimal solution of the LP relaxation is also the optimal solution to the original IP. In the branch-and-bound algorithm, we call the LP relaxation (11.2.3) of the IP (11.2.2) subproblem 1 (SB 1). After solving SP 1 (graphically, say, cf. the next picture) we find the solution to be z = 165/4 x1 = 15/4 x2 = 9/4 This means that we were not lucky: The decision variables turned out to be fractional. So, the LP relaxation (11.2.3) has (possibly) a better optimum than the original IP (11.2.2). In any case, we have found an upper bound to the original IP: Integrity Furniture Corporation cannot possibly have better profit than =C165/4.

Branch-And-Bound Method

194

x2 6 LP relaxation’s feasible region IP feasible point

5

4

3 LP relaxation’s optimum

2 Isoprofit line z = 20

1

0

0

1

2

3

4

5

6 x1

Next we split the feasible region (painted light green in the previous picture) of the LP relaxation (11.2.3) in hope to find a solution that is an integer one. We arbitrarily choose a variable that is fractional at the optimal solution of the LP SP 1 (the LP relaxation). We choose x1 . Now, x1 was 15/4 = 3.75. Obviously, at the optimal solution to the IP we have either x1 ≤ 3 or x1 ≥ 4, since the third alternative 3 < x1 < 4 is out of the question for IPs. So, we consider the two possible cases x1 ≤ 3 and x1 ≥ 4 as separate subproblems. We denote these subproblems as SP 2 and SP 3. So, SP 2 = SP 1 + ”x1 ≥ 4”, SP 3 = SP 1 + ”x1 ≤ 3”. In the next picture we see that every possible feasible solution of the Integrity Furniture’s IP (the bullet points) is included in the feasible region of either SP 2 or SP 3. Also, SP 2 and SP 3 have no common points. Since SP 2 and SP 3 were created by adding constraints involving the fractional solution x1 , we say that SP 2 and SP 3 were created by branching on x1 .

Branch-And-Bound Method

195

x2 6 SP 2 feasible region SP 3 feasible region IP feasible point

5

4 z = 20

3

2

1

0

0

1

2

3

4

5

6 x1

We see that SP 3 has an integer solution. Unfortunately, the integer solution of SP 3, z = 39, x1 = 3, x2 = 3 is suboptimal when compared to the non-integer solution, z = 41, x! = 4, x2 = 9/4, of SP 2. So, it is possible that the SP 2 has a further subproblem that has better integer solution than SP 3. So, we have to branch SP 2. Before branching SP 2 let us represent what we have done in a tree. The colors in the next tree refer to the colors in the previous picture.

Branch-And-Bound Method

196

SP 1 z = 165/4 x1 = 15/4 x2 = 9/4 x1 ≥ 4

x1 ≤ 3

SP 2 z = 41 x1 = 4 x2 = 9/5

SP 3 z = 39 x1 = 3 x2 = 3

As pointed out, we must now branch on SP 2. Since x1 = 4 is an integer. we branch on x2 = 9/5 = 1.8. So, we have the new subproblems SP 4 = SP 2 + ”x2 ≥ 2”, SP 5 = SP 2 + ”x2 ≤ 1”. When the solutions to these subproblems are added to the tree above we get the following tree (the color coding is dropped):

SP 1 z = 165/4 x1 = 15/4 x2 = 9/4 x1 ≥ 4

x1 ≤ 3

SP 2 z = 41 x1 = 4 x2 = 9/5 x2 ≥ 2

SP 4 Infeasible

SP 3 z = 39 x1 = 3 x2 = 3 x2 ≤ 1

SP 5 z = 365/9 x1 = 40/9 x2 = 1

Branch-And-Bound Method

197

We see that SP 4 is infeasible. So, the optimal solution is not there. However, SP 5 gives us a non-integer solution that is better than the integer solution of SP 3. So, we have to branch on SP 5. Since x2 = 1 is already an integer, we branch on x1 = 40/9 = 4.444. So, we get two new subproblems SP 6 = SP 5 + ”x1 ≥ 5”, SP 7 = SP 5 + ”x1 ≤ 4”. After this branching we finally arrive at the final solution where all the subproblems are either unfeasible. (There is no color coding in the boxes in the following tree. There is border coding, however. A thick borderline expresses a feasible IP solution, and a dashing red borderline expresses an infeasible case.)

SP 1 z = 165/4 x1 = 15/4 x2 = 9/4 x1 ≥ 4

x1 ≤ 3

SP 2 z = 41 x1 = 4 x2 = 9/5 x2 ≥ 2

SP 3 z = 39 x1 = 3 x2 = 3 x2 ≤ 1

SP 4

SP 5 z = 365/9 x1 = 40/9 x2 = 1

Infeasible x1 ≥ 5

SP 6 z = 40 x1 = 5 x2 = 0

x1 ≤ 4

SP 7 z = 37 x1 = 4 x2 = 1

From the tree above can read the solution to the IP: The SP 6 is the optimal

Branch-And-Bound Method

198

subproblem with integer solution. So, the solution to the IP (11.2.2) is z = 40, x1 = 5, x2 = 0. General Branch-And-Bound Algorithm We have solved a pure IP with branch and bound. To solve MIP with branchand-bound one follows the same steps as in the pure IP case except one only branches on decision variables that are required to be integers. So, solving MIPs is actually somewhat easier than solving pure IPs! We have seen all the parts of the branch-and-bound algorithm. The essence of the algorithm is as follows: 1. Solve the LP relaxation of the problem. If the solution is integer where required, then we are done. Otherwise create two new subproblems by branching on a fractional variable that is required to be integer. 2. A subproblem is not active when any of the following occurs: (a) You used the subproblem to branch on. (b) All variables in the solution that are required to be integers, are integers. (c) The subproblem is infeasible. (d) You can fathom the subproblem by a bounding argument. 3. Choose an active subproblem and branch on a fractional variable that should be integer in the final solution. Repeat until there are no active subproblems. 4. Solution to the (M)IP is the best (M)IP solution of the subproblems you have created. It is found in one of the leafs of the tree representing the subproblems. That’s all there is to branch and bound!

Solving Integer Programs with GNU Linear Programming Kit

11.3

199

Solving Integer Programs with GNU Linear Programming Kit

Solving (pure) IPs and MIPs with GLPK is very easy. The GLPK has all the necessary routines implemented and all you have to do is to declare which variables are required to be integers. To declare a decision variable as integervalued one simply uses the keyword integer in the declaration. Here is the GNU MathProg code for the simple IP (11.1.1): # # The IP (11.1.1) # # Decision variables # Both x1 and x2 are required to be integers var x1 >=0, integer; var x2 >=0, integer; # Objective maximize z: 3*x1 + 2*x2; # Constraints s.t. constraint: x1 + x2 <=6; end;

And here is the glpsol report: Problem: Rows: Columns: Non-zeros: Status: Objective:

ip 2 2 (2 integer, 0 binary) 4 INTEGER OPTIMAL z = 18 (MAXimum)

No. Row name ------ -----------1 z 2 constraint

Activity Lower bound Upper bound ------------- ------------- ------------18 6 6

No. -----1 2

Activity Lower bound Upper bound ------------- ------------- ------------6 0 0 0

Column name -----------x1 * x2 *

Integer feasibility conditions: INT.PE: max.abs.err. = 0.00e+00 on row 0 max.rel.err. = 0.00e+00 on row 0 High quality INT.PB: max.abs.err. = 0.00e+00 on row 0 max.rel.err. = 0.00e+00 on row 0

Solving Integer Programs with GNU Linear Programming Kit

High quality End of output

Here is the GNU MathProg code for the simple MIP (11.1.2): # # The MIP (11.1.2) # # Decision variables # Only x1 is required to be integer var x1 >=0, integer; var x2 >=0; # Objective maximize z: 3*x1 + 2*x2; # Constraints s.t. constraint: x1 + x2 <=6; end;

And here is the glpsol report: Problem: Rows: Columns: Non-zeros: Status: Objective:

mip 2 2 (1 integer, 0 binary) 4 INTEGER OPTIMAL z = 18 (MAXimum)

No. Row name ------ -----------1 z 2 constraint

Activity Lower bound Upper bound ------------- ------------- ------------18 6 6

No. -----1 2

Activity Lower bound Upper bound ------------- ------------- ------------6 0 0 0

Column name -----------x1 * x2

Integer feasibility conditions: INT.PE: max.abs.err. = 0.00e+00 on row 0 max.rel.err. = 0.00e+00 on row 0 High quality INT.PB: max.abs.err. = 0.00e+00 on row 0 max.rel.err. = 0.00e+00 on row 0 High quality End of output

200