Katta G. Murty The University of Michigan SPOHOCHSHOHOEEHOOHHOOHEHHLOLCOELOS John Wiley & Sons New York Chichester Brisbane Toronto Singapore e e ‘e Kd @ ,@ “e 4IMSS SCC CCCC COSTS SCHOO PEED HOOOOOOOOCEEOOEOOOE OOS Linear Programming Katta G. Murty The University of Michigan John Wiley & Sons New York Chichester Brisbane Toronto SingaporeForeword Few realize that linear programming is 8 revo: Iutionary dovelopment that permits us, for the first time in ur long evolstionary history. t0 make decisions about the complex world in hich we live that can approximate, in some Sense, the optimal or best dee'sion. Before the postwar periog (1947), decisions were being ‘made without gathering the relevant facts, or {ganizing them in an applicable way, or consid ring the Infinite number of alternative ways. to achieve desired goals that were tachnolag cally possible and die not exceed the available Fesourcas. Of course, before the advent of computers, such @ model-building appreash wouls nave been unthinkaole. Even with ‘most modrn metods, ii stil unthinkable to sift though all the ways of assigning, say, 70 Inlvidvals to 70 jobs an selecting the dest assignment for each person. However, by com: bining the computer witn methods sven as those discussed in this book, problems many times more complex are routinely solves every day all aver the word Pernape this Is why there was 80 much ex: eitement at a certain gathering in 1049 at the University of Chicago in which young mathoma ticians, economists, and statisticians presented their research on a new too! for attacking com: plex planning problems of the modern world Linear programming was that new tool. Since COWL SCO CCE OOOOH OOES OOOOH OE OOOOEEOOEOEEOOOCEESS then, four of those resent have received the Nobel Prize for research influenced by that cor ference. With tinear programming, one can Abstract the underiying essential similarities in the management of such seemingly slsparate systome a indusiial production, the flows of reeources inthe economy, erap rotation, azsig ment of individuals to jobs, oF the design of complex engineering systems (e.g. transport, water, energy. refinery scheduling). In my opinion, Katta Many nas succeeded in his alma providing students with an up-to-date, in-depth covorage ofall he important practical computational, and mathematical aspects of linear programming. He has done so ina clear, intuitively appealing, understandable way while maintaining an appropriate standard af rigor Only minimal mathematical background Is requires fam honored to have bean asked by Katta Murty to write tis foreword. It was my privilege to have hag him as a graduate student many years ago at Berkeley and to bo able to follow his important contriautions to the field since that time, both as a theoretician and as an expositor. George 8. Dantzig ‘tantra UniversityAIDOS TTT ST CCCCCCOS SOS OG OE OOO OOOO OOO OOO SEE EECCEEOS Preface wrRoDucrion 1 am gratetl for the enthusiastic reception ae appublaned by Weyin 378 Iroceves Sonstrctve suggestions for \meroving it tom Feaders all over fe word, whicn proved 19 be Crtremely vselul a8 | stare’ praparing the materia fortis Book As the Telugu poet J. S Sarma sac Sih T Ssh TS, BS too (TRS BS, OD Bod SOG, fou Osa $0d Da bo7 StonsbS (A good eric ts an authors best trienc). Most ofthe people pointed out that though the book twas quite comprehensive and the coverage fxtensive, there were stil several important topics not discussed. They suggested that in he Second edition an atfor should be made to make the caverage completo. Besides, severs new developments have occurred since 1976 ‘This rapid growin was made possible by the large number of very talented researchers who were strated to work in these areas. There fare reasons for this. Ono important reason is the existence of the very efficient simplex meth- ‘0d developed by George 8, Dantzig in #947 for solving linear programs, as well as the wide Avaliallty of software packages fori, Another is tho practical applicability of these areas. Today, linear programming and the related evelopments in network flow algorithms and integer and combinatorial pragramming make it possible for us to solve many of the complex optimization problems in our modern soviet, industry, and business with the use of comput fers. Another reason i We professional recogni tion given fo these areas by the Nobel Prize CCommitige when i swarces the i975 Nobel Brizo In Eeanomies to TC. Koopmans and L Kantaroviten for devoloping the application of linear programming to the economic problem of allocating resources ‘THE THREE BOOKS, | tried to expand the coverage of the book to include all hese new developments, as wel as ail otner important topies not discussed in the first edition. it 800" became clear thet these ‘dcitions would ereste avery large book. So it Became essential to divide into tree books — fone containing the material on linear progran ning, the second containing the material on hetwork and combinatorial programming, and the third on linear complementarity. These dv Sone also make practeal sense, since most Universities are offering separate courses in teach of hese areas, This is the book on linear programming. | have included inthis book all the material that s normally offered in courses with the words “ner programming” in their ‘He WHAT IS NEW? ‘The chapters on the elipsoid algorithm. the vector minima and ierative methods for solving linear Inequalities, and linear programming problems are new. Even though many of the Other chapter ttlos remain the same as those in the 1376 book, al of tam have buen very thoroughly revised to make the coverage com: plete, and ta include many new ilustrations and ‘THE OBJECTIVES 41 To provide an in-depth and clear coverage fof all the important. practical, technical, Computational, and mathematical aspects 0 linear programming and the transport tion problem 2 To discuss and ihustrate very clearly the rograms and to Rein develop sil in mod ing. 3. To discuss the theory and the geometry of linear programming systematically in an flamentary but rigorous manner so that a eager without much mamematical back ground can easily understand it. 4 To discuss clearly the algorithns for solving linear programs ané transportation rob tems, to present their oticent implementa tions for the computer, and 10 discuss their computational complexity. 5 To help develop akill in using algorithms Intoliigenty to solve practical problems. BACKGROUND NEEDED The background required te study this book is some familiarity with matrix algebea [which is eines in an undergraduate course), especialy ‘he concept of linear independence, anc the matrix reduction methoes for solving systems of linear equations. All this essential back- round material is reviewed in Sections 0.1 to 533 of the book for the sake of completeness SUMMARY OF CHAPTER CONTENTS The book begins with ¢ section entitled Notation, In which all the symbols and several terms are Getined, itis strongly recommended that the reader poruse this ection first and reler toi wherever trare is a question about the meaning of some symbol oF term Chapter 1 presents vasious methods used for formulating problems as linear programs. Many ilystrative examples are used. Section 1.1 is- ‘cusses very elemontary methods, Section 1.2 Seals with the more complicates methods based on piacewise linear approximations. This programming in parameter estimation during ‘modeling. Section | briefly presente tne var (ous approacnes for nancling multiobjective linear programming problems Section 1.8 ineluding, estimation of the data in the mode! ‘Chapter 2 discusses the simplex method for solving linear programs using canonical tab Toaux. This is the originally developed version fof the simplex method. It is computationally ineticient, but idealy suited for solving smal linear programs by hand computation and for introducing the base concepts of the simplex method tothe beginning student We alse cleary Feduced-graaient method when the problem is viewed as one involving the nonbasic variables only, with the basic variables eliminated using possible to branch off ram this and aiscuse the Feduced-gradient ane the generslized reauced Gradient methods of nonlinear erogramming as being generalizations ofthe simplex algorithm, Chapter presents the necessary mathemat cal, geometricsl, and linear algebra concent for stuaying linear programs, the structure of their sole of feasible solutions, and how the simplex method operates on them. This chapter is heavly rooted in linear algebra, and it ee cusses the mathematica! theory of linear pro- gramming, Relatec algortnms in linear algebra [o.g, the algorithm tor testing linear inceper ence, the algorithm for computing @ maximal linearly independent subset of 2 given set of vectors, etc) are al ciscussed here. Results on convex’ poiyhecta that are used in studying linear programming are also presented. Sec tions 31 and 8.2 form a review of the relevant concepts trom mal’x algebra and n-dimen sional geometry. Section 53 provides a review Of the matricreduction methods for solving sy toms of linear equations. Sections 34 to 2.10 Contain the basic mathematieal raeults in linear programming. Sections 311 ané 3.12 discuss the simplex method using matrix operations; the results in these sections are used later in Chapters & and 7 10 develop computationally fd. Section 218 discusses the very importantTIIII TIT TT CCT TTC CSO SC OOCOCOO OOH OOOOS concept of equivalence among linear program: ‘ming problems. Section 2.18 presents alge fithme’ for ranking the extreme points of Convex polyhedron in ineressing order of @ linear objective function Section 3.20 discusses methods for solving Iractional linear program tring problems (Chapter 4 deals with the duality theory of linear programming, The economic arguments Behind duauy are clesriy discussed. The aual variabie are clearly shown to be the marginal Values assceiated with the tems in the original linear programming mod! whon the dual prob: lem nas a unique opium solution, in this Chapter we alsa discuss the theorems o! alter fates or hinear equality or inequality systems, fhe stably of the linear programming mode! necessary and suiciont condiions fora linear program to have a unique optimum solution Sane Zero-sum two-person matrix games ‘Chapter 5 discusses the revised simalex method with Both the explicit form and the product form of the inverses. Here, we also present the infeasibility analysis of linear programming, ane the column generation pro- Eedure in modeling that the revised simplex format makes possible, Chapter 6 considers the dual simplex method using both the ca nonical tableaux ané the inverse tableaux. In the chapter discussion, we provide # geo" metrical lusration of the paths taken by the primal simplex and the dual simplex algo Fithms en the same problem, so thatthe reader fan compare and contras! these two algo- Fithms. Onapter digevsses numerically stable Implementations of the simplex algorithm bases on matrix actoriztions. These imple- mmentations improve the numerical accuracy land nelp o take acvartage of sparsity (Chapters 8 and @ discuss the various types o} sensitivity postoptimality analysis and ranging in linear programming, In Chapter 8 we also fiscuss the computational complexity of the parametric linear programming probiem and the properties of the optimum objective value function in a linear program when itis treates a « unaion of the Hgnthandsice constants SF he orgie ect costietens. When te cual Srobiem has aerate optimum solvens. he Gpnmum obiectwe value trates asa function Shine righsmanaside constants vector Bis nt Sorostaoioat nate In ths cas, an optimum sus Saaten cant be iterretes ase Vector of marginal values eesoiated wth the ema, as pointed out In rooronce [62 mes thttoots en ear progartmng a0 701 discuss tha fetintheirchaptrs on marginal analysis msc pointe = tb). tare are wo marginal mn iand te ane ote rate af hange i the fom por unt decrense in by, known a fhe eative and negative gina valves, In Sec Ton 18 we cleeves now to compute bot the postive and tne negative margin! values ef Erontycusing ne set of optmom solons of he Gua problem naper 10 considers degeneracy and the gration of eying. in degenerate problems, Wren can prevent ne (primal or dual) simon Xgontim tom terminating nate marr ot Hope The goomety of eogeneracy Is clear Scie, and varus mths fr event Syeing under Uagenoracy are oresente apter1 eats win ecient versions fo imal ané gua! simplex eto for handing Eoundee varebe inear programs. The genet= ization st fis fo the "generalized upper Sounding” (or GUS) constrains is presented tory sles) This naturally ead to “uctred Tow prouramming." an ares that deals wih thtcen’mpiomentatone of the simplex meth Sor solving large sate nes-programming trodels, which enpot te special suet of the mado Chapter 12 presents he cecompos- ton prnige nea grogremmingChaptr "S'veats wht special methods available fr handing anoportation proses. Al the prom trios Sr imo problem based on is speci Strucure are clearly derives. Efficient imple onmons ol the primal. and. ds simplex mmatous, sing te tee coresponcing fo the basis, ane the tripe labe! representation tor storing and manipulating the trees are. pre sented very clearly. Using this implementation very large Wvansparation problems are being applications (Chapter 1 ciscusses the worst case the ave ‘age computational complexity af the simplex method. Chapter 15 presents the poly omially Bounded ellipsoid algorithm for solv ing. system of linear inequalities, or near programs Most books on linear programming usually ignore iterative methods, One reason for this might be the tremendous practical success of the simplex matnod. Isading to the view that it Is futile to present other approaches. However, for tackling farge-seale unstructured lnoar pro: gramming models, teratve methods may pro vide 8 potentaly usetul alternative. Therefore, Wwe present a brie! survey of various iterative ‘methods in Chapter 16. Chapter 17 deals with methods for computing the vector minima in & mmultiobjactve linear programming problem, HOW To USE THE EXERCISES The exercises in the book are of several types. 1. Formulation problems. An effort is made to Inelude in Chapter + formulation problems from many of the areas of aplication that use linear programming, The insvuctor can formulate some of these problems in the class to ilustrate the wide applicability of linear programming and assign some others fs homework exercises) 2 Numerical problems. These require the ep plication of th algorithms ciscussed in the book. Solving these problems will give insight into the ways In which algorithms 3. Prools and other problems. These require @ deeper understanding of the materiat dis 4 True oF false questions. Several of these ‘questions in Chapters 8 and contain Statement and ask the reader to mark whetner it's true or false and to justify popular in the 1976 book, and the beginning Students seemed to learn a lot by doing them. The instructer could give a quie to the lass based on these questions sovera times during the term, ane thay can be uses to promote class discussion as well Exefcises that are listed in the misle of @ chapter ean normally be solved using the methods discussed inthe section in which they fanpear. A new sequence of exercise numbers Deging with eaen chapter (e.., Exercise 62 rolers to exercise number 2 in Chapter 6) WHAT DISTINGUISHES THIS BOOK FROM ‘THE OTHERS IN THE AREA ‘This book's main cistinguishing feature is its completeness and compronensivenass. Most Important topies in linear programming re discussed clearly ard in depth. The practical ‘modeling, mathematical, geometrical, algorit- ‘mic, aad computational aspects of linear pro gramming are all covered vary carey in Complete detail. Many of the books inthe area usually epecialize monly afew ol hese aspects For example, this is one of the few books that iscusses the implementation o! he primal si piax algoritnm for solving transporation prob: lems using tree fabels. Algorithms for solving diferent types of transporation models (those involving equality constraints only, or those Involving Inequality constraints, or the interval type of constraints, bounds oF the variables, ‘discussed separately and the material for reading betore the class and than spend the class ime going over the main Points, leaving the detalis to be read from the and the large number ofTHVT CCCCS exercises of various types make this book very Convenient for eter sel-study or for an it Strustor to teach from with very ite additional ‘tft on nie 0” ner part. WOW To USE THE BOOK IN A COURSE The material in Chapters + and 2; Sections 3.1 to 314 of Chapter 3; Sections 41 to 4:8 19 Chapter 4 Chapters 5,6, 7, 8,9, 10,and 18 torm the base core ofa one-term first-year graduate level course. in linear programming, and a Course can be covered inthis order. To teach mené emphasis on the mathematical theory Giscussed in Chapters 3 and 4. The same tourse can be taught to emphasize applications Of the material, especially the formulation, sigorthmie, and computational aspects of ln fer programming, The remaining material in the book cen be the basis for a second course on linear programming. Or anyone who hes taken the fst course gan study this material though sell-stdy fairy easily. By maintaining 2 fast pace, it's algo possible for @ welt Motivated class to cover the entire Book In a Slagle term. Tan undergraduate course on “inlraduetion 0 optimization” ean be taught using selected material from this book ané the companion book Network and Combinatorial Programming (relerence [194], abbreviated here as NCP) in the following order: Soctions 1.1 of this book tovering elementary methods of formutatian of linear programming problems; Sections 2.1 10 4.3 ofthis book to raview the background mate Fal of matrix algebra; Sections 4.1 10 44 and Sections 4.5.5 and 45.6 of this book, covering Guslity, marginal analysis (in particular, how to write down the cual of @ linear program in Standard form), and tne relationships of the primal and dual goltions associated with an bolimum basis forthe problem; Section 3.1 of NCP, discussing the Hungarian method 10° Solving the assignment prodiem (this provides 2 special aigoritim for a highly structures Tinear programming probiem, which is very fey to understand); Section 73.2 ofthis book iscussing the simpltied version of the primal sigoritim for solving the balanced transporta- tion problems; Chapter 2 ofthis book (when the flgoritime are discussed in this order, the Course moves tram simple, elegant algorithms for specially struclured problems to somewha complicated algorithms for goneral problems that lack stucture|; Sections 8.1 to 9B and 2s, discussing some sensitivity analysis, ane including generat linear programming an ansperiation problems; Chapter 12 of NCP, Covering some simple formulations of integer programming problems; Sections 151 10 15.7 of NCP, covering the branch-anc-bound approach and some supplemental material on applica: tions of dynamic programming ané nonlinear pragramming Of course, a wellmotivated person can also use this book for sel-study to learn linear pro- ‘gramming, Operstions researchers anc people ‘who use optimization in thir professional work Can use this book as a relerence for details of the methods or for developing improved slgoathms REFERENCES Rotorences are listed at the end of wach chap- ter, anda list of Dooks on linear programming is provided at the ond of Chapter 1. Because of the lack of space, the list of reteronees had to be kept very briel, However, most reterences sed in the preparation of this material have been cites Inpreparing thisbook have received advice and encouragement from several people. For this I'am heavily indebted to Mustala. Akai John Barthold, Pat Carstensen, R. Chandra. Ssesnaran, Sung Jin Chung, Yahya Fath, Akl Gana, David M Gay, Jack L. Golaberg, ihuyo Kaneko, Oli L- Mangasarian, Diana. P. O'leary. M./H. Pavtovi, Clovis Perin, Mike Plantholt, Loren Plateman, Steve Pollack, Ro: mesh Saigal, Art J. Schwarz, Bod Smith, W Allen Spivey, Kisus Teuemper, Layne Watson, fang Chrie Witegal, ‘The tinal version of the manuscript was completed during 1961-1982. !was an sabbati- cal leave trom the University of Michigan then and spent part o! that year at the Schoo! of Management ang Administration, University of Texas at Dallas; the ramaining part ofthe year | was at the ingian Statistica institute in Dei, India, on a project sponsored by the grant, INT-8109825 trom the U.S-India Cooperative Scienco Program, NSF Special Foreign Cur rency Program. I thank the University of Michigan for this eabbatical, the University of Texas at Dallas (in pericular, R. Chandra: sekharan of the School of Management and ‘Administraton) for inuiting me to spend a term ‘wits them, and the NSF (in particular, Osman Shinaishin, Program Director of US-India Co- operative ‘Science Program) for making it possible for me to vist the Indian Statistical From 1878 to 1982, the research carried out bymy graduate students and mysel in optimiza tion was partially supported by the Directorate fof Mathematical and Information Sciences of the US. Air Force Ottice of Scientiie Research Unger Grant AFOSR 76-9646. Several resuls, trom this research appear throughout the Book I'thank the AFOSR (in particular, our program manager, Joseph Bram) for this support and opportunity The manuscript was reviewed by Donovan Young ef Georgia Insitute of Technology. | thank him ‘or his excellent comments, They were very set in th final revision | am deeply indebted to George B. Dantaig for making lt passibie for me to study opera- lUons research, anc for graciously agreeing to write the foreword tor this 800K Most ofthe typing of the manuscript was done by Geraiaine Cox anc Jane Ouisiay at the Department af industrial and Operations Engi- nesring, The Unversity of Michigan, I thank them for theie patience an untiring ettort | would also lke o thank Bilt Stenguist, Eng neering Eaitor at Wiy, for his encouragement, Finally, | thank my wile Vijaya, my daughters Vani and Madhusri, ana Mother Adiakshmi for their patience and support. This bodk is ded cates to tem atta G. MurtyCOCK SOHSEHOSSSOSH OOOO OO OOOOH HOHOEHHLOOOE CEE Contents NOTATION, XV 1 Formulation of Linear Programs, 1 Sohing Liner Programs ia Modeling Mute Onjectve Problems, 27 Cenattucting Lina Progremming Mi Proctcal Appicabons, 22 nats Meant Soing + 2 The Simplex Method, 52 Irtvodueion, 62 Tanstorming the Probie eta the Standard Form, 53 Canonical Tebieno, $6 ‘ase witha Full rill Sass, 0 Outine ofthe Simplex Method for 2 Genera The Big Method, 8 3. Geometry ofthe Simplex Method, 91 clidoan Voetor Spaces: Dain somata! Concepts 9 ‘Basic Feasible Solutions and Extrome Points Conver Polynera, 16 ‘Ina Uelungas of Bane Feasiie Stutons In near Programming, 120 Simalen Algorithm. 12¢ ‘Ihe Wain Geometries! Argument nthe Simplex Algor, 137 an ‘nernate Optimum Feasibe Solutions ane why # an Aigo Required t Solve ase (of ne Simplex Method Using One Why Not voduee Two or More Veribles 8 Fractional Programmi, 18 Precisian in Computation, 164 4 Dually in Linear Programming, 182 nami of Dual Problems, 182 ‘bua Variables Aro the Marginal Values Prooiam Has # Unique Optimum Solution, 15 How to Wate the Dua! o! 8 General Linear Program, 186 COinerinerretatone and Apaietions "Necessary and Suticent Conlons for & Optimum Solution, 208 Stbiity of he Linear 5 Revised (Primal) Simplex Method, 231 51 52 Aeviges Simplex Algorithm with the Fevised Simplex Method Using Phase 55 ‘The Revisod Simplex Format and tn Column 8 The Dual Simplex Method, 249 se sa Simplex 23 Dual Simplex algorithm Wnen & Oval Simplex Algorithm 254. ” Simplex Algortnm, 285 Basis i not Known a the Slat. 238 Comparison of re Primal and the Dust ‘Aternate Optimum Solutions othe Dual Provlom 260 Stable Forms ofthe hod, 268, Roviow, 258 Cholesky Factoneation, 272 Use in Computer Coaes, 276 @ Parametric Linear Programs, 278 ts Inwocuetion, 278 The Paramotre Cost Simplex Aigorithn, 279. ‘oFFing an Optimum Basia or tne 8 the Optimum Obiocive ec Figh-tand. Sie Right tane-Sige Problem, 296 Right Hane-Side Preble 286. {810 ‘Te Siopo 0! me Optimum Objective Valve 18. The Seitaual Paametic Matha (Also 215 Marginal 8 Sonstvty Analysis, 210 Intosustion, 810 Inroducing a Now Activity, 910 83. nrogusing an Additonal qualiy Cantain, 32, 95 Cost Ranging of Nonbasic Cost Coeteien, 318 88 Cost Ranging ofa Basi Cost Coticent, 216 87 RightHane Side Ranging 317 gee tn the Input-Out 89 Cnanges in 2 Basic input Ourput Coatcion, 218 Sensitivity Analysis, 98 10 Degeneracy in Linear Programming, 323 430.1 Geometry of Depeneracy, 928 103. Summery of he Primal Simplex Algor 104 Do Computer Codes Use the Lenco Bin Rate ruie? S30105 Resolving Bayeneracy Using Only te 406 esottion of Cyelng In the Out Simpies Algor 1 11 Boundes.Variable Linear Programs, 42 113, Primal Simplex Method Using forking Bases, 360 Fesolution of Cyeing Under Degeneracy Inthe Bounsea-vaniasle Primal Simplex Aigoritin. 359 117 Gonorazad Upper Bounding (GUE), 98 12 The Decomposition Principle of Linear Programming, 371 121 robems witha Block Angular Structure, 271 122 interpretation in Soop of 12 The Transportation Problem, 380 198.1 The Balanced Transportation Prabiem: Theory. 380 122 Aovied Primal Simpiox Algorithm Balances Transportation Problem, 2 188 Unimodular and Tot ee Sranaporiaion Problem, 410 188 The Out Simple» Metnod fr the Balanced Transportation Problem, 16 inequallty Construits, 417 ‘96 Bounded Variable 44 Computationat Comploity of the Simplex Algorithm, 433, 1141 How Eticiant ia he Simplex algortnn? 433 342. Worst Case Computational Complexiy. 494 143 Average Computationa Complexity. 440 45. The Ellipsolé Method, 44 System of Open Linea lneguailes, 445 136. Elipslasyge Methods Using Otor 16 Horative Methods for Linear Inequallios ‘and Linear Programs, 456 164 lateduction 488 162 Aelonsion Methods, «56 164. Sparaty-Prasarving erative SOR ‘Agoritsfor Unear Programming, 483 185 An lterave Method for Saving a Primal 47 Vector Minima, 465 74. Oetinions ana Propetas ot 172 An Agari to Generate all he Exvema Point vw, 468 179 An Algvien lor Gonerating al the Eicient Faces of K. 886 Notations 2 Seto lines (ares or edges) na nawork. I This ents nore pa te (6:1. Tm 16-40 enoraored pai (the samicoo se Unless ofnewwise dalS69, mse are usualy EA.B.X.P.F,0.U,L Unless cterwise species NEPLO.FT.D1.U,H.4 Tresesymbole use fecioral const in mieanes, Sve Sec olynedron K. See Section 3. * ending 0 & convex polyhedron K. SeeCDI IT TIT TECTTCCCCC HOO OC OO OOOOC OOOOH OOOCOCOHOCEEEE un. Si, 8, YB) seas wed in Be Hane iy, Tre sel ot eages inte gran tn) etna ony for ee rubra a rere Forevarple(—aSt= 4 [83)= 5 Islets ony er ee muerso veresn Cal sie © SeUgt Fy an sloment fF" Alo “FF symbol EP are two sets, "Be Pr jan hat “Esa be tT 0 tha every ot union symbol, H are wo S015. DU His The emp sot The 4at containing ne elamens uma (Te minimu ruber among the et ts empiy mew con ne conveton batho > Lesicagraohicaly great HAT See Section 123. cop A minimal nesty dependent set feels ‘stensporaton array See section 3.16 a) Tsraterstomne Figure The jh igure in Chapter +The Relecence i /]_ The th relavanee inthe exereige Lf The th exe in Chapter ¢ exer keenin= 9) Soperscrits We use superserints to enumerate neva. asubsetot asst Bis salsto act of Ce that ir parameters) Isa tunction oftheTHTTTSTCTCEC SCC CCOOOSH OH OE OHOOOOHO EEO SOOOOCOOEHOEEEO agate variables (soe Seton 22), Ehminate PosiAi soos) A, CardinaityDetined only forsee. The carina of ‘Tian or sack constraints Ll be & point in AE Satisying an ineaualy constrain Hx) 20H 10) > 0, a consents) 2 0” v8 sd to be "ers, wnenever# futon Js) Nas to be mas 1 (f(a) Hence tne aigoritims iscuseed anit See Section 18 mete about ns 825) had asigniieetinivence Bigortnm casero rulgotor geting wreaures eres sain fr the propery fram. minimize 4x), sobeet Yo) 20 fo Ben 1 90) ‘a¥R(al=0.9{0)2 Otor!= tom, heh = tw) Hntely ng signet for sol The clase P of probleme This isthe class of al obleme for aolvng which there aus & pySSCCHSSSOSOHOEHSOHSOCHSSHSOKTHOHOCHOHSHEESCCHHOOHHHOKCHOOEHOE OEE Contents orATION, xv 2. The Smpex Metod, 52 Zh Onunwane spor ae fr» Gores! 2 Geometry of the Simplex Methos, 81 (Gennaro Cosap 81 Sareea ena By ne Eels ee Simplex Agr oP saa Tmoomensn a a Cones Payhewon 17 {130 Aturate Opium Foasiie Slaton ana 332 iy an Algthm Request owe ir ogra 8 ‘Simutarwovsy we oe eee? 18 4316 Aictage Method fr Sovog syste Sytem at nes noua, 30 ais Ehuvnan tia ropes 153 2 Etrera Pore | 46s Doty tory ina Progaming 18 ony. eae Une Priam Have a Uniaue ‘atu deen. 08 44¢ Sty fhe Lea 5 Revised (imal Simplex Method, 231 The Dua! Simpiex Method, 249 42 Oi Sipe Agrtin wn Da 69 seams sa dang fe Oa Sinper Agog, 55 ct Sip es an 2 ul esi Att Spa Shans ee 1 Numerialy Stable Forms of he Simplex Method, 28 4 ie Linear Programs, 276 83. Teta atmen dans tore ‘Sinpoe Ago 205 {690 Th Sl le pamm Ove Vue ta Tha butcut pramane Mea Aso Gana rat sa rn 1 Proertas oe Opin Obie Value ys trials Sano Pe 9 Senshviy Anaiyes, 210 38 args eine Ou oie in 0c en 10. degeneracy in Linear Programming, 323 101 oman o Oogeneray. 29 102 Reston Ging trFormulation of Linear Programs 1.1 FORMULATION OF LINEAR PROGRAMS 1.4.1 Introduction Optimization, or mathematical programming is the branch of mathematics dealing with methods for opti mizing (ie., either maximizing or minimizing) an objective function of n decision variables subject to specified constraints on these variables. We begin this book with an incident related to optimization. ‘Once | was planning for a trip to India. At that time my brotiter, who lives in India, had a four-year-old daughter and a boy who was just born. The question of what git to take for my brother's family arose, and my wife suggested that | take a baby seale, which they could use to keep track of the baby's weight. We got a really fancy one, and my brother became 50 excited on seeing it that he wanted ta check the baby’s weight immediately. We set it up and he put the baby on the pan, His four-year-old daughter was watching us with great curiosity, but she could not figure out what was going on, so she asked her father what he was doing. Just to tease her, he said, "You see, honey, your uncle from America likes your brother so much that he agreed to buy him from us at a cost of $1 per ounce of his weight, So | am weighing the baby to see how much money we will get by selling him to your uncle!” On hearing this, the litle girl looked quite shocked, and with a sad face, she asked, "You are not going to sell him now, are you, daddy?” From the worry in her face it was clear that she loved her brother very much; this pleased her father, and in order to console her, he said, “Don't worry, honey, | will never give your brother away to your uncle. | said it only to tease you.” Rolling her eyes in surprise, the little girl then said, "That is not what | meant, daddy. You see the boy is very small now, but he is gaining weight rapidly every week. If we wail for some more weeks, he will be much heavier, and will fetch much more money ,...” The attitude of the little gir in this story chapter is one of optimization, and it is almost universal Interest in optimization methods and their appli cations is very widespread now. In the days before digital computers, only very simple optimization models involving few variables and few constraints could be solved by the tools then available. So, in those days, a great deal of emphasis was placed on keeping the model very small and compact, as otherwise it could not be solved. In the last 50 years, a large number af algorithms to solve 4 great variety of optimization problems have been developed, and large digital computers for imple- menting and executing these algorithms have be- come widely available. Consequently, many types of optimization models can now be solved efficiently: therefore, practitioners are building many large- scale optimization modets and solving them almost routinely. The emphasis in constructing an optimi zation model these days is to keep it computable or tractable. One such model is the linear program- ming model, in which a linear objective function is to be optimized subject to linear equality and inequality constraints and sign restrictions (or lower and/or upper bounds) on the decision variables. We now have very efficient algorithms for solving linear programming problems, and very high-quality soft- ware for these algorithms is available at minimal cost, Many large-scale linear programming prob- ems (involving a few thousand constraints in several thousands of variables) are being solved daily very satisfactorily using this software. In this book we discuss linear programming in all its depth. In this chapter we discuss techniques for constructing linear programming models for prac- tical problems. To formulate a real life problem as allinear program is an artin itself, Even though there are excellent methods for solving a problem once it is formulated as a linear program, there is little theory to help in formulating the problems in this way. In a real life problem, several approximations may have to be made before it can be modeled as2 1 Formulation of Linear Programs a linear program. The basic principles will be illus- trated by considering a simple example. We will use the abbreviation LP for linear program 1.1.2. Product Mix Problem A-company makes two kinds of fertilizers, called Hi phosphate and Lo-phosphate. Three basic raw mate- flals are used in manufacturing these fertilizers as in the table given at the bottom ‘The raw material supply comes from the com: pany's own quarry. The net profit is the selling price minus the labor and other production costs. How much of each fertilizer should the company manu: facture to maximize its total net profit? This problem will be formulated using two different approaches, the diract approach and the input-output approach. 1.1.3. The Direct Approach Step 1 Prepare a list ofall the decision variables in the problem, This list must be complete in the sense that if an optimum solution providing the values of ‘each of the variables is obtained, the decision maker should be able to translate it into an optimum policy that can be implemented. in the problem of Section 1.1.2 the variables are x, = the tons of Hi-phasphate to be manufactured x; = the tons of Lo-phosphate to be manufactured Associated with each variable in the problem is an activity that the decision maker can perform. In this, problem there are two such activities: Activity 1: to manufacture 1 ton of Hi-phosphate Activity 2: to manufacture 1 ton of Lo-phosphate The variables in the problem just define the levels lat which these activities are carried out. When the Tons of Raw Material Required to Manufacture 1 Ton Raw Material __‘Hi-Phosphate 7 2 2 1 3 1 Net proit per ton manutactured 815 constraints binding the variables are written, it will be clear that there are additional variables (called slack variables in what follows) imposed on the prob: lem by the inequality constraints. Slack variables will be discussed in greater detail later. Formulation of the problem as an LP requires the following assumptions: Proportionality Assumption It requires 2 tons of raw material 1 to manutacture 1 ton of Hi-phosphate. The proportionality assump- tion implies that 2x, tons of raw material 1 are required to manufacture x, tons of Hi-phosphate for any x, 2 0. In general, the proportionality assump- tion guarantees that if a,, units of the ith item are ‘consumed (or produced) in carrying out activity / at unit level, then a,x, units of this item are consumed (or produced) in carrying out activity j at level x,; and if $c, of profit are earned in carrying out activity j at unit level, $¢,x, is the contribution of activity j to the net profit: for any x, 2 0 Additivity Assumption It requires 2 tons of raw material 1 to manufacture 1 ton of Hicphosphate and 1 ton of the same raw material to manufacture 1 ton of Lo-phosphate. The additivity assumption implies that 2x, + x, tons of raw material 1 is required to manufacture x, tons of Hi-phosphate and x; tons of Lo-phosphate for any x, 20, x, 20. The additivity assumption gen- ‘erally implies that the total consumption (or produc- tion) of an item is equal to the sum of the various auantities of the item consumed (or produced) in carrying out each individual activity at its specified level. The additivity assumption implies that the ob- jective function is separable in the variables: that is, if the variables in tne model are x)...» %, Maximum Amount of Raw Materials Available Lo-Phosphate per Month (ons) 1 1500 1 1200 ° 500 s10 Se COCOOOOOOOOO OOOO EOOOOOO OO OOOOEOOCOOOOOOO OOOH OOCOOEOSPHCHOHSHSSHSHSHHSHOHSHSSHHHSSHOHOSSHHOHSHEHOHOHOOCHOHHOHECHOOOSCEE and the objective function is 2(x),.... %4) = 21x) ten z(x) can be written as a sum of n functions, ‘each of which involves only one variable in the model fie, as 2(%,) ++: +2,(%,), where z{x) is the contribution of the variable x, to the objective function] The proportionality and the additivity assumptions together are known as the linearity assumptions. The Implications of the Proportionality and Addi- tivity Assumptions In most real lite problems, the proportionality and additivity assumptions may not hold exactly. For example, the selling price per unit normally decreases in a nonlinear fashion as the number of units purchased increases. Thus, in for- ‘mulating real life problems as LPs, several approxi- mations may have to be made. it has been found that most real life problems can be approached very closely by suitable linear approximations. This, and the relative ease with which LPs can be solved, have made it possible for linear programming to find a vast number of applications. The proportionality and the additivity assumptions automatically imply that all the constraints in the problem are either linear equations or inequalities. They also imply that the objective function is linear. Continuity of Variation Assumption It is assumed that each variable in the model can take all the real values in its range of variation. In real life problems, some variables may be re- stricted to take only integer values (eg., it the variable represents the number of emply buses transported from one location to another). Such re- Strictions make the problem an integer programming problem. integer programs are much harder to solve than continuous variable LPs. They are nat dis- ‘cussed in this book. See (1.71, 1.92, 1.94, 1.104, 1.103] for integer programming, Step 2 Write down all the constraints and the objec- tive function in the problem. Nonnegativity Restrictions The variables x,, x, have to be nonnegative to make any practical sense. In linear programming models in general the nonnegativity restriction on the variables is a natural restriction that occurs be- 1.1 Formulation of Linear Programs 3 ‘cause certain activities (manutacturing a product, ete.) can only be carried out at nonnegative levels lems and the Associated Constraints There may be other constraints on the variables, imposed by lower or upper bounds on certain prod- ucts that are either inputs to the production pracess or outputs from it. Any such product that leads to a constraint in the model is called an item. Each item leads to a constraint on the decision variables, and conversely every constraint in the model is asso- ciated with an item, Make a list of all the jtems that lead to constraints in the prablem. In the fertilizer problem of Section 1.1.2 each raw material leads to a constraint, for example, the amount of raw material t used is 2x, + x, tons, and this cannot exceed 1500 tons. This imposes the con- straint 2x, + x; $1500. Since this inequality just compares the amount of raw material 1 used 10 the amount available, itis called a material balance inequality. The material balance equations, or in- equalities corresponding to the various items in the problem are the constraints of the problem, Each material balance inequality in the model con- tains in itself the definition of another nonnegative variable known as a slack variable, For example, the constraint imposed by the restriction on the amount of raw material 1 available, namely 2x, + x35 1500, can be written as 1500 ~ 2x, ~ x, 20. It we define x, = 1500 — 2x, — xz, then x, is a slack vari able representing the amount of raw material 1 re- ‘maining unutilized, and the constraint can be written in the equivalent form 2x, + x; + x, = 1500, where x20. Objective Function Now write down the objective function, By the pro- portionality and additivity assumptions, this is guar- anteed to be a linear function of the variables. This completes the formulation of the problem as aa LP. Itis as follows: Maximize (x) = 15x, + 10x, tem Subjectto 2x + xy S 1500 Rawt xt 8 1200 Raw 2 x 500 Raw3 x20 20 ay4 1 Formulation of Linear Programs Example 1.1 A Blending Problem A refinery takes four raw gasolines, blends them, land produces three types of fuel. Here are the data. Raw Gas Octane Available Price Type Rating arrels per Day _per Barrel ro "4000 $31.02 86 5080 33.5 st 7100 36.35 99 4800 3875, ‘Minimum Selling Octane Price Rating (Sibarrel) Demand Pattern 1 95 48.18 At most 10,000 barrels/day 2 %0 42.95 Any amount ean be sola 3 85 4099 Atleast 15,000 barrelsiéay ‘The company sells raw gasolines not used in making fuels at $38.95/barrel if its octane rating is over 90 land at $36.85/barrel it its octane rating is less than 90. How can the company maximize its total daily Profit? In formulating this problem, we find that the variables are xy = barrels of raw gasoline of type i used in making fue! type j per day for i = 1 to 4 and j= 1.2.3 yj = barrels of raw gasoline type / sold as it is in the market The octane rating of fuel type 1 should be 295. Assuming that the proportionality and additivity assumptions hold for the octane rating of the blend, the octane rating of this fuel, if a positive quantity of this fuel is made, is (68x, + 86x, 191%), + 99x45)! (140 #25 + tq). Thus, if a positive quantity of this fuel is made, the constraint on its octane rating requires (68x), + 86K,, + 91x, + 99x. MO +1 + X41 + Xay) 295, oF, equivalently, 68x, + 86x, + BH, 4 9k, — 950K, Ha, FA, + Ho) BO. This Inequality will also hold when the amount of this fue! manufactured is zero, Hence, the octane rating con- straint on this fuel may be represented by this linear inequality. If we proceed In a similar manner, we ‘obtain the formulation given at the bottom, By defining the variables as the amounts of various ingredients in the blend, blending problems can be formulated in this manner. However, ifthe variables fare defined to be the proportions of the various ingredients in the blend, remember to include in the ‘model the constraint that the sum of the proportions of the various ingredients in a blend should be equal to one, Maximize 45.15(%,) 4 X24 + ty} Xa) 42950012 + a2 +2 + Head 4 40.99(445 + a3 + Xa9 + Kaa) + ¥1(85.85 ~ 31.02) + y3(36.85 ~ 99.15) + y3(88.95 — 36.35) + y4(98.95 — 98.75) — 91.02(x, +42 + Xi9) = 8.150024 + Xan + Hay) ~ 96.95(%y5 4 Xy9 + Hy9) — 9B.750Kay + Ker + Mead Subject to 681, + 86x25 + 91x, + 99Xqy — 95(K Xa HX + Xa) ZO 68k: + 86Xp + 91R5, + 99K) — 90UH, Xe +2 + Xa) ZO Bx, + 86X;5 + 91X55 + 99x45 — BSL + X2y +445 + Kes) ZO Xt tia ty, = 4000 Yay ta tas ty = 5050 Nyy Fa by bY = 7100 Xa thar that Ve Ai tay ta ty 4300 $ 10,000 Aig tay # Hy3 + Xap 2 15,000 ° for all (andj SOHOHOHOHOHOHOHOHSHSHOHOHOHHSHHHOHSHHHHOHHHSHHNHHOHHHHHOHHHEHCEOCHELESHOHHOHSHSHSHSHSHHSHSHHHHOHSHSHHSSHSSHHHOHSHOHSHSOHHOHHOHHOHECHHOEOSOCE 1.1.4. The Input-Output Approach In this procedure the lists of all the possible acti- vities and the items defining the constraints are first prepared. The list of activities should include all the possible actions that the decision maker can perform in the problem. Units should be fixed for measuring the level at which each activity is carried out. The levels at which the activities are performed define the variables in the problem. ‘An item in the problom is any material or resource ‘on which there is either a requirement or a limit on its availability or use. Each item leads to a constraint in the problem. The list of items should include an item called dollar that defines the objective function. It money is required to carry out an activity at unit evel, dollar is an input for carrying out this activity. Ht money is produced as a result of carrying out an activity, dollar is an output from it. Even it the optimized item is different from money, we will use the term dollar to describe it. To write the constraints in tabular form, define x, as the level at which the jth activity is carried out, and assign the /th column in the tableau to corre- ‘spond to this activity. Likewise associate a row of the tableau with each item. If a,, units of item i are required as an input for carrying out the jth activity ‘at unit level, enter +a,; in the (i, ) position (Le., in the ith ow and jth column) of the tableau. On the ‘other hand, if a,, units of item j are produced as an ‘output by carrying out the jth activity at unit level enter —ay, in the (i,j) position of the tableau. Thus, inputs to activities are entered with a+" sign, and ‘outputs from activities are entered with a" —" sign (405). In the objective row corresponding to the item “dollar” all costs are inputs and, hence, are entered 14 Formulation oftinear Programs 5 with a4" sign, and all protits are outputs and are considered as negative costs and are entered with a ” sign. The constraints are completed by entering the limits on the availability or requirements of the items, on a right-hand side column of the tableau. The objective row gives the coetficients of the objective function that is to be minimized. When formulated in this manner, the fertilizer product mix problem leads to the tableau given at the bottom ‘The blank entries in the tableau are zeros. This is equivalent to the model formulated earlier, Xj, XX are the slack variables corresponding to the inequality constraints obtained in the earlier model Example 1.2 A Multiperiod Problem A nationalized company can either manulacture capital goods or consumer goods. The company’s production is being planned over a five-period horizon, At the beginning of period 1 the company has 100 units of capital goods and 50 units of consumer goods in stack. Each unit of capital goods can be used to produce either three new units of capital goods at the expense of three units of ‘consumer goods to be provided from stock or 10 units of consumer goods, over an uninterrupted two-period interval, at the end of which the unit of capital goods again becomes available for use. Once capital goods are produced they can be used for production in subsequent periods. It necessary, capital goods can also be lett idle. The company will sell everything in its stock at the beginning of period 6 and close up. The company can also sell either capital goods or consumer goods in any period according to the following price schedule. Which production plan will best maximize the total returns? ‘Activites and Levels ToLeave T Ton Unwulized To Make 1 Ton of Raw Material So Aight. irphosphate _Lo-phosphate Huphosphate __Lo-phosphi 728 OM toms x“ % | Constants Raw material? 2 1 1 1500 law material 2 1 1 1 1200 Flaw material 3 1 1 500 Dollars cost 15 -10 =| 2x) minimize6 + Formulation of Linear Programs 1 vnit apis petwity in pened | > inperiod t= 2 ——! 3 unite consumer ‘onde im period Figure 11 Price in Monetary Units per Unit Goods During the Period Period Capital Goods Consumer Goods 7 « 04 2 6 os 3 a os 4 6 015 5 a 01 6 3 on We will formulate this problem by using the input-output approach. During each period the com- pany can possibly engage in the following activities. 4. Assign one unit of capital goods to manufacture three units of capital goods over the next wo periods. 2 Assign one unit of capital goods to manufacture 10 units of consumer goods over the next two periods 3 Leave one unit of capital goods idle until next period 4. Sell one unit of capital goods. ‘Sell one unit of consumer goods from stock. 8 Carry over one unit of consumer goods in stock until next period Of course the company cannot engage in activities 1 ‘or 2 during period § because the company has to close down at the beginning of period 6. Let x,, be the level at which activity jis carried out during the ith period The items are the capital goods and the consumer {g00ds in stock in the various periods. Here is the list of items: Capital goods available at the beginning of period 1, Consumer goods available at the beginning of period i, for various values of i, and dollar. Clearly the iter capital goods at the beginning of period 3 is an input into activities 1, 2, 3. 4 during period 3 and it is an output from activity 1 during goods in period F +2 Le in period in peiod 2 poriod 1, Also one unit of capital goods assigned to produce new capital goods at the beginning of period ‘finishes this job by the end of period 2 and becomes, available for use again at the beginning of period 3. ‘Thus the output from the activity of assigning one unit of capital goods at the beginning of period 1 to produce capital goods is four units of capital goods (itself and the three new units of capital goods it has produced) at the beginning of period 3. Figure 1.1 represents some of the input-output features in this problem. From these considerations we arrive at the formulation of the problem in Tableau 1.1. Each of the constraints in the problem is a material balance equation; this formulation could also have been obtained by direct arguments. Example 1.3 A Diet Problem In the diet model a list of possible foods that can be included in the diet is provided, together with the ‘nutrient content and the cost per unit weight of each food. A diet is a vector that specifies how much of ‘each food is to be consumed per day. It is a feasible diet if the amount of each nutrient contained in it is greater than or equal to the specified minimum daily requirement (MOR) for that nutrient. The probiem is to find a minimum cost feasible diet. Inthe following diet problem the nutrients are starch, protein, and vitamins, and the foods are two types of grain with data given below. Nutrient UnisiKilogram ot Grain Type (MOR of Nutrient Nutrient 7 in Units Starch Protein vitamins Gost (Sa) ‘ot food -é& é rePey pepe] [ae fo-| sree fre-f re ae 2 5 a 3 | = fm aie eoauee H ro w= hye arr Tae Vb neaiaen ‘e ‘oe e e e ‘e \° (e e e e ‘e ie e e ‘e ve (e@8B 1 Formulation of Linear Programs The activities and their levels in this model are for / = 1, 2: Activity j: to include 1 kg of grain type j in the diet. Associated level = x,. So x, is the ‘amount in kilograms of grain j included in the daily diet, j= 1, 2, and the vector x = (x, %)" is the diet The items in this model are the various nutrients, each of which leads to a constraint in the model. For example, the amount of starch contained in diet x is 5x, +7xz, which must be 28 for feasi bility. This leads to the following formulation, Minimize z(x)=0.60x, 40.95%, item Subject to 5x, #74228 Starch 4x, 42x,215 Protein 2x, 4,29 Vitamins and x20 Comment 1.1: A diet problem involving 77 difer- tent foods, to meet the minimum requirements of nine different nutrients at minimum cost was formulated by G. J. Stigler ia "The Cost of Subsistence,” Journal ‘of Farm Economics, vol, 27, 1945. The formulation led to an LP in 77 nonnegative variables, with nine inequality constraints. Stigler did not know of any method for solving the LP at that time, but ne ob- tained an approximate solution using a trial and ‘error procedure that led to a diet that met the mini- mum requirements of the nine nutrients considered in the model at an annual cost of $39.93 at 1939 prices. Alter Professor George B. Dantzig developed the simplex algorithm for solving LPs in 1947, Stigler’s diet problem was probably one of the first large problems to be solved by the simplex method, {and it gave the true optimum diet for this model, with an annual cost of $38.67 at 1939 prices. So the trial and error solution of Steigler was very close to the optimum. Stigler’s original diet model contained no constraints to guarantee that the diet is palatable, and does not allow much room for day-to-day varia tions in diet that contribute to eating pleasure, and hence is very hard to implement, The basic mode! can be modified by including additional constraints to make sure that the solution obtained leads to a tasteful diet with ample scope for variety. This sort of modification of the model after looking at the op- timum solution to determine its reasonableness and implementability, solving tne mosified model, and ‘even repeating this whole process several times, is fairly typical in practical applications of linear pro- gramming Human beings have the philosophy of live to eat as well as eat to live. And, if we can afford it, we do not really bother about the cost of the food. It is ‘also impossible to make a human being eat tho same diet every day. For all these reasons, it is not practical to determine human diel using an optimiza- tion model However, itis much easier to make cattle and fowl consume the die! that is determined as being optimal for them. For this reason, the diet model is used extensively in farms for determining optimal cattle feed, chicken feed, turkey feed, etc. Since the avail ability and prices of the foods keep varying all the time, large-scale farms run their diet model several times each year to determine the optimal feed mix with current data Example 1.4 The Transportation Problem In the transportation model we have a set of nodes or places called sources, which have a commodity available for shipment, and another set of places called demand centers, which require this commo- dity. The amount of commodity available at each source (called the availability) and the amount required at each demand center (called the require: ment) are specified, as well as the cost per unit of transporting the commodity from each source to each demand center. The problem is to determine the quantity to be transported from each source to each demand center, so as to meet all the require- ‘ments at minimum total shipping cost. Consider the problem where the commodity is iron ore, the sources are mines 1, 2, where the ore is produced, and the demand centers are three steel plants that require the ore, with data as given below. Unit Cost of Shipping Ore from Mine to Stee! Plant Amount of Ore (cents per ton) ‘Available at 72 a Mine (tons) Mine + ° 6 2 108 Mine 2 “wo 9 197 Tonsolore 7113398 required at stee! plant The activities in the transportation model are: to ship one unit of the commodity trom source i to demand center j, over i, j. It is convenient to i @ e e e e e e e e@ e e e e @ e@ e e e e e e e@ e @ e e @ e e @ e e e e e e e } baSOCHHSHSHSSSHSNHSHHSHSHSHHSHSHSHSHOSHHHHHSHEHOHSHOHOHOHHOHESHOOSECE represent the variable corresponding to the level at which this activity is carried out by the double subscripted symbol x,, Thus, x, represents the ‘amount of iron ore in tons shipped from mine / to steel plant j, The items in this model are the ore at various locations. Consider ore at mine 1. There are ‘only 103 tons of it available, and the amount of ore shipped out of mine 1, which is xy bX. +X. cannot exceed the amount available, leading to the constraint x4 +%X2+ "iy $103. Likewise, con- sidering ore at steel plant 1. there are at least 71 tons of it required, and the amount of ore shipped to steel plant 1 has to be greater than or equal to this, amount, leading to the constraint x, +3) 271 This leads to the formulation given at the bottom. The Special Structure of the Teansportation Problem ‘As an LP, the transportation problem has a very ‘special structure. It can be very compactly repre- sented in the form of an array, in which row 7 cor- responds to source 7, and column j corresponds to demand center j. The cell in row / and column | of the array, denoted by (/,/), corresponds to the ship- ping route from source / to demand center j. Inside this cell record the decision variable x;,, which represents the amount of commodity shipped along this route, and enter the cost per unit shipped on this, route in the lower right-hand corner. The objective function in this model is the sum of the variables ln the array multiplied by the cost coefficient in ‘the corresponding cell, Record the availabilities at the sources in a column on the right-hand side of the array: and similarly the requirements at the demand centers, in a row at the bottom of the array. Then ‘each constraint in the model is a constraint on the ‘sum of all the variables either in a row or a column, 1.4 Formulation of Linear Programs of the array, and it can be read olf easily trom the array as Shown below. Array Representation of the Transportation Problem ‘Stee! Plant rez) a Miet |x ta | te Mine? foe ae | may 197 Br 213) 29 x20 (or alll,) Minimize cost ‘Any LP, whether it comes from a transportation or a different context, that can be represented in this, special form of an array, is called a transportation problem. In general, the constraints may be equa- tions or inequalities, In a general LP, even when all the data are integer valued, there is no guarantee that there will bbe an optimum integer solution. However, the spe- cial structure of the transportation problem makes the following theorem possible. Integer Property in the Transportation Model Ina transportation model, if all the availabilities and requirements are positive integers, and if the prob- lem has a feasible solution, then it has an optimum solution in which all the decision variables x, assume only integer values, Caution: The above statement does not claim that an optimum sotution for the transportation prob- em with positive integral availabilities and require- ments must satisly the integer property. There may Minimize 2(x) = 9eyy + 16xX,3 + 2Bey5 + 14x3) + 292 + 19X29 Item Subject to Xt at Xo $103 Ore at Mine t Kt rt XS 197 Ore atMine 2 xy + om = 71 Ore at Plant t Xa Xe 2133 Ore at Plant? Xs vay 2 96 Ore at Plants and x20 foralli=1,2, /=1,2.310 1 Formulation of Linear Programs be many alternate optimum solutions for the prob- lem, and the statement only says that at least one of these optimum solutions satisties the integer property. This integer property is proved in Chapter 13, and using it, a special algorithm is developed for trans- portation problems. Example 1.5 The Assignment Problem This problem, known as the marriage problem, was ropc..ed as an application of linear programming to sociology in the early 1950s, We consider a club of sociologists consisting of tive men and five women who know each other very well. Based on their Personal opinions, we are told that the amount of happiness that the ith man and the jth woman derive by spending a period of time x, together is cx, where the ¢,, are tabulated below. N 7 2.3 4~*S 1 7% 6 0 2 oF 2 9 8 ay e 3 & 9 398 4-0 oo 5 7 8a These are the rates at which happiness is acquired by the various possible couples, per unit time spent together. It c, is positive, the couple is happy when together. if c,, is negative, the couple is unhappy when together, and so acquire unhappiness only. Determine how much time each man in the club should spend with each woman, so as to maximize the overall happiness derived by all the members of the club. To keep the model simple, we assume that the remaining lifetimes of all club members are equal, and measure time in units of this lifetime. There are 25 activities in this model, and these are for 1.{=1 to 5: Activity: man i and woman i to spend one unit of time together. Associated level = x,. Thus x, i the fraction of their lifetime that man i and woman j spend together, for i, j = 1 10 5. The items in this model are the various lifetimes fof each member of the club. For example, man 1's, lifetime leads to the constraint that the sum of the {ractions of his lifetime he spends with each woman should be equal to 1, thatis, x;4. + Xi2 + X13 + Xa + X13 = 1. This marriage problem is the following transportation problem Honan] Man t{2)afads Hp) xa | one | ti fet 2 [ote |e | | ee | ee Bf) | ae | oe fae 4 fom | a | oe ft) as fad 5 ee fad - — {or all.j; maximize objective It is a special transportation problem in which the ‘umber of sources equals the number of demand centers, all the availabilities and requirements are 1, and all the constraints are equality constraints, Since all variables are 20, and the sum of all the variables in each row and column is required to be 1, all variables in this problem have to lie between O and 1. From the integer property mentioned eariier, this problem has an optimum solution in which all x, take integer values, and in such a solution all x,, are equal to 0 or 1. One such solution, for example, is io %) (12) In this solution man 1 (corresponding to row 1) ‘spends all his litetime with woman 5 (corresponding to column §), So in this solution we ean think of man 1 being assigned to woman 5, etc. Hence such an integer solution is known as an assignment, and the Problem of finding an optimum solution, which is an assignment, is called the assignment problem. In the optimum assignment, each man lives happily ever alter with the woman he is assigned to and vice versa, and there is never any divorce When | took the linear programming course trom Professor George B. Dantzig at the University of California, Berkeley, he told us the following story about this result. He was delivering a lecture on linear programming one day in the city of Hollywood, tte CO OOOOOOOHOOOEOOOOOOOHOOOOOOOCEESOOOOCOHSOEEEEEthe home of movie stars and glamorous fashion models, the city which probably has the highest divorce rate in the world, That was in the early days of the simplex algorithm, and the lecture was given wide publicity as a talk on a very significant new development, The city's newspaper, the Hollywood Daily, sent one of their seasoned reporters to cover the talk, The reporter sat through the entire lecture, but could not understand any of the theory presented. When Dantzig looked at him at the end, he was sitting with a glum face, which seemed to convey his dilemma of how to write this up as an exciting news story, To cheer him up, Oantzig told him, “Let me tell you an application of linear programming that has ‘sex appeal.” “Now you are talking,” the reporter said with a smile. Dantzig continued, “Our work with linear programming models leads us to a mathem: ‘cal proo! that itis possible for all of us to live happily married without any divorce by choosing our mates: optimally.” "You say without any divorce?” the re- porter asked, and shaking his head in the negative, continued, "Man, you have been working with the wrong kind of models.” ‘The conclusion that there exists an optimum mar- riage policy that maximizes the overall club's happi- ‘ness without any divorce is very interesting. Extend- ing this logic to the whole society itself, one can argue that there exists a solution pairing each man with @ woman in society that maximizes the society's, happiness without any divorce. Natural systems have a tendency to move toward optimum solutions, and it such a divorceless optimum solution exists, one would expect it to manifest itself in nature. Why, then, is there so much divorce going on in society, and why Happiness scared by ‘he coupe Time spent tgethar by a couple Figure 1.2 The inappropriateness of the proportionality assumption in the mariage problem. 11 Formal jon of Linear Programs 11 Js the frequency of divorce increasing rather than declining? This seems to imply that the conclusion obtained from the model—that there exists an ‘optimal marriage policy that maximizes the society's happiness without any divorce—is false. Ifitis false, some of the assumptions on which the model is constructed must be invalid, The only assumptions used in constructing the model are the proportiona- lity and the additivity assumptions. Let us examine ‘each of them caretully ‘The proportionality assumption states that the happiness acquired by a couple is proportional to the time they spend together, that is, it behaves as in the thin line in Figure 1.2. In practice, though, a couple may begin their life together in utter bliss, but develop a mutual dislike as they get to know each other over time. After all, the proverb says: “Fam! larity breeds contempt.” The actual happiness acquired by the couple as a function of the time spent together often behaves as the thick nonlinear curve in Figure 1.2. Thus the proportionality assumption is not reasonable in the marriage problem. The additivity assumption states that the society's happiness is the sum of the happiness acquired by the various members in it. In particular, this states that @ person's unhappiness cancels with another person's happiness. In reality these things are quite invalid, History has many instances of major social upheavals just because there was one single un- happy person. The additivity assumption is quite inappropriate for determining the society's happi- ness as a function of the happiness of its members, Finally, the choice of the objective of maximizing society's happiness itself is quite inappropriate. In determining their marriage partners, most people are guided by the happiness they expect to acquire, and probably do not care what impact it will have on society, It is extremely hard to force people to do something just because it is good for society Political idealogies based on societal objective functions that ignore individual objectives are not really acceptable to many people. This is illustrated by the following definition that appeared in a big city newspaper one day: “The government released a statement that “the current level of unemploy- ment in the country is acceptable”. What it really means is that the government economist who pre- pared this statement is confident that his job is ‘secure and that he is not worried about joining the list of unemployed.”12 +. Formulation of Linear Programs In summary, for studying the marriage problem, the linearity assumptions and the choice of the objec: tive of maximizing the club's happiness seem very inappropriate, and hence models for this problem based on these assumptions can lead to very unnatu- ral or even wrong conclusions. This problem is dis: cussed here mainly to provide an example where the linearity assumptions are totally inappropriate. Comment 1.2: The first formulations of the trans portation problem date back to the late 1930s and the early 1940s, See L. V. Kantorovitcn, Mathemati- cal Methods in the Organization and Planning of Production, Publication House of the Leningrad Stale University, 1939, translated in Management Science, vol. 6, pp. 366-422, 1960; F. L. Hitchcock, “The Distribution of a Product from Several Sources to Numerous Localities,” Journal Mathematics and Physics, vol. 20, pp. 224~230, 1941; and T. C. Koop- mans, “Optimum Utlization of the Transportation System,” Econometrica, vol. 17, nos. 3 and 4 (Sup- plement), 1949, For a detailed coverage on the early history of linear programming developments, see the book by G. 8. Dantzig, [1.61] and his paper {1.8} What Information Can be Derived trom a Linear Programming Model? In Section 1.2, Chapter 2, and in later chapters we present algorithms for solving LPs. Sometimes there may be alternate optimum solutions for an LP. In Section 3,10 we prosent methods for computing all fof them, Using these methods, a suitable optimum solution (one that satisties some conditions that may not have been included in the model, but which may be important) can be selected for implementation. ‘An infeasible LP is one that has no feasible solu- tion, When this happens, there must be a subset of conditions in the model that are mutually contradic tory (maybe the requirements of some items are so igh that they cannot be met with the available resources), and these conditions have to be modified in order to make the model feasible. The methods Proportion of the Metal in Alloy Metal 7 2 7 05 06 2 os. o Setting price of alloy per ton s10 6 discussed in Sections 5.4 and 62 help in identifying such a subset of conditions, After making the neces- sary modifications, the new model can be solved. twill be very usetul for decision making if we can find owt the rate of change in the optimum objective value per unit change in the requirements or the availabilities of each item (e.g., in the fertilizer prob: lem, what is the effect on the maximum net profit, of an increase of 1 ton in the availability of raw material 12) These rates are called the marginal values associated with the items, or the dual vari ables, or the shadow prices of the items, These are the variables in another linear programming problem that is in duality relationship with the orig- inal problem. In this context the original problem is called the primal problem and the other problem is called the dual problem. The derivation of the dual problem is discussed in Chapter 4, The use of the ‘dual solution in a marginal-value analysis (or mar- ginal analysis) is discussed in Sections 4.6.3 and 8.15. When the original problem is solved by the revised simplex method (Chapter 5), the optimum dual solution is also obtained automatically as a by- product, without any additional effort. By pertorm- ing a marginal analysis using the optimum dual solution, the decision maker can determine what the most critical resources are and how the require- ments or resource availabilities can be modified to arrive at much better objective values than those possible under the existing requirements or re- source availabilities. By providing this kind of infor- mation, the linear programming model becomes 4 very valuable planning tool Exercises Formulate the following problems as LPs. 1.1 A nonferrous metals corporation manutac- tures four different alloys from two basic metals, The requirements are given below. Determine the optimal product mix to maxi: mize gross revenue. Total supply of Meta! 3 a per Day oa oa B tons o7 09 4 tons, 18 4 ~ SOOHOHSOHOHHHPHSHSHHOHHOHOHOHHSHHHHHHOHHESOHHOHHOHHOHEOHOEOBESSCHCHOHHSHSHHSOHHSHSHSSSHOHSSHHHSHSHHSHHHSOKHHSEOHHOHOHOHHHOHSOOHOHOCLE 12. Formulations Using Piecewise Linear Functions 13 Paint Type Thinner Tyee Data 7 2 a a 2 Cost (Saal) @ 7 si @ To Viscosity (CP) wo 78007, 2 ‘vapor pressute (PSI) 02 04 oso 30 Britiance content gsiga) | 30 2 9 0 | 0 ° Durability content (asigalj | 2000 1500 1000500 | ° Property Requirement Reason Viscosity S400 ‘So thal a Single coat would éo. Between 15 and 30, Between 2 and 4 PS! 2575 Britance content Vapor pressure Durat ity content So that surtace is properly retlectve For proper drying in 8 For weather resistance, 1.2. BLENDING PROBLEM. Determine an optimum blend for the paint to paint a huge skyscraper. The paint to be used can be obtained by blending four raw paints and two thinners. Data oa them and requirements on the blend are given above 4.2. FORMULATIONS USING PIECEWISE LINEAR FUNCTIONS 1.21 Convex and Concave Functions Let f(x) be a realvalued function defined over points x= (x,,....x)F eP CRY, where Tis either R° or a convex subset (see Chapter 3 for a definition of this term) of RY. Then f(x) Is said to be a convex function if for any x8 = (xh...) P= (4, x2yFef and 0 < a <1, we have Hax! + (1 — abe) < alls!) + (1 aye) (13) fo ed on the eeal line, Figure 1.3 A convex function ‘The inequality (1.3), which defines the convexity of @ function, is called Jensen's inequality after the Danish mathematician who first discussed it, The important property of convex functions is that when you join two points on the surface of the function by @ chord, the function itself lies underneath the ‘chord on the interval joining these points (see Figure 1.3). Similarly, i g(x) is a real-valued func- tion defined on the convex subset I of Rit is said to be a concave function iff for any x', x* eT, and Osa 1, we have glax! + (1 —a)x?) & agtx!} + (1 adgtx*) (1.4) Glearly a function is concave iff its negative is con- vex. Also, a concave function lies above the chord on any interval (see Figure 1.4). Convex and con- cave functions figure prominently in optimization, ‘Some of the optimization algorithms can be mathe- matically proved to work only under suitable con- vexity assumptions on the objective and constraint 4G) Figure 14 A concave function defined on the real tine14 1 Formulation of Linear Programs functions. The reader can verity the following prop: erlies of convex and concave functions, from the detinitions given above. 4A real-valued function defined on the real line (i¢., a function of one real variable) is convex iff its slope. oF the first derivative, is monotone non- decreasing with the variable, It is a concave function if its slope is monotone non-increasing with the variable, 2 Let Ax) fox)... h(t) be all real-valued functions defined onthe convex subset T= R* The function O(x)=ayfy(x) +-*-+ahx) is a linear combination of these functions. It f(x), 1,(2) are all convex and ay,...,.2, 2 0, (x) Is convex. If f,(2)....,4(x) are all’ concave and 2,,-...%2 0, (x) is concave. 6(x) is con- vex if, for every f satistying @,>0, f(x) is convex, and for every t satistying 2, <0, f(x) is In optimization literature, mathematical optimality conditions have been proved, and nice algorithms constructed, for optimization problems in which a Convex function has to be minimized, or a concave function has to be maximized over a convex set. Hence such problems are considered to be nice problems and mathematical programmers reler to them as convex programming problems. I an opti- mization problem requires the minimization of a non- ‘convex function (or equivalently, the maximization of @ nonconcave function), it is a hard problem and, in general, the only known algorithms that may be able to handle such problems might be enumerative algorithms, which could be quite ineticient, particu- larly when the probiem size is large. In mathematical Programming literature, these hard problems are Felerred to a noncanvex programming problems, The reader can verity that the objective function in an LP is both convex and concave. So LPS be- long to the class of nice convex programming prob- lems. We have nice algorithms for solving LPs, and these are discussed in later chapters. A real-valued function 0x) delined on a convex Subset FcR" is said to be an affine function iff itis both convex and concave; that is, if it satisties Dox! + (1 ~ ahx2) = 26 (x) + (1 — 2)8() for every pair of points x", x? in T and 0 $a $1. Given the affine function 0x) it can be shown that there exist feal numbers co, cy,...,c, such that 6lx) = Cy bias boo + Ory Piecewise Linear Functions Defined on the Real Line Let (i) be a real-valued function defined on the interval i$ A$ J of the real line. It is said to be a piecewise linear function it this interval can be Partitioned into subintervals such that in each sub. interval 0(2) is afin: that is, if there exist values Aas Ay =<, Such that OU) = 5 4kA Tory SA + Yous. They are Subject to the constraints OsnsHt Osy.Sxt— (16) osysm—K OSI and for every f, it y, > 0, then each of y, is. equal to its upper bound x} — xf", for all. (1.7) ist When the variables y\.....¥,4; are defined in this manner, z,(x,) is clearly equal to cly, + FO Mery Example 1.6 Consider the ease where z\(x,) has the following slopes: Interval of, 3 10-25 5 2520 7 1.2. Formulations Using Piecewise Linear Functions 15 In this case, x, will be partitioned as y, + ¥2 + Ys where 0 Sy, 510,05; $15,0Sys, and implies y= 10 and y, y>0 ys>0 implies yy Then 2\(x,) can be expressed as 3y, + Sy: + TVs The variable x, can now be eliminated from the model by substituting 4) = y, +--+ yoy wherever it appears in the model. In the objective function 2,(x,) 18 replaced by ely, +--+ 6;"1¥, 44. Notice that the new objective function is linear in the new variables, The constraints (16) on the new variables are included with the other constraints inthe mode! The following facts guarantee that in any optimum solution of the transtormed model. the constraints (1.7) are automaticaly satisiod: (1) We are trying to minimize the overall objective funetion; and (2) the slopes discussed satisty the conditions cf < chen eet IF the slopes do not satisty condition 2, and the objective tunction has to be minimized, then the Constraints (1.7) on the new variables have to be specifically included in the model; since these con- straints are not linear constraints, the transformed ‘model is not linear programming model Example 1.7 . Consider the function z,(x,) defined in Example 1.6. Let x, be any nonnegative number. Keep x, fixed and consider the following optimization problem Minimize ‘Subject to Osy,s10 0 By, + Vy + TVs nent yan 18) 2515 OSys In (1.8), the slopes of the variables yy, yz. ¥s i the objective function are strictly increasing in that order. This implies that the optimum solution for (1.8) Nem H=wH=o W osx,s10 yy=10 y2=x, 10 yy=0: Ex, 10 y2=18 yen, 25it 25 This implies that the optimum objective value in (1.8) is z(x,) specitied in Example 16, for any x, 20, and that condition (1.7) holds automatically in the optimum solution for (1.8). Hence, in any objective function to be minimized, in which z,(*,)16 + Formulation of Linear Programs appears with a positive coetticient, we can replace x, yy. + ¥2 + Ys and z,(x,) by 3y, + Sy + TVs, where the variables y,, y,, ys are constrained by the bounds given in (1.8) I the contribution of some other variable to the objective function is also piecewise linear, partition that variable too in a similar manner. If the original separable, piecewise linear objective function to be minimized is convex, constraints on the new vari- ables of the type (1.7) can be ignored in the trans formed model, and the transformed model is a linear programming model in terms of the new variables. The same technique can be used for transforming any problem in which a separable, piecewise linear objective function is to be maximized, subject to linear constraints, into an LP, provided the objective function is concave (Le, if the slope of each com- ponent function in the objective is monotone de- creasing with the variable) Example 1.8 ‘A company manufactures three different products, 4, 2, 3, using limestone as the basic raw material ‘The company has its own limestone quarries, which ‘can produce up to 250 units of limestone per day at a cost of $2/unit, If the company needs additional limestone, it can buy it from a supplier at a cost of Sofunit. The regional electric utility has recently adopted a modern stepwise rate system to discourage wastage. Icharges the company $30 per unit for the first 1000 units of electricity used daily, $45 per unit for 500 units per day beyond the initial 1000 units, and a hefty $75 per unit for any amount beyond the initial 1500 units of electricity used per day The region's water distribution authority charges at the rate of $6 per unit of waler used per day Up to 800 units, and $7 per unit for any amount used beyond 800 units per day. The company buys fuel from a supplier at the rate of $4/unit, but the energy conservation laws restrict the company from Using more than 3000 units of fuel per day. ‘The company's labor force provide 680 man-hours of labor per day during regular working hours, and the regular wages for these laborers are paid directly by the company’s parent organization and do ‘not cost the company itself anything directly. How- ever, ithe company needs more than 640 man-hours of labor per day, they can get up to a maximum of 160 more man-hours per day by asking the laborers to work overtime, for which the company has to pay itself at the rate of $12 per man-hour. The remaining data in the problem are tabulated at the bottom We will now formulate the problem of determining how much of each product to produce daily, so as to maximize the company's daily net profit. First, let us ‘examine the structure of the objective function. Let Yur Yer Yes Yee ¥e denote the number of units of lime: stone, electricity, water, fuel, labor, respectivaly, used by the company per day. We denote by the symbols 2((y,), 2y.), ete. the costs to the company (Siday) of these respective inputs. See Figures 1.6 and 17 The fuel cost and the sales revenue from the sale of products 2 and 3 are linear. The costs of limestone, electricity, water, and labor are all piecewise linear land convex, but all these things enter the objective function of net profit to the company with a —1 coefficient. The sales revenue trom the sale of product 1 is piecewise linear and concave, and this enters the net profil function with a coefficient of +1 Thus the overall objective function, net profit, is Units of inputs Needed to Produce (One Unit of Product Product | Limestone | Electricity | Water | Fuel | Labor Selling Price T 2 a 7 T 2 | $800/unit for me Test 50 units; $250/unit beyond 50 unite per day 2 1 2 vw yt 1 | $380/unit to an upper limit of 100 units/say a a 5 2 | 3 + _ | sasorunie SOCHHOHCHHOHOHSHOHONHOHOHOHSEOHEOHHOHHLOSa) (s75/unie cont Slope $30 pp, 10001500 Electricity units used per day Figure 1.6 Cost to the company of electricity used daily Its a convex piecewise linear function. Notice that the slope of 2.(y,) 8 monotone increasing with y, piecewise linear concave, and since this is to be maximized, this problem can be modeled as an LP. The primary decision variables in this model are clearly x, = number of units of product / manulac- tured and sold per day, j= 1, 2, 9. The limestone used per day is yy = (14/2) +x, + (9xy/2}, in terms of the primary decision variables. Since the cost of limestone has a single breakpoint at 260 units, we detine the variables ¥4, ¥.2 by 09) OSy: 2) = Wert Ss Here y.,, yu2 Feler to the amounts of limestone used per day in the two intervals for limestone within Which its cost is linear. In (1.9), the variable y, can easily be eliminated. In the following model, the variables Yor. Yeas Yost Vers Yea! Yous Yori Xiu Mia have similar interpretations, Maximize (800%, + 250%, 3) + 350x, + 450x, = (es + Si2) ~ Bes + A5¥e2 + 75Ye9) = ye + Wea) = We ~ 12¥e Subject to 12 Formulations Using Piecewise Linear Functions 17, S250luaie ; ‘ss0arunit 30 Produ 1 sold per day, Figure 1.7. Sales revenue per day from production anc Sale of product 1 111s a piecewise linear concave function, Notice that its slope 1s monotone decreasing KF 8-9, = 0 2x, $Y Ya = x Mea OS ys $280, 0S y5 Oy, 5640, 0S y,, 5 160 05x, $50,05%5 OS 5 100,0Sx, Exercises 4.3. A firm manufactures four products called P, P,, Ps, Py. Product P, can be sold at a profit fof $10 per ton up to a quantity of 10 tons. Quantities of P, over 10 tons but not more than 25 tons can be sold at a profit of $7 per ton. Quantities beyond 25 tons earn a profit of only $5 per ton Product P; yields a profit of $8 per ton up t07 tons. Quantities of P, above 7 tons yield a profit of only $4 per ton. Everyone who buys P, also buys P, to go along with it, This is described later. Both products P,, P; can be Sold in unlimited amounts18 1. Formulation of Linear Programs P, is a by-product obtained white producing P,. Upto 10 tons of P, can be sold at $2 profit per ton. However, beyond 10 tons there is no market for Ps, and since it cannot be stored, it has to be disposed of at a cost to the firm of $3, per ton. P, is a by-product obtained while producing P,. Also, P, can be produced independently Every customer who buy @ tons of Phas to buy Ar tons of P, to go along with itor every 8 2 0. Also, P, has an independent market in unlimited quantities. One ton of P, yields a profit of $3 per ton ifit is sold along with P,. One. ton of P, sold independently yields a profit of $2.50 per ton. Production of 1 ton of P, requires 1h of machine | time plus 2h of machine 2 time. One ton of P, requires 2h of machine 1 time plus 3h ‘of machine 2 time. Each ton of P, produced automatically delivers 3/2 tons of P, as a by- product without any additional work. Each ton of P produced yields 1/4 tons of P, as a by- product without any additional work. To produce 1 ton of P, independently requires 3h fof machine 3 time. The company has 96 h of machine 1 time, 120, h of machine 2 time, and 240h of machine 3 lime available. The company wishes to maximize its total net profit. Formulate this problem as an LP and justify your formulation 1.2.3 Minimizing the Maximum of Several Linear Functions Gonsider an optimization problem that is ike an LP with the exception that there are k different linear or affine cost functions instead of one. Suppose the objective is to minimize the maximum of these k affine cost functions. Let x=(ry,.....%4)% repre: sent the column vector of variables in the problem and let cf -+clx,....ch + cx be the & affine cost functions, where o = (ch, ch....,%) 18 the row vector of cost coefficients in the rth function. We hhave to find an x that minimizes f(x) = maximumfcs + 68x, 63 + Cx, «0h + 8x) subject to the constraints in the problem. We do not have f(x) given by an explicit formula, but for any given vector x, f(x) can be computed easily from the above equation. That is why the funetion f(x) defined in this manner is known as the pointwise maximum (or supremum) of the affine functions cf, + c'x, ¢ to k. This function is not in general separable, but it can be shown to be piecewise linear and convex on rt (see Section 8 14 where this is proved). Since f(x) is piecewise linear and convex, the problem of mini- mizing f(x) subject to linear constraints can be transtormed into an LP. Problems like this often appear in practice. For example, suppose we want to finish a project in minimal time. There may be k different agencies working independently on segments of this project Let x be the vector of variables in the problem, ch +e" may represent the time taken by the rth agency by adopting the solution vector x. Then the overall time taken for the project is f(x), defined above While dealing with an uncertain situation, a method in which decisions are taken so as to mini- mize the maximum cost that may be incurred (or to maximize the minimum profit that may be made) is, known as a worst case analysis method. Worst case analysis often leads to models of this type. To transform this into an LP, define a new vari= able x, .; and introduce these constraints. Kero} ex 20 i Let X=, Xe)", Augment the con- straints in the original problem with these con- straints. Then minimize 2(X) = x, ,, subject to all the constraints. The new objective function is linear in the variables and, hence, the new problem is an LP. WX = (i,q Byes)” 1S an optimum solution of this new LP, % = (8... , 8)¥ is an optimum solution, fof the original problem and %,., = (3) is the minimum value of f(x) in the original problem. Exercises 1.4 Justily the preceding statement. 1.5 Consider the probiem in which ore has to be transported from m mines to n plants. The amount of ore avaitable at the ith mine is a,tons ‘and the amountof ore required at the th plantis bj tons. Ya, = J, 6,. Exactly one truck ean be eccccccccbe SHOOHSHSHHOHSSOHSOHDOHRHHOSHSNHSHSHHSOHOHHHHSSSCE OEESSCHOHSHHSHSHSHHSHHHOSEHSHHSHOCHHSHHSHHOHOHSHSHOHHOHOHOHHOHOHOLEOELE®E used to make these shipments, and the truck is restricted to making no more than one trip from ‘a mine to a plant. What is the minimum capacity truck that can do this job? What is the optimal shipping schedule? Formulate this as an LP. Consider the problem Minimize Hx) = maximum (2x, — 9x3, — 7x, + Bx) Subjectto x, — 95x, 28 9x, + 15x, 2 100 x,andx, 20 When transformed as above, this problem becomes the LP Minimize x3 Subjectto x, —2x, + 3x, 20 yt xy 35028 8x, + 15x, 2 100 x and x, 20 Weg +elx,... 165+ €% are given atfine functions, the funetion defined by g(x) = minimum {c3 + ox, 6 + 6%), is known as the pointwise mini- mum (or infimum) of the functions of +c'x. cj) +6%x. A problem of maximizing the pointwise ‘minimum of several given linear functions, subject to linear constraints, can also be transformed into an LP using exactly similar arguments. The function 9(3) defined above is not in general separable, but it is piecewise linear and concave. Suppose it is required to maximize this function subject to some linear constraints. Define a new variable x,., and introduce the additional constraints, x,.,'—c} ~ clx 50... ker — ch —€% $0, The problem of maximizing x,.; subject to these additional con- straints, and the linear constraints in the original problem, is an LP that is an equivaient problem Mlustration of the Worst Case Analysis Consider the fertilizer product mix problem dis- cussed in Section 1.1.2, Suppose we are determining the product mix for next year. Let the net profits per 1.2. Formulations Using Piecewise Linear Functions 19. ton of Hi-phosphate and Lo-phosphate fertilizers be $c, and $c», respectively. In Section 1.1.2, we had (64, €3) ~ (38, 10). Suppose we do not know at this slage wit the exact net profits from these fertilizers will be next year. However, assume that itis known from market analysis that the vector (¢,, ¢,) will be either (15, 10) or (10, 15) or (12, 12) Assume that the remaining data in the problem will be the same next year as in Section 1.42. If x,, x, are the amounts {in tons) of Hi-phosphate and Lo-phosphate fertilizer respectively, manufactured next year, the total net Profit will be 15e, + 10x, oF 10x, + 15x, oF 124, + 12x,, depending on the fertilizer net protit vector at that time. So, the minimum net profit will be A(x) = Minimum {15x, + 10x, 10x, + 15x, 124, + 1243} Under this uncertainty, the worst case’ analysis dictates that we choose (x,, x3) to maximize A(x) subject to the constraints in (1-1). Transforming this into an LP leads to: Maximize x3 Subjectto x, ~ 15x, — 10 $0 xy 10x, ~ 15x, 50 xy — 12x, 2ey + = 1800 xe x) 51200 (1.19) 1 (%, X3, X3) IS an optimum solution of this LP, then % = (%,, %) is an optimum solution of the worst case problem, and ¥,=h(Z) is the optimum objective value in it 1.2.4 Minimizing a Nonnegative Weighted Sum of Absolute Values In some cases we may have to solve an optimization problem of the form: Find (xy... X) 0 Minimize vod = Fo . ay Subjectto S ayx,6, for where |xJ is the absolute value of x, and all the G; are given nonnegative numbers, The absolute value function |x| is a piecewise linear convex function with its only breakpoint at x,~0 (see tom20 1 Formulation of Linear Programs sib 2 Ie ‘ Figure 1.8 The absolute value function is piecewise i and convex Figure 1.8). So when all c, 2 0, the objective function in (1.11) is separable, piecewise linear, and convex, land hence problems such as this can be transtormed into LPs, The variable x, is unrestricted in sign in the problem. Every real number can be expressed as the difference of two nonnegative numbers. Hence, we can express the variable x, as the difference of two nonnegative variables, as in xp 20 x20 (112) By detining tx 30 itxy>0 ay) tx 20 itx<0 we can even guarantee that xpxy =0 (1.14) thatis, atleast one of and x; is always zero, When defined in this manner, x; is known as the positive part and x; as the nagative part of the unrestricted variable x). if xf and x; satisfy both (1.12) and (1.13), clearly |x| = xj" + xy. By using this equation, (1.11) i transformed into the problem, Find x = (xt. xf) and x = 0 xy) to Minimize U(x, x7 S ete +7) Subjectto Si aya? x)= 6, for/=ttom (145) af 20.x7 20 forall) sfx =0 forall) 6 Let ¥*, #” be any feasible solution to (1.15) that may oF may not satisty (1.16). Let ¥, = min(z,', ¥, ). ap aay ~~ 9 for all j. Clearly (8°, #7) 1 also @ feasible solution to (1.15), it satisfies (1.16) and since all ¢, > 0, ue) WR") (aan) ‘Thus the property that ¢, 2 Ofor all jimplies that i the original problem (1.11) has an optimum solution, then there is an optimum solution to (1.15) that automatically satisties (1.16). When the LP (1.15) is solved by the simplex method, an optimum solution for it that satisfies (1.16) automatically will be ob- tained, provided (1.15) has a feasible solution. From the optimum solution of (1.15) an optimum solution of (1.11) Is obtained by using (1.12). Hence (1.11) is equivalent to the LP (1.15). Example 1.10 This example illustrates (1.17). Consider 2(x,) = 6x) Transforming this as above, we get Ux, x1) = Get + xz), where MPaxpam 20, xP BO (1.18) Give a specific value to x,, say x, ~ ~20. Then in order to satisty the constraints in (1.18), we need (ef =, x7 = 20+ 2), for any a 2 0. In this general solution to (1.18), U(e/, ¥;) has the value of 6(20 + 2a), and the minimum value for U(x}, x;) is attained by setting x 0, leading to the unique solu- tion (x? =0, x; = 20), which satisties (1.16). The same phenomenon holds for any other real value of 1.6 If'some c, <0, show that (1.17) may not hold, ‘and hence the transformation of (1.11) into the LP (1.15) will not work. Note 1.1: it is possible to handle problems slightly more general than (1.11) by the methods discussed here, Let c) + Y4=; 0%, f= 1 to r be ¢ given affine functions. Consider the problem of min- mizing the general objective function oy = Fans Stel + F oiley + F ex1.2. Formulations Using Piecewise Linear Functions 21 subjectto linear constraints on the decision variables A. i= 1100. If, 9, are nonnegative for all jt, this problem can bo transformed into an equivalent LP by the motnods discussed above For this, detine each of the affine functions ¢} + T..¢)x) a8 a new vari- able, say y, by including additonal linear constraints cht Sexy. =0 te ttor (1.19) in the system of constraints for the problem, In terms, of x, and y,, the objective function Ox) = Ww v= YE dy Y Ge + Y gly It x, is restricted to be nonnegative in the system of constraints, then jx,| can be replaced by x,: otherwise, express x, a8 the difference of two non- negative variables, x? —x), and y, similarly as yi = yc. Assuming that f, 2 Ofor all j= 10, g,> 0 for allt = 1 tor, replace [x by x7 + x). replace x, in the linear part of the objective function and in the constraints by x) —a;. replace |y| by yy ty. in the objective function, and replace y, in the con- straints (1.19) by yz —y, (see Example 1.11, below) Example 1.11 Consider the following problem. 1.25 Applications of Linear Programming in Curve Fitting Consider the following problem. In a laboratory a ‘chemical reaction was conducted at various temper- atures and the yield was measured. The data are given below. Tomperawret 588 wo, 92 96 Itis expected that the yield y(t) can be approximated by a cubic expression in terms of the temperature namely, ay+ayt+ayt +a? Let a= (ay, a, a), and fla) =a tat tay tay, The coefficients ay, a4, a;, a3 are the parameters in the model to be estimated using the data, The curve- fitting problem, or the parameter estimation prob- Jom, 18 the problem of determining the optimum values for these parameters that make the value of the fit fla, ) a8 close as possible to the observed yield y(t) over the value of t used in the experiment In almost all practical applications, the functional tori of the curve or surlace to be fitted is determined from practical considerations or theoretical rea- soning, and the values of the parameters in the Minimize (x) = 3x, — 4x, + 8x3 + Spo] + 2)e| + 9]—19 — 124) 4 Sx, — 74] + 19)15 ~ 3x, — 4x, + 6 Subject to Bx, 2x, + 19x) = —9 (120) Be, 3x; + 9x = 10 First we define two new variables y, =~ 13 ~ 12x, +x, — Thy, yy = 18 — Sx, ~ 4x, + Oy, Since all the variables are unrestricted, we then substitute x) = x} ~ x 2, where for j= 1103 andy, fort ¥, ace all restricted to be nonnegative. The transformed problem is: Minimize 8x — #7) = Aix uj + Subject to ay ~ Sa, + 25a; — 1254; —u, +¥, = 80 ag ~ 9a, + 9a, 27a,—u,+¥,= 92 Br at a ayy ty ay aust (12) at at at aus tus Up. Wi 20 forall, ag, ay, ay, ay unrestricted Nao. 4, 1% = 1105) is anoptimum solution of the LP (1.23), then d = (Gp, 4, dy.) is an optimum parameter vector, and (4,1) = d9 + ayt + 4,0" + ayt? is the best fit under this measure. ‘The minimum objective value in the LP (1.23) gives the minimum Vaiue of the measure of deviation L,(a), and Gepending on its magnitude, one can determine whether the cubic ft obtained is a reasonably good ft or not. If itis not considered a good fit, a different functional form or modet can be tried. Chebyshev Approximation Now let us consider the problem of determining the values of the parameters a= (a, ay, a:, a3) that minimize the measure of deviation L,.(a), This is the problem minimize La) aximum {lag + ayt + a3t? + as? ~ yah eS} (124) over all possible real values for ay, &,, a3, 43. By the same arguments as those used earlier, and the discussion in Sections 1.2.3 and 1.24, itis clear that this problem is equivatent to the LP Minimize z Subjectto z 4.20 foralli=1tosay — Say + 25a, ~ 1254, — 4, v= 80 ay 9a, + 9a, ~ Bay vy + = 92 ay a+ ap— ay Uy + ¥y=96 (1.25) 8 Huet v4 = 98 ay+ a+ apt ay—Ustvs= 100 U,.%20 forall, ap, &y, ay, ay unrestricted As belore, if, (a a Ay 2 Oy Wy 1 to §} is an optimum solution of the LP (1.25), 14,1) = ay + At + yf + aye is known as a mini- ‘max solution for the curveritting prablem (because it minimizes the maximum deviation of the cubic fit from the observed yields at values of ( used in the experiment), or a Chebyshev approximation. Least Squares Approximation The problem of determining the values of the param- eters a= (ao, a. 2;, 24) that minimize the measure of deviation L,(a) is minimize L(a) Day + ayt tayt +ay? — yin}? (1.26) over all possible real values for a, a. a3. a For obvious reasons, problem (1.26) is known as @ east squares problem and this method is known as the method of least squares. Problem (1.26) can be solved by setting each of the partial derivatives of Ly(@) with respect to each 9, to zero, since there are no constraints on the parameters a, Curve-Fitting Problems with Constraints on the Values of the Parameters In most practical curve-fitting problems, there may be constraints on parameter values. In that case, the parameter values should be determined to minimize the chasen measure of deviation subject to the constraints. Comparison of the Various Measures of Deviation The L,- and the L,-norms are not differentiable at points in the parameter space where some de Viation term is zero, and this term changes sign depending on which direction we move in the parameter space from that point. The L;-norm does not have these problems and itis in general differen- tiable everywhere, Belore the development of the simplex algorithm for solving linear programs in 1947, most optimization methods were calculus 1.2. Formulations Using Piecewise Linear Functions 23 based and relied exclusively on the use of partial derivatives. Hence in statistics and other applied sciences, traditionally the method of least squares had a historical advantage, and it has become well established as the standard tool for solving curve- fitting problems. However, when there are known Constraints on parameter values, the least squares problem becomes a constrained nonlinear pro- gramming problem. Traditionally this is handled by ignoring the constraints. If the resulting uncon- strained least squares solution violates the con: straints on the parameter values, that solution is somehow modified to another solution which satis fies the constraints. This procedure is in general uunsatistactory, and since most practical curve-fitting problems tend to have constraints on parameter values, it becomes very hard to use the method of least squares to solve them satistactorily, If the parameters to be determined in the model ‘occur linearly in it (as in the example discussed above}, and if the constraints on parameter values are either linear equality or inequality constraints, the problem of determining the best L,- or Le- approximation can be transformed into an LP and solved efficiently by the simplex method (see Chap- ter 2 and subsequent chapters). Nowadays, in most computing centers alt over the world, excelient and reliable computer codes to solve linear programs are available, and they work very well even on fairly large-size problems; hence the use of an Li= or La-approximation (as opposed to the least squares approximation} in linear curve-fitting prob- lems is becoming widespread. If the parameters to be determined appear nonlinearly in the model (lor example, if we wish to approximate the yield y(t) by 4 function of the form g(t) = a cos(B(log 1"), where 2. B. y are the parameters), or if there are nonlinear constraints on parameter values, the use of a least squares approximation with some nonlinear pro: gramming algorithm is preferred (see Relerence [1.30 Uses in Solving Practical Optimization Problems To determine an optimum solution in a practical problem, we first construct a mathematical model of the problem, and then find the optimum solution by solving the model. if the mathematical model involves only linear or affine functions of the decision variables, the model will be an LP and can be solved by the methods to solve LPs discussed in the24 1 Formulation of Linear Programs. sequel, However, the model itself might involve several parameters that may have to be estimated using actual data. Io constructing linear models, this parameter estimation itself can be done by using L,- or L,-approximations and a linear pro- ‘gramming formulation for tinding these approxima- tions as discussed above. For example, in optimizing the operations of a chemical plant we may have a constraint that the yield y in the plant should be above a certain specified minimum level. Sup- pose the yield depends on the temperature ¥,, flow rate x, and prassure x. Suppose we want to approximate the yield by the affine function ay +4yx) + a:x) + 2yxy, where ay, a, a2, a) are the parameters in this model. The best values for these parameters can be estimated from data on the yield using Ly or L_-approximations and a linear programming formulation of it as discussed above. 1.2.6 Modeling with Excesses and Shortages In many practical problems (e.g., inventory mod- cling), we are required to compare a linear function of the problem variables, say, a,x, (which in inventory models may be the amount of production) with a known constant, say, (which will correspond to the demand in inventory models). Y'ax, > b, there is an excess situation and a cost is incurred, which is equal to c,().a,x) — 6}. In inventory models: this will be the Rotding cost. On the other hand, it ax; r) for = 1 to m. Note that a plant cannot operate on avertime unless the regular time is, fully used up. The transportation cost from the ith plant to the jth market is ¢,, dollars per ton of cement. All the cement manufactured during a season can be transported to the market in the same season. At the Ith market there is a demand of d; tons of cement for J=1 to n, There is no restriction that the demand should be met; however, each ton of demand met in market j fetches a price of a, dollars and each ton of demand left unsatistied at market j results in a penalty of 8, dollars to the firm. If the total amount of cement shipped to market / exceeds the demand at that market, the excess can be sold to a warehouse at a price of y; dollars per ton (note y, . Similarly the shortage and excess amounts at the jth market are obtained by expressing the difference between the total supply and demand tor cement at this market asthe diferenco of excess and shortage, both of which are nonnegative numbers, that i Shh ay ~ 4) =u,—v,, and then the net cost at this ‘Market can be expressed as 8,7) ~a,{d,—~¥)) ~ yu) Hence eliminating the constant terms trom the objective function, the overall problem is, Minimize Saud sind + E05, +a, — yu) Subject to Vy Yur rue fori=ttom mn Yxyraytyy forj=1ton YuS a fori=ttom Yu Sb, fori=ttom Xie Vis You Uy Yj ZO for alli, j Exercises 1.8 There are three types of machines that can be used to make a chemical, The chemical can be manufactured at four different purity levels. Data are given inthe table atthe top. The chem- ical Is used by the company internally; if the company does not manufacture enough of it, it ‘can be bought at a price, which is called the penalty for shortage. Assume that there are 30 production days per month at the company. Formulate the problem of minimizing the overall cost as an LP. 1.3. SOLVING LINEAR PROGRAMS IN TWO VARIABLES Les involving only two variables can be solved by drawing a diagram on the Cartesian plane. Given an LP, a feasible solution is a vector that specifies @ value for each variable in the problem, which when substituted satisfies all the constraints and sign restrictions. An optimum solution is a feasible solution that either maximizes the objective function ‘or minimizes it (depending on what is required to be done) on the set of feasible solutions. To solve an LP involving only two variables, the set of feasi ble solutions is first identified on the two-dimensional Cartesian plane. The optimum solution is then identi- fied by tracing the line corresponding to the set of feasible solutions that give a specific value to the objective function and then moving this line parallel to itselt AS an example consider the fertilizer product mix problem (1.1). The constraint 2x, +x, $1500 requires that any feasible solution (x, x,) to the problem should be on one side of the line 2x, + x; = 1500, the side that contains the origin (because the origin makes 2x, +x, = 0 < 1500). This side is26 1. Formulation of Linear Programs indicated by an arrow on the line in Figure 1.9. Like- all the constraints can be represented on the diagram. The set of feasible solutions is the set of points on the plane satisfying all the constraints, ‘and sign restrictions; this is the shaded region in Figure 1.3 For any real value of Zo, 2(x) = Zo represents all the points on a straight line in the two-dimensional plane. Changing the value of z, translates the entire fine to another parallel line. Pick an arbitrary value fof 29 and draw the line z(x) = zp on the diagram. IW this line intersects the set of feasible solutions, the points of intersection are the feasible solutions that give the value 2, to the objective function. the line does not intersect the set of feasible solu- tions, check whether increasing zo, or decreasing it, translates it to a parallel fine that is closer to the set of feasible solutions. Change the value of zo appropriately until a value for Zo, Say 29, is found such that the line 2(x) =z has a nonempty inter- section with the set of feasible solutions. Then go to the next step. In the fertilizer problem, it 2; SPOEEOOOOSESESOHOESSOSOLOOSHEEHHHOSOEHOOHSOHOEHCELESCOCO OOOEEEOOOSEOEOEOOOOOOEOOCHOCOOOLOLOEEEOOBLOCE 3000, the line 2(x) = ~3000 does not intersect the set of feasible solutions at all Also, the line z(x) = zp moves closer to the set of feasible solutions as zo increases from 3000, and when z) =0. the line 2x) = 15x, + 10x; = 0, has a nonempty intersection with the set of feasible solutions. If 2(x) has to be maximized, try to move the line 2x) = 2 in a parallel fashion by increasing the value of Z» steadily from 2 as far as possible white stil intersecting the set of feasible solutions. W 4, is the maximum value of zq obtained in this process, it is the maximum value of z(x) in the problem and the set of optimum feasible solutions Is the set of feasible solutions that lie on the line 22) = Zo ‘On the other hand, if the line 2(x) = 2) has a nonempty intersection with the set of feasible solu- tions for every 2225, then z(x) is unbounded above on that set. In this case 2(x) can be made to diverge to +00 and the problem has no finite optimum solution. If the aim is to minimize 2(x), then steadily de- ‘crease the value of zp from z) and apply the same kind of arguments. For the fertilizer problem, as 2, is increased from ©, the line 15x, + 10x, = zy moves up until 2 19,500. For any value of za > 13,500 the line 2(x) = 29 does not intersect the set of feasible solutions. Thus.the optimum objective value is $13,500 and the optimum solution of the problem is (xy. x3) = (800, 900), Exercises 1.9 Solve the following LPs by drawing diagrams. (a) Maximize 30x, + 20", Subjectto ot 2B 1 (b) Minimize Subject to 14 Modeling Multiple Objective Problems 27 (6) Minimize x, + 3x Subjectto 2x, + x, 2 10 a+ 4 S20 x)= 2 S10 xt 590 x20 420 (6) Minimize 10x, + 3x, Subjectto x, + x) 220 x 5 6 x 22 x S12 net fe) Maximize 2x, — 2x Subjectto 2x, + $2 81 x20 () In the following problem show that the variables x5, x, x5 play the roles of slack variables. Show that the constraints in tho problem can be transformed into inequality constraints in the variables x, x3, by ‘eliminating these slack variables. Use this and solve this problem by drawing a diagram. Minimize x, 3x, Subject 0 2x, + x, 4x5 at - for all = 1108 Note 1.2: LPsin higher dimensional spaces (23) cannot be solved by drawing diagrams, but the simplex algorithm of Chapter 2 provides a systematic procedure for solving them. However. we will gain a lot of intuition on how the simplex algorithm works by referring to and visualizing the concepts of ‘geometry. 1.4 MODELING MULTIPLE OBJECTIVE PROBLEMS In all the problems discussed so far, there was a single, well-defined objective function to be opti mized. In some problems the decision maker may28 +. Formulation of Linear Programs want to optimize two or more objective functions simultaneously; these are called multiple objective problems. Example 1.13 Consider the fertilizer manufacturer's problem for- molated in Section 1.1.3. The fertilizer manutacturer leacly wants to maximize its profit 2"(x) = 15x, + 10x,, which is in dollars. Now, suppose that in the world of ertilizers the prestige of a company goes up with the amount of Hi-phosphate fertilizer it manu factures. Our manulacturer may consider it very important to maintain the image of its company as a prestigious company, and in this case it should also want to maximize the amount of Hi-phosphate fert izer produces by its company, which is 2%(x) = x,, in tons. In this case, we have two different objective functions 2'(x) = 15x, + 10x, and 24(x) = x;. both of which the manufacturer would lke to see maximized simultaneously, subject to the constraints derived in Section 1.1.3. This is a multiobjective problem In a multiobjective problem, if there exists a feasible solution x', which simultaneously optimizes every one of the objective functions, x* is clearly the optimum solution for the problem, Unfortunately, this phenomenon is quite rare. There is usually no single feasible solution that simultaneously provides the best value for each objective function. AS an it lustration, in Example 1.13, 2"(x) is maximized by the Unique feasible solution z = (300, 300), yielding a value to 2'(x) = $13,500. The problem of maxi mizing 2%(x) has alternate optima, with the points x{2) =(600, 2) 05 $ $00, all being optimal to it, giving a value to 2%(x) of 500 tons. Clearly, there is no single feasible solution in this problem that simultaneously optimizes. (here maximizes) both 2(x) and 2%), Very often, in multiobjective problems, the various objective functions conflict with each other. An op: timum solution for one objective function might turn ut 10 be. very undesirable solution for another objective function. This conflict among the various objective functions in a multiobjective problem is illustrated by the cartoon in Figure 1.10 There is no theoretically satistactory concept of optimality yet in multiobjective problems in which the objective functions contlict with each other, but many practically reasonable approaches have been developed for handling them. We discuss these ap- proaches here brietly. We begin with a few detini- tions first Let K denote the set of feasible solutions (ie. set of solutions satistying all the constraints and restrictions on the decision variables), and (x) P(x)... OX), the various objective functions in a multiobjective problem, We assume f 2 2, as other- wise it is a single objective problem. Then the vector fx) = (P10), POD, "09) is Known as the vector- valued objective (or criterion) function. in this prob- lem. A feasible solution ¥ €K is said to be a vector minimum, if there exists no other feasible solution x€K that satisfies (x) <1"(2), for r= 1 to {with the inequality holding strictly for at least one r, vector maximum, if there exists no other feasible solution xeK that satisties 1"x) 2 1"(2), for r= 1 to f,with the inequality holding strictly for at least one r So, given a vector minimum x, any move trom x that strictly decreases an objective function from its value at ¥, automatically leads to a strict increase in another objective function from its value at % ‘Thus a vector-minimum solution may be a desirable solution to seek when itis required to minimize each objective function in the problem. A similar argument shows that a vectorsmaximum solution may be a desirable solution to seek when it is required to maximize each objective function in the problem. Example 1.14 The set of feasible solutions of the fertilizer problem formulated in Section 1.1.3 is drawn in Figure 1.9. For the two-objective version of this problem discussed in Example 1.13 above, it can be verified that every point on the thick line segment joining x! = (900, 900) and x? = (600, $00) is a vector-maximum feasible solution. A vector maximum over K for the vector function (x), is-a vector minimum over K for the vector fune- tion f(x) and vice versa. We now discuss some of the approaches that can be used to handle multi- objective problems, 1.4.1 The Vector-Minimum Approach Let f(x) = (F(X)... £109) be the vector of objective functions. If it is required to make one of these SOROS OHOOOHNSHSHSOSSHOHSHSHHHOKHHOKCHOHSEHOOHHHHOHHHOHHOHBLOSOSPCOHHHCOHOHESHSCHOOCOESEKECHOHO HOHE EET HOCH SHOOKEOSOOEELE 14 Modeling Multiple Objective Problems 29 Figure 1.10 A multiple objective problem with conflicting objective functions, (Orawing by Ziegler, copyright © 1980 The New Yorker Magazine, Inc). functions as large as possible, replace it by its negative in the vectar. Atter this modification, every function in the vector f(x) is required to be made as low as possible, Then we could take a vector mini mum of f(x) over the set of feasible solutions as @ solution to this multiobjective problem. In this case, the vector minimum is also known as a pareto optimal solution or efficient solution or a nondomi- znated solution, and this method of choosing a vector minimum as the solution is known as the pareto optimality criterion. (One drawback of this approach is that even though the concept of vector minimum is mathematically very appealing, usually there are many vector minima, and there exists no satistactory theory for ‘comparing and choosing the best among them. tall the constraints are linear, and each objective function in f(x) is affine, the problem of finding a vector minimum for f(x) over the feasible region is known as a linear vector-minimization problem. In linear problems a vector minimum can be found very ctficiently, see Chapter 17 or references {1.37—1.48} and efficient methods even exist for computing all the vector minima, However, so far, these methods have remained as mathematical tools that have not found many uses in practical applications. 1.4.2 Approach Based on Ordering the Objective Functions in Priority Order Let f(x) = (F(a), ...1"()) be the vector of objective functions, each of which has to be made as low as possible, This approach can be used if there exists @30 1 Formulation of Linear Programs clear priority order among the objective functions. so that one of the objective functions is recognized as the most important among all, another is recognized as being the most important among all but the first, etc. Let f(x)... Fx) be the objective functions in this priority order. When such universally recognized priority order exists arsong the objective functions, proceed in stages as follows, First, in stage 1 mini- mize (x) (the most important objective function) over the set of feasible solutions. If the optimum solution for this problem is unique, accept it as the ‘optimum solution for the multiobjective problem and terminate. Otherwise let K, denote the set of alter~ rate optima, each of which minimizes M(x) over the original set of feasible solutions. Go to stage 2 In general, if the method does not terminate in stage 1, we let K, denote the set of alternate optimum solutions of the stage r problem. Then in stage r + 1, minimize f"*1(x) over K,. If the optimum solution of this problem is unique, terminate by accepting it as the optimum solution of the multiobjective problem, Otherwise fet K,., denote the set of alternate ‘optimum solutions of the stage r + 1 problem. Go to stage r + 2 and continue in the same way. If all the constraints are linear, and each objective function is linear, this approach can be implemented very effi- ciently using the simplex algorithm for solving linear programs (see Section 3.10 and Exercise 9.72). If a natural priority order exists among the objective functions, this is the best approach to use. This is known as a sequential procedure, and it can be shown (see Chapter 17, in particular Exercise 17.6) that the solution obtained under this procedure is a vector minimum for this multiobjective problem Requirements tor Using the Other Approaches All the other approaches for handling multiobjective problems discussed below require that each ob- jective function be expressed in some common comparable units. For example, in Example 1.13, 2x) = 15x, + 10x, measures the profit in dollars, and 2%(x) = x, measures the Hi-phosphate fertilizer production in tons, and these two objective functions are not comparable. However, we can change the second objective function into the profit derived in dollars from Hi-phosphate fertilizer production, PQ) = 15x,. Maximizing 2°(x) is the same as maxi- ‘mizing f(x) and vice versa. When the vector of ob- jective functions for this problem is taken to be 2'(x) and f(x), both objective functions are in comparable units of dollars. Using similar ideas, itis often pos- sible to express all the objective functions in a multi- objective problem in common units. We assume that this is done for using any of the approaches dis- cussed below 1.4.3 The Worst Case Approach Let (x) = (00)... (@)) be the vector of objective functions, each of which is required to be minimized, all the objective functions being in common units, De- fine for each x, 6(x) = maximum (f(a)... (2)}. the pointwise supremum function ofall the objective func tions. This approach takes the feasible solution that minimizes 8(2) as an optimum solution of the mult objective problem. if all the constraints and the ‘objective functions are linear, the problem of mini- mizing 6{x) can be transformed into an LP as dis- cussed in Section. 1.2.3. This approach tries to minimize the maximum of the different objective values, and hence can be interpreted as a worst case approach. 4.4.4 Approach Based on Constructing a Weighted Average of the Different Objective Functions Let 09 = (Pa). f()) be the vector of objective functions each of which Is required to be minimized, all of them being in common units, For r=1 to f determine a weight, 4,, for the rth objective fune- tion, 0 Metatturaica! coke Ls coke screenings ‘The raw material and labor availability exhibit ‘seasonal variations. During the months 10, 11, 42, 1, 2,3 the company can hire at most enough labor to produce 1200 and 600 tons per month during regular time and overtime, respectively. In months 4, 5, 6, 7, 8, 9 these labor capacities are 800 and 500 tons, respectively. The prod: uct manufactured during a month can be sold anytime during the next month or later. Stor- ‘age costs are $1,00/ton from one month to the next for the product. Raw material cannot be Conor tabar Lint on aw {ton 1 Froucton tte arity (ensign oman omand Sion pie swonn | Regie tne | Ovotine | _ ‘one ot prod ton) ten) 7 cy % 29 9 n 2 | ring ese © ‘mo ‘8 a | one we tos " 4 7% ® ‘00 $00 a ay ‘0 ES % | fone to oo» FA 7 0. oo» & ‘ "00 oo» = * t Ee s v0 & = 0 ‘20 ~» h ‘ too 2 ke so vo ®38 1.23 1 Formulation of Linear Programe stored. It has to be used up in the month in which it is obtained. Operations begin in month 1 with a stock of $0 tons of the product. At the end of month 12 the company should have a stock of at least 50 tons of the product Determine an optimum production schedule. 3. An agency controls the operation of a system consisting of two water reservoirs with one hydroelectric power generation plant attached to each, The planning horizon for the system is. a year, divided into six periods. Reservoir 1 has a capacity for holding 3500 kilo acre-tt(ka-tt) of water and reservoir 2 has a capacity of 5500 ka- ft Atany given instant of time, ifthe reservoir is at its full capacity, the additional inflowing water is spilled over a spillway. Spilled water does not produce any electricity, During each period some specitied minimal ‘amount of water must be released trom the reservoirs to meet the downstream require ments for recreation, irrigation, and navigation purposes. However, there is no upper limit on the amount of water that can be released from the reservoirs. Any unreleased water is stored (up to the capacity of the reservoir) and can be Used for release in subsequent periods. All water released from the reservoirs (even though it is released for recreation and other purposes) produces electricity Itcan be assumed that during each period, the water inflows and releases occur at a constant rate. Also, on an average 1 att of water released trom reservoir 1 produces 310 kWh of electricity and 1 att released from reservoir 2 produces 420 kWh. At the beginning of the year 1.24 reservoir 1 contains 1800 ka-tt of water and reservoir 2 contains 2500 ka-tt of water. The same amounts of water must be left in the respective reservoirs at the end of the year, The electricity produced can either be sold to a local firm (called a class | customer) or to class Il customers. Class | customers buy electricity on an annual basis, they require that specified percentages of it should be sup- plied in the various periods. They pay $10 per 1000 kWh, Class II customers buy electricity on a period by period basis. They will pur- chase any amount of electricity in any period at $5 per 1000 kWh. The other data for the problem are given in the table at the bottom. Operate the system to maximize the total an- nual revenue from the sale of electricity. ‘A company has a permit to operate for tive seasons, It can manufacture only during the first four seasons and in the fitth period itis only allowed to sell any leftover products. it can manulacture two types of products. One unit of product 1 requires five man-hours in preparatory shop and three man-hours in finishing shop. Each unit of product 2 requires six man-hours in preparatory shop and one man-hour in finishing shop. During each season the company has at most 12,000 man- hours in preparatory and 15,000 man-hours in the finishing shop (only during the first four seasons). The product manufactured during some season can be sold anytime from the next season onward. However, selling requires some marketing effort and it is expected that 0.1 and 02 man-hours of marketing effort are required to sell 10 units of Inflows In Kilo Aero Fe into Reservoir Minimum Release rom Reservoir Percentage of Annual Enargy Sold to Class | Customer Perios 1 2 1 2 10 Be Delivered in Period 7 ary 200 ~«G0 10 2 tart 235 200578 2 3 se tam 200 995 18 4 46 731 2001495, 22 5 32 an 200 ssa 2 6 159 497 200392 10 —(S. Parikh)Exercises 39 Expected Selling Price per Unit of Normal Rate Maximum Man-Hours Overtime Rate Season per Man-Hour at Normal Rate per tan-Hour Prosuct) Product 2 2 & +00 320 320 35 3 sé 300 seo 25 40 4 st 00 $20 30 40 5 sio 1000 $20 19 30 products 1 and 2, respectively. Man-hours for 1.28 A contractor is working on a project, work on 1.25 ‘marketing effort can be hired at the rates given in the table at the top. There is no limit on the fumber of man-hours that can be hired for marketing effort at the overtime rate. fa unit of product is available for sale during a season, ‘but is not sold in that season, the manufacturer thas to pay carryover charges of $2 per unit to put it up for sale again in the next season. ‘The selling prices in the various seasons are given in the table above. How should the com= pany operate in order to maximize its total profit? ‘The table at the bottom gives the composition which is expected to last for a period of T weeks. It is estimated that during the jth week, the contractor will need u, man-hours of labor, J = 110 T, for this project. The contractor can fulill these requirements either by hiring laborers aver the entire T week horizon (called steady labor) or by hiring laborers on a weekly basis each week (called casual labor) or by employing a combination of both. One man- hour of steady labor costs ¢, dollars; the cost is the same each week. However, the cost of ‘casual labor may vary from week to waek, and it is expected to be ¢,, dollarsiman-hour, during wok] ]= 1 How ean he tlt abot Uf warious ode woes in makingcoreste. Tre ECR. = 110 T- How can efi slab other material in each food is fiber, water, etc, a ‘The company blends these food materials and —(S. Kedia and G, Ponce Compos) mnakes ho kinds of cereals Inthe process imakes tna Kinds cereals Me 003 tay pare are m retires wh te Mh reinay of minerals and vitamins are completely lost having the capacity to supply @; gal of fuel from the mix. For each 100 kg of foods added ree are ° cues “en 8 Geran tor ins pee in the blend, the blending process adds 5 kg and the demand in the jth city is b, gal fis inte lang, he blending proces a the tacon ot a galt uel coved’ Cereal type 1 sells for $1.50/kg. It should asprin wn ‘a of Tue traves contain at least 22% of protein, 2% of minerals. e inoty bn city f a a delivery truck. Find a and vitamins, and at most 30% of starch by ‘etal es! vote schedule that aed the weight. Cereal type 2 sells for $1.00/kg. It Q om of fuel consumed by the delivery ‘should contain at least 30% starch by weight, ucks, What ioe optimal prodec! mt lor the 128. A tamer has tree tas. He can grow the company? Gitoren crops on them Daa for the coming calgon oy Won ine Germs —_vaby fond ___Proein__Serch__‘aminace. tra’) per Day : 6 @ 7 2 of v0 2 f = : = oa 0 5 2 % 2 a oat 0 ‘ z iB 3 & oss mo40 1. Formulation of Linear Programs Water Available in a-t Gost per att of Water Beyond Farm Usable Acreage at s25ra-tt ‘Amounts in Previous Column 7 “400 1200 $50 2 600 200 $60 450 1100 $70 Maximum Acreage Water Consumption Yieia crop Farmer Can Prant (asta) unitsia Selling Price per Unit a 500) 6 30 '825/anit up to 10,000 ‘$20/unit beyond 10,000 units 8 700 5 100 ‘t5/unit up to £40,000 units ‘$8/unit beyond £40,000 units c 380 4 75 45/unit up to 20,000 units 42/unit beyond 20,000 units season are given in the table at the top. To harvested. The farmer can sell corn at the rate maintain a uniform work load among the three of $2.50 per bushel in the wholesale market. farms, the farmer adopts the policy that the He can also raise poultry. Poultry is measured percentage of usable acreage planted must be in poultry units. To raise one poultry unit re- the same at each of the three farms, However, quires 25 bushels of corn, 20 man-hours of any combination of the crops may be grown at labor, and 25 tt! of shed floor space. He can any of the farms. The only expenses the farmer either use the corn that he has grown himself counts are for the water used at the various ‘or buy corn trom the retail market. He gets farms. How much acreage at each farm should corn at the rate of $3.50 per bushel trom the be devoted to the various crops to maximize the retail market, He can sell at the price of $175 farmer's net profit? Per poultry unit in the wholesale market up to 1.29 Atarmer can lease land upte maximum of 1000 £200 units. Any amount of poultry over 200 units a. He has to pay $5 per acre per year it he leases up to 600 a. Beyond 600 a, he can lease at $8 per acre per year. He grows corn on the land. He can grow corn at the normal level or at an intense level (more tilizer, frequent irrigation, etc.) Normal level yields 70 bushels per acre. Intense level yields 100 bushels per acre. The requirements are given in the table at the bottom. Harvesting requires 0.5 man-hours of labor per bushel Requirements per Acre per Year Labor (man-hours) IMatorats (seed, ertiizer, water etc.) sells for $160 per unit. He has only one shed for raising poultry with 15,000 1? of floor space He and his family can contribute 4000 man- hours of labor per year at no cost. It he needs more labor, he can hire it at $3 per man-hour up to 9000 man-hours. For any amount of labor hired over 3000 man-hours, he has to pay $6 per man-hour. Maximize his net profit 1.30 A.company makes products 1 and 2. Machines 1, 2, and 3 are required in the manufacture of ‘Normal Level ‘Intense Level 6 2 $20 $35 SOHHOHHSHOHOHSHOHSHHHSHSHNHSHHHHHHSHHHOHOHHOHOHHHOHHHHHECOHCOEOESthese products. The relevant data are tabulated below. The manufactured products may be within specification or not. Product i, which is within specifications, is denoted by W,, and the same product that is outside the specifications, is denoted by R,,/ = 1,2. Both W, and R can be sold in the market. The market price of Ris, c/unit for i ~ 1, 2. W,can be sold olther as W, or itcan also be sold as R,. IfW, is sold as W,, it sells forc,,/unit, andiifitis sold as R,, itis sold at the same price as R,. However, R, cannot be sold as W, When product /is manufactured, an average proportion p, comes out as W,,/ = 1.2.L,.é,are the minimum demands for W, and R,. respectively. K, and k, are the maximum amount ot W, and R,, respectively, that can be sold. Determine a production and marketing schedule for the company that maximizes its to all the folal sales revenue subject ‘constraints Machine-Hours Required fo Manufacture One Unit of Product on Machine Product 1 28 1 38 2 2 7 45 Machine-hours available fn the machine 350 9701420 131 1.32 Exercises 41 There are 14 food ingredients available. They can be classified into seven groups. It is required to mix them into a minimum cost diet satistying the following constraints: (i) Protein content should be at least 20%; (i) fat content should be at least 3%; (i) fiber content should be at most 12%; (iv) Ca content should be between 1% and 2%; (v) Ph content should be between 0.6% and 2%; (vi) Ca content should bbe greater than or equal to the Ph content Data and other constraints are given below. A company manulactures products 1, 2, and 3 with labor and material as inputs. Data are given in the table at the top of page 42. The material required in this process is obtained from outside, There are two suppliers for this material. The amount of material availabie, called x, in the table, depends on the supplier chosen. The small supplier can provide at most 25 tons of material per month, The big supplier will only supply if at least 35 tons of the material is taken per month. By company regulations, the company has to choose one of these two suppliers and get the material only from it. The company would like to adopt the policy that maximizes ils monthly profit. Can this be found by solving one LP? It not, discuss, how you can solve this problem. Range of Range ol Percent of Nutrient Contant cost nornuiont Grose Group Food Ingredient | Protein Fat Fiber Ca__—Ph ‘| porton (9%) we 1 1 Oried beet pulp 9 05 2% 07 005| $64 41020 Stoo 2 Dried citus pulp 6 3 % 2 at | 3 — tt020 2 3Ground yellowcorn | 8542502925) $5 Tt0.25 ONO 4 Ground oats 2 45 2 of o4 | sa ttoas 3 8Corn molasses 35 06 01 | 19 Sto18 —Sto1d 4 Wheat midatings we 4 8 oF o9 | 6¢ Sto20 101020 T Wheat bran we 4 5 01 12 | — 5t090 5 BDistiller grains % 85 9 015 068 | 77 Sits 21025 ‘9Corn gluten teed m2 8 03085 | 11025 6 10Cottonseed meal a 4s ora |e 103531035 11 Linseed olt meat a1) 8035 8 | Bs 11095 12 Soybean oil meal 4 05650208 | 8 tt 85 7 18Ground limestone 3 © 05 | 10 Oto 14 Ground phosphate a 6 _tto2 (G. Brigham)42 1. Formulation of Linear Programs Thoms Requirement to Produce 1 Ton of Product tom ' {Labor (man-hours) Material (tons) Profit ($!t0n) 1.33. The following table gives some of the data 134 1.95 collected during a ballistic evaluation of a new armor plate material, tis the thickness of the target armor plate in inches and v is the striking velocity of the projectile (in feet per second) required to penetrate the plate. vis the control parameter. Estimate tthe thickness of the plate that a projectile with striking velocity v would penetrate) as a function of v. Use the data to fit the following curve: t= av + ay(ayy)", where 4), 43, @ are the coefficients to be estimated a), @, ay should be all nonnegative and a, should lie between 0 and 1, Discuss a practical method for obtaining the values of ay. ay, ay that give the best fit, v.00 2800~—aas0~—«ma00 ~~ 4087 4596 4645 4.697 5.002 SEPARABLE CONVEX PROGRAMMING PROBLEM. Consider the following problem: where f(x) = 6x7" and ¢,, a), 4, k, are given positive numbers, and 0-< 4 0 be a given vector of positive weights. For each vector y (us---sJq) ER", detine the function Liy) by LU) = maximum {uy i= 1 ton}. Given a vector a= (a4... 0,) ERY itis required to find a vector x = (x...) to minimize L(a 9), Subject t0 x; —Xi4) $0, i= 1 ton ~1 Formulate this problem as an LP. A probiem of this type arises when tying to estimate the failure rate ofan item as a function ofits age, An experiment can be run with several new items putin use at time 0. Time can be divided into unit intervals, ana we observe the number sti ‘working atthe beginning of each interval. fis the observed number of items still working at time point f,~ f. 8 the number of items that failed inthe ith interval, and ((, fs )/t) = a, is the observed failure rate at the age of i time units, Suppose trom practical and engineering considerations we know that the failure rate increases with age (this is known as the IFR, or the increasing failure rate property). The actual failure rates observed a... a, may not be increasing due to random flyetuations in the observations. An optimum solution x of the above problem provides a vector satistying the IFR property that is closest to the observed17 vector of failure rates a. and hence x can be taken as the estimate of the vector of failure rates of the item at various ages. How does the formulation change if each x, is required to be nonnegative, in addition, and also less than or ‘equal to one? IV. A. Ubhaya (1.32, 1.93}) COAL DISTRIBUTION PROBLEM. Formulate this problem as an LP. A region contains a large umber of small individual coal mines. The quality of coal varies trom mine to mine considerably because of the heterogeneous nature of coal production in te area. The major coal market in the region is for thermal generation of electricity that requires large guaranteed deliveries of coal, consisting of relatively uniform grades. The objective is to indicate mine operating levels to the coal producers, such that their net delivered price of the make-up coal on a BTU basis is maximized, subject to the condition that the mixed coal product must satisfy the ulilily requirements Let x, denote the tons of coal shipped from mine / (total of m mines) to the jth collection point (a central collection points). We are also given the values of the following data ele. ments; q, = daily production capacity (tons) of | 0:m; b = projected daily demand (tons) by the utility: a,, 5, Wis Was fu 6, = percentage (by weight) of ash, sulfur, moisture, volatile matter, ash fusion, and fixed carbon, respectively, in the coal from the ith mine; a, 8, w ~ permissible maximum ash, sulpher, and moisture percentage (by weight), respectively. in the mix; v, f= permissible minimum volatile matter and ash fusion percentage (by weight), respectively, in the mix; c), 6, =permissible lower and upper Exercises 43, bounds for percentage (by weight) of fixed carbon in the mix; 2, = delivered price of coal at collection point per ton; ,, = shipping charges Per ton of coal trom mine / to collection point | (6. Manula and ¥. Kim {1.24)) 1.98 PARTICULATE BLENDING OF CHEMICALS: There are { balches of chemicals available Te following data are available. wy ~ weight (tons) of material in the kth batch, ay, aan, fq, (071d = 1) 18 the fraction of material (oy weight) in the kth batch in the various partcie S25; by, by... .bal50, = 1) is the desired fractions of blend (by weight) in the various article sizes; w = weight (tons) of the blend to be produced. It is desired to blend material tom the various batches, so as to produce a blond with fractional size-weight distribution as close tothe desired as possible. Formulate this problem as an LP. Discuss how the form lation changes if it is required to minimize the ‘maximum deviation trom the desired fraction in the blond over all the sizes. In particular, formulate the problem with the data at the bot- tom of the page 1.99 THE BETTING PROBLEM. A linear pro- ‘grammer is planning to spend a day at the race course. He will have $b available for betting. On the day ot his proposed visit, here is only one race planned in which k horses numbered 1 2,....k are competing. ithe bets a dollar on the ith horse, his gross payott from this bet will be 0 if the ith horse does not come first in the ace, or $2, ifthe ith horse comes first in the race, for! = 1tok. The numbers 2y,..., 4 are known positive integers. He is, of course, allowed to bet any nonnegative amount on any number of horses. His problem is to determine how much to bet on each of the horses in the Fracion by Weight of Various Sizes Batch 1 2 7 O10 O40 2 020 0.10 3 020 08 4 oa a5 5 oe a7 Desired traction in blend 0480.90 34 Tons Available o30 020100, 050 020 250 oo2 010 «50, 030015500 org 00s 200, 015010 800 tons required (A.A. Klimpe! (1.20))44 4) Formulation of Linear Programs race (i.e., how to divide his available $b for bets fon the various horses) s0 as to maximize his minimum net gain, irrespective of whichever horse comes first in the race. Formulate this as an LP. Write down the special case of this problem when b = 100, k= 5, and (24, 22, 25 4, a5) = (2, 5, 1, 7,3) (8. Vajda (1.110) 1.40 A PRODUCTION PLANNING PROBLEM. A 141 ‘company has to make deliveries of a product to its dealers at the end of each month, At the beginning of the year the company knows that they have to deliver d, tons of this material at the end of the /th month in that year, i = 110 12. Material produced during a month can be Gelivered either that same month itselt or stored in inventory and delivered in some other month. Itcosts the company $¢,/ton to carry the material in inventory from one month to the next. The year begins with 0 tons in inventory and is required to end with O tons in inventory at the end of month 12. The company has to determine how much material should be produced each month of the year. If the company decides to produce x, tons in month 1 and x,, fons in month i + 1, it incurs a cost of C_]x, ~ %,.| 0 adjust the production levels at the beginning of month i+ 1. Itis required to determine how much to produce in each of the 12months of the year soas to minimize the total of inventory and production-tevel adjustment costs over the year. Formulate this problem as an LP. Write down the problem for the specific case when d = (30, 40, 20, 70, 80, 90, 100, 30, 150, 200, 150,150), and ¢, ~ 10, ¢ (A. J Hottman and W. W. Jacobs (1.16}) Consider the following system of linear Inequalities, It is required to find a feasible solution x to this system that makes each of the four constraints to be satistied as an equation as nearly as possible. Formulate this problem as an LP X- 2et Que Kt STE Dey + yA Way t Axe t Bx 4x, Bx, + 19K; HBA, + TK, + 5x5 + 12g Sty + Bxy— Bx # 19K + Bx, ~ Ty x20 forallj~tto6 1.42 Consider the following system of linear 1.43 ‘equations. It is required to find a vector x that makes the left-hand side in this system as close to the righthand side componentwise as possible, Formulate this problem as an LP. Xt Wns Bt = 19 2ey— Sry t Ary + Sx Axy = 18 = F181, — 10x, Bx, = 22 1, 4 13x, 4 22K, — 17x; = —10 Yx, — 12x 10xy~ 2 -8 Bye Xt 3x Aten, = 13 KW + 1g + B+ Oy = It AGGREGATE PLANNING PROBLEM IN A STEEL ROLLING MILL. This application arises in the production planning process of a large steel rolling mill, where approximately 4000 to 6000 cast rolls of 300 different types, each requiring an average of 10 operations. are machined in a large job shop each year. The order book is first balanced in an aggregate manner using the linear programming model discussed here, such that monthly or other periodic demands are approximately in tine with the manufacturing capabilities. Detailed Production scheduling and planning is then carried out using heuristic methods. We are given the following data and delinitions of the decision variables for the aggregate planning problem. The constraints are the machine-time availability constraints and supply-demand constraints (the number of rolls ofa given group machined and shipped in a period cannot exceed the demand for the group in that period). The objective is to maximize the total tonnage of all rolls machined and shipped during the planning peried. ignoring the integer requirement on the variables s,,, formulate this problem as an LP. Data given are: N = the umber of months (or periods) into the future for which planning is being done; M = the number of groups of rolls under consideration during the planning period; K = the umber of different types of machines; 0,, = the number of rolls of group i demanded in period j; Py = the processing time (hours) required by each roll in group Fon machines of type k: Ay = the processing time available in hours in period j fon machines of type k: W,,=the averageSPHOSOHHSHSHSSSHSHHHSHSOHSHHSHSHOHHHSSHSHHSHHEHSHHHSHSHOHHSCOHOLHLO 146 145 Exercises 45 Maximum Quantity for Shipping Day Desired trom Bonus Offered Bidder Point Shipping Point per Barre! Company ae 70.000 barrels ‘010 8 10,000 barrels 008 Company D A 10,000 barrets 020 8 10,000 barrets $0.15 shipping weight ota rollin group ’demanded in period j. The variables are: x,, ~ the number of rolls in group / that should be machined and shipped in period j. —{8.K. Jain, K. L. Scott, E.G. Vasold [1.18] CRUDE OIL SALES. This problem arose in the sale of crude oil produced at. the Naval Petroleum Reserves at the Elk Hills (A) and ‘Buena Vista (B) fields in California, Amaximum of 10,000 barrels per day Is up for sale from ‘each of these fields. Two companies, C and D, bid for the purchase of this crude oil according to the offers tabulated at the top. The bonus here is the amount the company agrees to pay over an announced minimum base price per barrel. The U.S. Government is con strained by a law not to sell more than 19,000 barrels per day from both these fields to- gather to any single purchaser. It is required to determine how much crude oil to sell 10 each of companies C and D from each of the fields A and B, 0 as to maximize the total bonus obtained, subject to these constraints. Formulate this problem as an LP. AIRPLANE FUELING PROBLEM. An airplane ‘has a route to fly that takes it trom city rtor + 1. {ton — 1, and then at the end from city n to city 1.Forr = 1 ton, w, isthe expected weight of the plane in units (including the body plus crew, passengers, and baggage, but excluding fuel, estimated by using the known passenger inputs ‘and departures at the various cities) when in flight trom city r to r+ 1. We also have the following data for each r= 1 ton: c, =the cost of aviation fuel per unit weight at city r, in dollars; k, = the maximum amount of avia- tion fuel that can be purchased in city F in weight units. Let F units be the maximum. weight of fuel that can be loaded into the 18. L Jackson and J. M. Brown (1.171) plane. By FAA regulations the plane must carry a specified amount of y units of reserve {uel, when it takes off rom any city on the route, beyond what it needs to complete the immediate flight segment and landing at the next cily on the route. Denote by x, the amount of fuel in weight units purchased and loaded into the plane at city /,¢ = 1 ton, Let W, denote the total combined weight of the plane, including Iuel, at takeot! fcam city ron the route ton 1 Assume that the amount of fuel (in weight units) needed to take ol from city r and fly to and land in the next city, city ¢ + 1. in the route (define city n +1 =city 1) is aW,, F=1 tan (wnere x, are given positive numbers determined by using the flying distance and time between cities r and +1. ele). Formulate the problem of deter- mining x... .%, optimally so as to min mize the total fuel cost for completing this trip, as an LP. Also, write down the special ‘case of this model when the data aren = 5, y= 25, F = 226. 7 is 350 ow 2 25 5 300 on 3 or @ 310 oor 4 08 @ 00 oa 5 35 35 25 0.15 2. Forr = 11ton, lett, be the expected duration fof the fight rom city rto city r + 1. In (1) we ‘gave simple linear formulas for determining the fuel consumption in the various tlight ‘segments in the route. If you think that this only gives a crude approximation, consider the following. Suppose that the fuel consumption is to be determined by a more accurate procedure described below. When46 1. Formulation of Linear Programs the plane takes off and keeps tying, fuel will be burnt and the weight of the plane continuously decreases by the amount of spent fuel. At any point of time t, when the plane is in the air, let V(Q) denote the combined weight of the plane, including the fuel at that time point. Suppose the rate at which fuel is burnt at that time point is AV(t) dt, where Lis a given positive constant. Use this to determine the amount of fel spent by the plane to fly the various flight segments. in the route, With this, construct a model for determining the optimum values of Xi. %qy $0 a5 to minimize the total fuel cost for completing this trip. 3. An airline operates a large number of routes with a fleet of airplanes. Discuss how the model discussed in (1) can be extended to determine which plane should buy how much fuel in which city and when, so as to minimize the total fue! bill for the airline, —(M. Queyranne [1.25}) 1.46 A RESOURCE ALLOCATION PROBLEM. The federal government has agreed to provide funds tor a project to transtorm some untrained and unemployable people into occupationally trained employable individuals through a process of training, a participant completes the training period for a particular program, there is a positive benefit as a result of the cost incurred in the training. If the participant does not complete the training, there is no benefit, but a fractional cost of training is incurred ‘anyway, depending on how much of the period he or she remained in taining, The problem is to allocate the funds and participants among the various programs to maximize the total expected benefit, subject to the budget and other constraints. Available data is tabulated at the bottom of this page (some of the values ace estimated from Input information for a community of about 700,000 people consisting primarily of disadvantaged persons, of whom 8 maximum of 500 are expected to apply for 8 training program of 12 weeks duration). The Description Training Program = cost of program In dollars per participant accepted to take it the increased annual income in dollars that @ participant can expect to receive. it he or she completes the program successtully 46, =the probability that an applicant will quality to be accepted as a participant in the program the probability that a participant will complete the program successtully ihe or she begins it 4, =the probability that a client who completes the program will succeed in obtaining a job {oiving him or her the expected increased! lannval income) fractional cost incurred per participant beginning inthe program but not completing it successfully ‘maximum number of applicants that can be ‘screened by the program cision variable, the number of applicants sent to the ith program = decision variables, dolars of budget allocated to the ith program Wk. jet 2 3 4 Work On-the-Job Classroom Public Experience Training ‘Training _Service Job ‘960 2400 2680 ‘3600 8920 10,400, 10.400 7800 08 os or os oa o72 033 ose oss o7 oa oa 0.25 os oar 0.26 180 315 20 50 ” n ” ns Bandyopadhyay, .K. Sungupta, and R. K. Bogh (1.3}) SOHOSHOHSHOHOOHOHSSHOHSHOHSNSHHHSHHHHSHEHOHHOHHHHOHHSHOHOHHOHOHEOCOEHEOHE147 total budget available is $500,000, The ex: pected benefit is the expected increased an- fual income earned by successful partici. pants, Ignoring the integer requirements on the x,, formulate the problem of determining the x, y to maximize the expectes total benelit as an LP. ‘Transform the following mathematical program into an LP. Justity clearly that the solution ob- tained from the transformed LP solves this problem. In this problem, if there is an ad tional constraint |x, ~ 2x, + 3x, ~ 30) ‘could you still ranstorm the whole problem into a single LP? Why? Discuss briefly an approach for solving the problem with this additional ‘constraint Minimize (x) = 6x, + 5x — 8x5 xy + 2Q Ty St ox, — 5x, — 20) S 20 forall Subject to REFERENCES On Applied Linear Programming 14 12 13 14 1s 16 'S. Advani, “A Linear Programming Approach to Air Cleaner Design," Operations Research (March/April 1974), 295-297. L.A. Amold and D. Botkin, “Portfolios to Satisfy Damage Judgements: A Linear Pro- ‘gramming Approach," Interfaces 8, 2 (Feb- ruary 1978), 38-42, 4J.K. Bandopadhyay, S. K. Sengupta, and R. K, Bogh, "A Resource Allocation Model for an Employability Planning System," Interfaces 10, 5 (October 1980), 90-94 F.M, Bass and RT. Lonsdale, ““An Explora~ tion of Linear Programming in Media Selec- tion,” Journal of Marketing Research 3 (May 1966), 173-188, F.G.Bland, “The Allocation of Resources by Linear Programming,” Scientific American 244, 6 (June 1981), 126-144. J Byrd, Jr. and LT. Moore, “The Application fof a Product Mix Linear Programming Modet a7 18 19 1.10 116 Reterences 47. in Corporate Policy Making,” Management Science 24, 13 (September 1978), 1842-1960. A, Chaenes, C. Colantoni, W. W. Cooper, and K. O. Kortanek, “Economic Social and Enter- prise Accounting and Mathematical Models, The Accounting Review 47, (January 1972), 85-108, G. B, Dantzig, “Reminiscences About the Ori- gins of Linear Programming,” Operations Re- search Lottors 1, 2 (April 1982), 43-48. A. Dutton, G. Hinman, and C. 8. Millham, “The Optimal Location of Nuclear Power Fa- cilities in the Pacific Northwest,” Operations Research (May/June 1974), 478-487. 4.F. Engel and A. Warshaw, “Allocating Ad- vertising Dollars by Linear Programming, Journal of Advertising Research 4, 3 (Sep- tember 1964), 42-48. A. Geottrion, “The purpose of Mathematical Programming is Insight, not Numbers." Inter- faces 7,1, 1976, pp. 81-92 ©. A. Glassey and V. K. Gupta, “A Linear Pro- ‘gramming Analysis of Paper Recycling,” ‘Management Science (December 1974), 392~ 408, J.J. Glan, “A Mathematical Programming Approach to Bee! Feedlot Optimization,” ‘Management Science 26, § (May 1980), 524~ 535, P. Gray and C. Cullinan-James, “Applied Op- timization—A Survey," Interfaces 6, 3 (May 1976), 24-29 RLV. Hartley, “Linear Programming: Some Implications for Management Accounting,” ‘Management Accounting 51 (November 1968), 48-51 A.J.Hotiman and W.W. Jacobs, “Smooth Patterns of Production,” Management Sci- ence 1, 1954, pp. 86-91 B. L. Jackson and J. M. Brown, "Using LP for Crude Oil Sales at Elk Mills: A Case Study.” Interfaces 10, 3 (June 1980), 65~70. S.K. Jain, K.L Scott. and E.G. Vasold, “Orderbook Balancing Using a Combination of Linear Programming and Heuristic Tech- niques,” Interfaces 9, 1 (November 1978), 55-6748 1.19 1.20 124 122 1.23 1.24 1.26 127 1.28 1.90 1 Formylation of Linear Programs W.J. Kennedy and J.€. Gentle, Statistical Computing, Dekker, New York, 1980. A.A. Klimpel, “Operations Research in the Chemical industry, Parts { and Il," Chemical Engineering 80, 9 and 10, 1973, pp. 103-108 and 87-94. D.S. Kochhar and G. Allora-Abbondi, “Chan- nel Allocation on a Cable Distribution Sys- tem,” Computers and Industrial Engineering 4, 1980, pp. 173-184 D. B. Kotak, “Application of Linear Program- ming to Plywood Manutacturer,” Interfaces 7, 41 (November 1976), 56-68. A. F. Love, “An Application of a Facilities Lo- cation Model in a Prestressed Concrete In. dustry,” Interfaces 6, 4 (August 1976), 45-49, G.Manula and ¥. Kim, “A Linear Program- ming Simulator for Coal Distribution Prob- lems,” Proceedings of the Symposium in Operations Research in the Mineral Indus- tries, University Park, P., 1966. W. Marcuse, L. Bodin, €. Cherniavsky, and S. ‘Yasuko, “A Dynamic Time Dependent Model for the Analysis of Alternative Eneray Poli cies,” In Operational Research, North- Holland, Amsterdam, The Netheriands 1975. J.J.Moder and S.E.Elmaghraby (Eds), Handbook of Operations Research Models and Applications, Van Nostrand Reinhold, New York, 1978. E.F. Peter Newson, (Ed.), Management Sci- ence and the Manager, A Case Book, Prentice-Hall, Englewood Cliffs, N.J., 1980. A. H. Philipson and A. Ravindran, “Applica- tions of Mathematical Programming to Metal Cutting,” Mathematical Programming Study, 11 (1979), 116-134, M. Queyranne, “The Tankering Problem’ CBA Working Paper series 1982, University of Houston, Houston, Texas. J.R Rive, The Approximation of Functions: Linear Theory, Vol. 1, Addison-Wesley, Reading, Mass., 1964. E.L. Summers, “The Audit Staff Assignment Problem: A Linear Programming Analysis, The Accounting Review 47 (July 1972), 443-453, 1.92 133 1.34 1.95 1.38 On Mul 1.37 1.39 1.40 141 142 1.43 1.44 1.45 V.A. Ubhaya, “Isotone Optimization, I" Jour: nal of Approximation Theory 12, 2 (October 1974), 146-159, VLA. Ubhaya, “Isotone Optimization, 1, Journal of Approximation Theory 12, 4 (December 1974), 315-331 H.M. Weingartner Mathematical Program: ‘ming and the Analysis of Capital Budgeting Problems, Prentice-Hall Englewood Cit, NJ, 1963 T.K. Zierer, A.W. Mitchell, and T. R. White, “Practical Applications of Linear Program- ming to Shell's Distribution Problems," Inter- faces 6, 4 (August 1976), 13-26 Communications in Statistics, Part 8, (Special Issue on Computations for Least Absolute Values Estimation) 86, 4, 1977 jective Programming J.L. Arthur and A. Ravindran, “PAGP, A Partitioning Algorithm for (Linear) Goal Pro- ‘gramming Problems," ACM Transactions on Mathematical Software 6, 9 (September 1980), 978-986. A.Charnes and W.W. Cooper, “Goal Pro- gramming and Multiple Objective Optimiza- tions, Part |," European Journal of Operations Research 1, 1, 1977, pp. 39-54. J. L. Cochrane and M. Zeleny (Eds.), Multiple Criteria Decision Making, University of South Carolina Press, Columbia, S. C., 1973, JL. Cohon, Multiobjective Programming and Planning, Academic, New York, 1978. G. Fende! and T. Gal (Eds.), Multipte Criteria Decision Making, Theory and Application, Springer, Berlin, 1980. C.L Hwang and A. S. Masud, Multiple Objec: tive Decision Making Methods and Applica tions: A State of the Art Survey, Springer, New York, 1979 J.S.H.Kornbluth, "A Survey of Goal Pro gramming,” Omega 1, 2, 1973, pp. 193-205. S.M. Lee, Goal Programming for Decision Analysis, Auerbach, Philadelphia, P., 1972 J. Sponk, Interactive Multiple Goal Program- ming: Applications to Financial Management, Martinus Nijhott, Boston, Mass., 1981SOOHHSSSSHSSHSSHSHSSOCHSHHSHOHOHOHSHOSHSHSHHHOHHOHOOHOHHOSOHOOOCEOCE 148M Starr and M. Zeleny, (Eds.), Multiple Cri- teria Decision Making, North-Holland/TIMS Slugies in Management Science, vol. 6, Amsterdam, The Netherlands, 1977. 1.47 8.Zionts and J. Wallenius, “An Interactive Programming Method for Solving the Multiple Criteria Problems,” Management Science 22, 6, 1976, pp. 652-663, 1.43. Zionts (Ed,), Proceedings of the Conter- ence on Multiple Criteria Problem Solving (Butfalo, NY. 1977), Springer Verlag, New York, 1978. Selected Books on Linear Programming and Related Areas. 1.49 JR. Aronofsky, J.M. Dutton, and M. T. Tayyabkhan, Managerial Planning with Lin- 1° Programming in Process Industry Opera- tions, Wiley-lnterscience, New York, 1978. 1.80 T.S.Arthanari and Y.Dodge Mathematical Programming in Statistics, Wiley, New York, 1981 4.51. M.L. Balinski (Ed), Mathematical Program- ming Study 1, Pivoting and Extensions: tn Honor of A. W. Tucker, North-Holland, Amster- dam, The Netherlands, 1974. 1.52 M.L,Balinski_ and €,Hellerman (Eds), Mathematical Programming Study 4, Com- putational Practice in Mathematical Pro- gramming, North-Holland, Amsterdam, The Netherlands, 1976. 1.583. M. Bazaraa and J. J. Jarvis, Linear Program ming and Network Flows, Wiley, New York, 1977, 154 E M.L.Beale, Mathematical Programming, Isaac Pitman, London, 1968, 1.55. S.P. Bradley, A.C. Hax, and T. L. Magnanti Applied Mathematical Programming, Addi- son-Wesley, Reading, Mass. 1977. 1.96 A. Charnes and W.W. Cooper. Management Models and Industrial Applications of Linear Programming, Vols. | and Il, Wiley, New York, 1961 1.87 A. Charnes, W. W. Cooper, and A. Henderson, An Introduction to Linear Programming, Wiley, New York, 1953, 1.88 N. Christofides, Graph Theory, An Algo rithmic Approach, Academic, New York, 1975. 159 1.60 161 162 1.63 1.64 1.65 1.68 167 168) 170 amt 172 173 174 Roterences 49 V. Chvatal, Linear Programming, W. H. Free- man & Co., San Francisco, California, 1988. L. Cooper and D. Steinberg, Introduction 10 Methods of Optimization, Saunders, Philadel- phia, 1970 G.B. Dantzig, Linear Programming and Ex tensions, Princeton University Press, Prince: ton, NJ., 1963, G.B Dantzig, MAH. Dempster, and M Kallio (Eds), Large-Scale Linear Program- ming (Proceedings of an IIASA Workshop, CP.81-S1-2 Volumes, June 2-6, 1980), The International Institute tor Applied Systems Analysis, Laxenburg, Austria, Dorfman, PA, Samuelson, and RM. Solow, Linear Programming and Economic Analysis, McGraw-Hill, New York, 1958. N. J. Driebeck, Applied Linear Programming, Addison-Wesley, Reading, Mass., 1969 S.E.Elmaghraby, Some Network Models i Management Science, Springer, New York, 1970. G.D.Eppen and FJ. Gould, Quantitative Concepts for Management: Decision Making Without Algorithms, Prentice-Hall, Engle- wood Ciitts, Nu., 1978, S.Even, Graph Algorithms, Computer Sci- ence Press, Potomac, Md., 1978. L.A. Ford, Je, and D. A. Fulkerson, Flows in Networks, Princeton University Press, Princeton NJ, 1962. 0. Gale, Theory of Linear Economic Models, ‘McGraw-Hill, New York, 1960, M.R. Garey and D.S. Johnson, Computers and Intractability: A Guide to the Theory of NP-Completeness, Freeman, San Francisco, 1979 R. S. Garfinkel and G. C. Nemhauser, Integer Programming, Wiley-lnterscience, New York, 1972 W.W. Garvin, Introduction to Linear Program ming, McGraw-Hill, New York, 1960, S|. Gass, Linear Programming: Methods and Applications, 4th ed., McGraw-Hill, New York, 1975. D.P. Gaver and G.L. Thompson, Program- ‘ming and Probability Models in Operations50 178 1 17a 179 1.80 1.81 182, 1.83 1.84 1.85 1.86 1.98 1.89 1.90 1 Formulation of Linear Programs Research, Brookes/Cole Publishing Co. Monterey, Calif, 1973 R.L.Graves and P.Wolle (Eds), Recent Advances in Mathematical Programming, McGraw-Hill, New York, 1963 M.R. Greenberg, Applied Linear Program: ‘ming for the Socioeconomic and Environmen- tal Sciences, Academic, New York, 1978. G. Hadley, Linear Programming, Addison: Wesley, Reading, Mass., 1962, F.S, Hillier andG. J. Lieberman, Introduction to Operations Research, 3rd ed., Holden-Day, San Francisco, 1980. T.C. Hu, Integer Programming and Network Flows, Addison-Wesley, Reading, Mass., 1970, P. A. Jensen and J. W. Barnes, Network Flow Programming, Wiley, New York, 1980, J.F. Kennington and R.V.Helgason, Algo- rithms for Network Programming, Wiley-Inter- ‘science, New York, 1980, T.C. Koopmans (Ed.), Activity Analysis of Production and Allocation, Wiley, New York, 19859, H.W. Kuh and A.W, Tucker (Eds), Linear Inequalities and Related Systems, Princeton University Press, Princeton, N.J., 1956, L.S. Lasdon, Optimization Theory for Large ‘Systems, Macmillan, New York, 1970 E.L.Lawler, Combinatorial Optimization— Networks and Matroids, Holt, Rinehart and Winston, New York, 1976, RL. Levin and R. P. Lamone, Linear Program- ming tor Management Decisions, Irwin, Homewood, Ill, 1969 A.W. Llewellyn, Linear Programming, Holt Rinehart and Winston, New York, 1964 D.G. Luenberger, Introduction to Linear and Non-Linear Programming, Addison-Wesley, Reading, Mass., 1973. ALE. Machol, Elementary Systems Mathe- matics: Linear Programming for Business and the Social Sciences, McGraw-Hill, New York, 1976. E.Minioka, Optimization Algorithms for Net works and Graphs, Dekker, New York, 1978 191 1.92 1.93 194 1.95 1.96 197 1.98 1.99 1.100 4.101 1.102 4.103 1.106 1.107 4.108 B.A. Murtagh, Advanced Linear Program- ‘ming, McGraw-Hill, New York, 1981 K.G.Murty, Linear and Combinatorial Pro- gramming, Wiley, New York, 1976. K.G. Murty, Linear Complementarity, to be published K. G. Murty, Network and Combinatorial Pro- gramming, to be published. 8. Noble, Applied Linear Algebra, Prentice: Hall, Englewood Clits, N.J., 1969 W. Orchard-Hays, Advanced Linear Program- ‘ming Computing Techniques, McGraw-Hill, New York, 1968. D.P. Phillips, A. Ravindran, and J. Solberg, Operations Research: Principles and Prac- tice, Wiley, New York, 1876. P.H.Randoiph and H.D. Meeks, Applied Linear Optimization, Grid, Columbus, Ohio, 1978 8.8, Rao, Optimization Theory and Applica- tions, Wiley Eastern Ltd, New Delhi, 1978. AI. Rothberg, Linear Programming, North Holland, New York, 1979. T.L Saaty, Mathematical Methods of Op- erations Research, McGraw-Hill, New York, 1959. M. Sakarovitch, Notes on Linear Program- ming, Van Nostrand Reinhold, New York, 1971 H.M.Salkin and J, Saha, Studies in Linear Programming, North-Holland, Amsterdam, The Netherlands, 1975, H. M, Salkin, Integer Programming, Addison Wesley, Reading, Mass., 1975, DM. Simmons, Linear Programming for Operations Research, Holden-Day, San Fran- cisco, 1972, M, Simonnard, Linear Programming (Trans~ lated from French by W. S. Jewell), Prentice- Hall, Englewood Clits, NJ. 1966, W. A. Spivey and A, M. Thrall, Linear Optimi- zation, Holt, Rinehart and Winston, New York, 1970. JE. Strum, Introduction to Linear Program- ‘ming, Holden-Day, San Francisco, 19721.108 1110 aati a2 4.113 H.A.Taha, Integer Programming: Theory, Applications, and Computations, Academic Press, New York, 1975. . Vajda, Mathematical Programming, Addi- son-Wesley, Reading, Mass., 1961. ©. Van de Panne, Linear Programming and Related Techniques, North-Holland, Amster ‘dam, The Netherlands, 1971 C.Van de Panne, Methods for Linear and Quadratic Programming, North-Holland, Am ‘sterdam, The Netherlands, 1975, H.M. Wagner, Principles of Operations Re- search with Applications to Managerial Deci- sions, Prentice-Hall, Englewood Clitts, NJ, 2nd ed., 1975. 1314 1115 1.116 war 1.118 Reterences 51 H.P. Williams, Mode! Building in Mathemati- cal Programming, Wiley-lnterscience, New York, 1978. N. Wu and A. Coppins, Linear Programming and Extensions, McGraw-Hill, New York 1981 M.Zeleny, Linear Multiobjective Program: ming, Springer, New York, 1974 S.Zionts, Linear and Integer Programming, Prentice-Hall, Englewood Cliffs, NJ, 1974 S.1. Zukhovitskiy and L.1. Avdeyeva, Linear and Convex Programming, Saunders, Phila- delphia, 1966.“ee 2.1 INTRODUCTION In an LP we have to optimize the objective function Subject to linear equality and/or inequality con- straints, and sign restrictions on the variables. If there are no sign restrictions on the variables and ‘no inequality constraints, the LP will be of the form Minimize 2(x) = ox Subjectto Ax =b where A,b. c,are given matrices or vectors. Itcan be shown (see Exercise 3,17) that in this case, if there is a feasible solution, z(x) is either a constant on the set of feasible solutions, or it can be made to diverge to —a on this set. Hence, such LPs are trivial, and not of practical interest. Hereatter, we will assume that the LPs being discussed consist of at least one inequality constraint or a variable restricted in sign {As a naive approach for solving such problems it may be tempting to use the familiar technique in elementary calculus of trying to solve the system of ‘equations obtained by putting the partial derivatives of the objective function equal to zero. This technique does not work hare because of the constraints. Even the standard Lagrange multiplier technique does not apply directly to this problem because we have to deal with inequality constraints, and the Lagrange multiplier technique only handles equality con- straints, We take recourse to the Karusn—John~ Kuhn-Tucker theories to handle inequality con: straints. Unfortunately these theories provide only certain necessary conditions that every optimum solution must satisfy, and they do not provide any systematic method for solving these conditions. in any case, itis clear that any method that can find an optimum solution of an LP must contain in itself a method for finding a feasible solution satisfying a system of linear equality and inequality constraints, can be shown that this problem of finding an initial 52 The Simplex Method feasible solution can itself be posed as another LP. Solving such a linear programming formulation ‘seems to be the only practicable approach for finding a feasible solution, satisfying all the constraints in the problem. Thus, none of the classical methods in calculus or linear algebra offer any approaches for solving LPs or for solving systems of linear inequal- ities. Special techniques had to be developed to tackle these problems. This chapter discusses one of these special techniques, known as the simplex method due to George B. Dantrig (1947) Recently, a special direct method for solving sys- tems of linear inequalities (which does not need the transformation of this problem into an LP) known as the ellipsoid method has been developed. The ellip soid method belongs to a class of algorithms known 48 variable metric methods in nonlinear program- ‘ming literature. By applying the ellipsoid method to solve the system of necessary and sufficient op! ‘mality conditions corresponding to an LP, we can solve the LP. The ellipsoid method is polynomially bounded (in the sense that the worst case computa- tional effort needed to solve an LP or a system of linear inequalities with integer data, is bounded above by a polynomial in the size of the problem), The simplex method is a pivotal method (based on carrying out a sequence of pivot operations on the matrix of coefficients of the system of constraints) {nd it is not polynomially bounded in the worst case. However, for solving practical LP models, the sim- plex method seems to be much more efficient than the ellipsoid method, In this chapter we discuss the simplest form of the simplex method to establish rotation and develop ideas. Other forms of the sim- plex method are discussed in later chapters (see, for example, Chapters 5, 6, 7, 10, and 11). The ellipsoid method is discussed in detail in Chapter 15, Mtwould be convenient ifthe final answer to a prob- lem can be obtained either as a simple formula, in2.2. Transforming the Problem into the Standard Form 53 terms of the data, or in some tabular form from which the answers can be looked up easily. It will be shown in Chapter 3 that this is generally impossible for LPs LPs can only be solved by means of an algorithm, which is a step-by-step procedure that leads to an answer to the problem in a finite number of steps. 2.2. TRANSFORMING THE PROBLEM INTO. THE STANDARD FORM Betore applying the simplex method on an LP, all the ‘constraints in iton which pivot operations are carried ‘out must be transformed into equality constraints. ‘The reasons for this are described below. Also one may want to either maximize or minimize the objec- tive function. It becomes very cumbersome if we have to discuss versions of the simplex method for ‘each of these different types of LP models sepa- rately. Instead, we discuss the simplex method for solving LPs in a particular form known as the stan- dard form, and show here how every type of LP ‘model can be transformed into this standard form, In an LP model, sign restrictions on a variable ‘would require that the variable be either nonnegative (or nonpositive, as specified. We refer to such condi- tions as restrictions in the LP model. In the simplox ‘method, pivot operations are not carried out on these restrictions. They are held separately and the method tries to satisty these restrictions implicitly. All the other conditions in an LP model (whether it involves one or many variables) will be referred to in this chapter as constraints. Constraints are all the conditions on the decision variables in the model which are included in the pivot operations when the ‘simplex method is used to solve the problem. For the version of the simplex method that we discuss here, all conditions other than sign restrictions on individ ual variables, are called constraints. In Chapter 11, we discuss a special efficient imple- mentation of the simplex method known as the bounded-variable simplex method, ia which both ‘upper bound and lower bound conditions on indi- vidual variables are treated as restrictions, and con- straints refer to only those conditions that involve two for more decision variables. In that method, pivot operations are only carried out on the system of constraints in the model, each of which involves two for more decision variables To transform an LP into the standard form, carry out the following steps in the order listed Transtorm Lower Bound Constraints into Nonnegativiy Restrictions I there is any variable in the problem that has a required lower bound other than zero, for example, x, 24,, where Z, is some specified number. elimi- nate this variable from the model by substituting x, = ys +4; wherever it oceurs in the model. Clearly the constraint xj 2 ¢, is equivalent to requiring xy =y.+%s, 20. In this manner all lower bounded variables in the problem can be replaced by variables with lower bounds of zero. The lower bound constraint on x, gets replaced by the nonneg- tivity restrition on the variable y, Transforming all the Constraints into Equality Constraints Convert all the remaining inequality constraints (including upper bound constraints on variables, if any) into equations by introducing appropriate slack variables, A constraint of the form, © aie, 6, becomes Lats t tera On Meer BO and J, a,x, 2 6, becomes ¥ asx) ~ Hore, x.4. %ys2 afe the slack variables, The coel- ficient of the slack variable in the transformed ‘equality constraint depends on whether the original inequality constraint was a < or a > inequality. Slack variables are always nonnegative, and represent tunutilized capacities, etc. For example, in the above constraints, the slack variables are X,,, = 2; ~ aX, aNd X43 = SF; aX, —Dy. IN some Textbooks, the name slack variable is only used for variables ‘such as x,., above; and the variables such as x,,; are called surplus variables. Here we use the term slack variable for both these kinds of variables, The slack variables are as much @ part of the original problem as the variables used in mod- fling the constraints. Slack variables remain in the problem throughout, and their values in an optimal54 2 The Simplex Method feasible solution of the problem give significant information Note 2.1: Each inequality constraint in the origi- nal problem leads to a different slack variable. Nonnegative Variables In the standard form we require all the variables to be nonnegative. It the original problem contains some unrestricted variables, they can be handled by two methods. In one method each unrestricted variable is transformed into the difference of two nonnegative variables, as in Section 1.2.4, For ex- ample, let the variable x, in the problem be unrestricted in sign. Any real variable can be ex- pressed as a difference of two nonnegative vari- ables. So we can substitute x, = x! — x;, where Xf 20 and xz 20. It x7 — x; is substituted for x, wherever it occurs in the original LP and the new variables xj and x; are restricted to be nonnega- tive, the resulting LP is equivalent to the original one. Each unrestricted variable can be handled in a similar manner. This method handles unrestricted variables conveniently, but it is net used much in practice because it increases the number of vari- ables in the problem, Also if this method is used, the resulting LP will invariably be unstable (see Section 4.8) The other method for handling unrestricted vari- ables takes advantage of their unrestricted nature and eliminates them from the problem, Suppose x, is an unrestricted variable in an LP. Since x, is @ variable appearing in the problem, the coefficient of x, in at least one of the constraints must be nonzero, Suppose it is the ith constraint, which is ax; + aX) $91 +9QX,=b,, where a4 #0, This con: straint is equivalent tox, = (b,— aise; — ,X,}/@,.. Since there are no further restrictions on the variable x,, once the optimal values of the other variables x,,...,x, are obtained, the associated ‘optimum value of x, can be obtained by substituting the values of x, to x, in the above expression. Eliminate x, from the problem by storing this ex: pression for x, somewhere, by substituting this ex- pression for x, in all the other constraints and the objective function, and by eliminating the ith con- straint from further consideration, All the unrestricted variables in the problem ca be eliminated from the problem ia this manner one by one. Each unrestricted variable eliminated in this way reduces the number of variables in the LP by one and reduces the number of constraints by one; this makes the order of the problem smaller and hence easier to solve Exercises 2.1 Explain why variables that must be nonnega tive cannot be eliminated from the problem in the above manner, It there is @ variable in the model that must be nonpositive, say x, <0, let x, = —y, and eliminate x, from the model by using this equation; y, is a nonnegative variable. After these transformations, all the variables in the transformed model are non- negative variables. All restrictions in the model are nonnegativity restrictions on the variables. Minimization Form Eliminate constant terms in the objective function. This does nol change the sel of optimum solutions for te problem. Then express the objective function In the minimization form. Inthe original statement at the problem, i the objective Is maximize 20) = Th.1 ch, change the problem withthe new objec tive a8 minimizo z{x) = - 20) = Siny ey, Tie change does not afect the set of Optimal teasble folulions ofthe problem, and the equation [man mmum valve of z(3}| = ~ {minimum valve ofr], ean be used to got the maximum valye of the original objective function Nonnegative Right-Hand-Side Constants Now the LP has been transformed into the following form Sen Minimize 21x) These are the 110m constraints in the problem These are the sign resirictions on the variables. suectto fayyat, i and x20 Multiply the ith equation on both sides by —1, if b, is negative. As a result of this, all the right hand constants in the equality constraints become nonnegative.COCOHHOOEEOHOOSOOEOOOLOOKCHOOETCOEOHEOOHOOOELEOEEOCE 22. Transtorming ine Problem into tne Standara Form 55 Standard Form The problem is now in standard form. It may be written in detached coefficient tabular form. Tableau 2.1 Original Tableau x noe ° an a ° be ay ay a o b 2a Bg Bay Be ° be a Gr 8 oe 1 ° 442 Or all, minimize 2 Each row in the tableau (except the last row) rep- resents a constraint in the problem. The ith row rep- resents the constraint ayyx, +aaXz4: °° taut, =b, The last row in the tableau represents the equation 6.x, +2 + 6k, — 2 = 0. This row is called the cost row, or the objective row. The ¢, are known as the original cost cootticients; the numbers by,..-, Bq are known as the original right-hand side constants. land the numbers a,, are known as the original input-output coetlicients. The column vector asso- cated with x) in Tableau 21 (ayy... 6)" is known as the original column vector of x, in this problem. The variables in the original tableau Xpee-o%y are all the variables in the model, in- cluding slack variables (excepting possibly some un- restricted variables that may have been eliminated from the problem). A feasible solution is any vector X= (¥...-%) that satisfies all the constraints and the sign restrictions. Figure 2.1. System of inequalities. Why Inequality Constraints Should be Transformed into Equality Constraints Before Applying the Simplex Method In any system of equations we can (1) multiply all the coetlicients on both sides of an equation by any nonzero number and (2) multiply aa equation by 2 reai number and add it to another equation. These transformations do not change the set of feasible solutions of the system. For example, consider the system: 1 3x, — Te xtog 3 ‘Add two times the second equation to the first ‘Then multiply the second equation by four. This leads to the following system, which can be ver fied to have the same solution 5x, — 8x = 25 4x, + oxy = 92 When transtormations of types 1 and 2 are carried out on a system of inequalities, the set of feasible solutions is normally altered, For example, consider the system of inequalities x, 20, x,20 in RP (As discussed above, these are treated as sign restrictions in the simplex method, and not included in the pivot operations. However, we choose this simple system purely to illustrate what can happen when operations of types 1 and 2 are performed on fa system of linear inequalities.) The set of feasible solutions of this system is the nonnegative quadrant in R?, the shaded region in Figure 2.1. Now trans- orm the system by adding two times the second inequality to the first. The new system and its set of feasible solutions are given in Figure 22. Clearly Figure 22. Transformed system,56 2 The Simplex Method the transformed system does not have the same set of feasible solutions as the original system However, when a system of inequality constraints is transformed into a system of equations by intro- ducing the appropriate slack variables, then trans- formations of types 1 and 2 can be carried out on the resulting system of equations, and these will not change the set of feasible solutions of it The simplex method uses transformations of types. 1 and 2 on the system of constraints to find an op- timum solution. Hence it is essential that the system of constraints of an LP on which pivot operations are performed be transformed into a system of equa: tions by introducing appropriate slack variables be- fore applying the simplex method on it. Example 2.1. Transforming a Problem into Standard Form Consider the following LP Maximize See Ft Yet WS Yor Subject 10 Bry + 2k + Va— Vet Vote 2x, He Bet 4 P+ Wet Yo FW Stet = 7 Ye B-5 x <9 GX Z0 Ye SO Yur You Ys unrestricted First the lower bounded variable yg is eliminated by substituting y, = —5 + x,, where x, is a nonnegative variable, in all the constraints and the objective function. Replace the constraint y, > —5 by the onnegativity restriction on x,. The transformed first constraint iS 3x; + 2x; + Xe — s+ Ve +¥r2 —4 + 5-1. Introducing the stack variable s,, this con- straint is equivalent to 3x, + 2X; + Xe Ys + Ye + ¥>~ 5, = 15S, 2 0. From this constraint, we derive Ve = 13x —2xy— ky + Ys — Yr +S). Substitute this expression for y, in the objective function and all the constraints, thereby eliminating y, from the system. These transformations change the fourth constraint into =x, ~ 7x; ~ 4x3 — Xe + ¥s— Yn + 28, = 10. From this we have y, = —10~ x, ~ Tey 4x, — x, +y, +25). Substitute this expression for ¥> in all the remaining constraints and the objective function, thus eliminating y, trom the system. The original constraints 1 and 4, trom which the expres- sions for the unrestricted variables ys, y; are ob- tained, are eliminated from the system Slack variables are now introduced to transform the femaining inequality constraints into equations. Transform the nonpositive variable y, by substituting Ys= Xs, where x, 1s a nonnegative variable. Eliminate constant terms trom the transtormed objective function, and multiply itby ~ 1 totranstorm it into the minimization form. in the resulting system, thefirst constraint is —x, ~ 15x, — 9x, — 24 — x4 + 45, +5; = —31, Multiply both sides of this by —1. to make the right-hand side constant positive, The resulting problem in detached coefficient tableau form is as follows ¥, 2 Ofor all /,5,,5,,5 20, 210 be minimized 2.3 CANONICAL TABLEAUX, 2.3.1 Nonsingular Linear Transformations Each row of the tableau represents a constraint in the problem. From linear algebra it is well known that nonsingular linear transformations, that is trans: formations such as the following, transform the system into an equivalent system, 11 Multiplying al the entries in a row of the tableau by a nonzero constant, which is equivalent to multiplying both sides of the corresponding con- straint by that nonzero constant. 2 Adding a constant multiple of a cow of the tableau to another row, which is equivalent to adding a constant multiple of a constraint to another one. Since they operate on the rows of the tableau, trans- formations of this type are known as elementary row operations. Given the equivalent system obtained after making several elementary row operations, a reverse series of elementary row operations can be performed on the equivalent system to yield the ‘original system again. Neither the set of feasible solutions nor the set of optimum solutions of the S9HSOHOHSHOHSHSLRESHSSOSHSHOHOHSHHHSHOHSKHHOHSHSHEHLOHOSSEHLESHSCLHCESCSESCOSHSOCOOEEEHHOEOEOHOEEOOCEOLESOEOSOOOHOEEOOEEEOOE 23 Canonicat Tableaxe 57 Flow a ay by Flow 2 a ay, ae bs alfiow 1) + Row? Bay) $9 Dy) AH, ob, + by ‘Row 2) Bas, Ba, 8 Bos problem is changed by making such transformations. Hence we take the liberty to pertorm such transtor- ‘mations. The main approach of the simplex algorithm is to move trom one basic feasible solution to another by making such transformations until an optimum solution is obtained Note 2.2; Row operations are always pertormed element by element using corresponding elements in the respective rows. The tableau at the top trates how a row operation is performed using two rows, rows 1 and 2. The third row in this tableau gives (row 1) + row 2, where 2 is any real number The fourth row in this tableau is Brow 2), where B is any nonzero real number. A numerical example is given below. Row 1 + 08 5 3 Flow 2 12 4 1 =sifow p+ Row2 | 5 1-8 Pivot Operation A pivot operation or a pivot step on a system of equality constraints, or on the tableau representing the system in detached coetficient form, is uniquely specified by specifying the pivot element for the operation. The pivot element has to be nonzero. (It cannot be zero, because the pivot operation requires divisions with the pivot element as the denominator } Consider the pivot operation on the system in Tableau 2.2, with the element a,, a the pivot element. The row containing the pivol element 2,, (ie, the rth constraint in Tableau 2.2 here) is called the pivot row, and the column containing the pivot element (j.e., the sth column in Tableau 22 here) is called the pivot column in this pivot operation. ‘The pivot operation consists of a series of elemen: tary row operations on Tableau 2.2, o transform the pivot column into one containing an entry of +1 in Tableau 22 Pivot Operations on a System of Equality Constraints, With the Pivot Element Circle an an B a8 =, Pivot row Pivot column the pivot row, and an entry of 0 in all the other rows, IW requires the following 1 For each / 4 r, subtract a suitable multiple of the pivat row from the ith row so that the entry in the pivot column in the ith row becomes zero, i Divide the pivot row by the pivat element. Since this is a system of equations, these opera- tions fead to an equivalent system of equations. The ‘equivalent system obtained after this pWvot operation is given in Tableau 2:3, where for j=1 to m By) = 8] By = 8) ~ 8,8, Jy OF TAL By mB = aby ori Ar See Example’2.11 below for an illustration of the pivot operation, The simplex method is called a pivot ‘method, since it solves an LP by carrying out sequence of pivat operations on the system of con- straints in the LP Tableau 23 Equivalent System Obtained Atter the Pivot Operation a a Bad aa: ay 0