LIBRARY OF THE MASSACHUSETTS INSTITUTE OF TECHNOLOGY ^]AS3. MULTIPLE CRITERIA PUBLIC INVEST>ENT DECISION MAKING BY MIXED INTEGER PROGPJ^^IMING* Jeremy F. Shapiro WP 788-75 May 1975 *Multiple Criteria Decision Making Proceedings of a Conference Jouy-en-Josas, France May 21-28, 1975 Jointly sponsored by Centre d'Enseignement Superieur des Affaires (C.E.S.A.) and European Institute for Advanced Studies in Management (E.I.A.S.M.) Edited by H. Thiriez and New York: 1976 S. Zionts Springer-Verlag, Berlin-Heidelberg. MULTIPLE CRITERU PL'BLTC INVESTMENT DECISION MAKING BY MIXED INTEGER PROGRAMS, ING-<^ Jeremy 1. F. Shapiro IKTRODVCT ION Multicriterion optinization has long been an integral part of quantitative The n>odels which have received the most attention models for public planning. consist of concave utility functions to be maximized (in a vector sense) subject to The convex structure of such models permits the development convex constraints. of con-.plete theories characterizing efficient solutions, and well-behaved algorithms for finding these solutions. In this paper, we consider a different class of Dublic planning probleras for which the assumption of convex constraints, or non-decreasing marginal costs of production, is not valid. There are at least three circumstances when this is the First, there can be public investment alternatives such as dam or power case. Second, there can be important plant construction involving large fixed costs. returns to scale in construction and production costs. Finally, public investment, decision problems involving new technologies can have associated diffusion and learning curves which are S-shaped; that is, increasing marginal costs are ex- perienced at first, followed by a consolidation phase when marginal costs are decreasing. models. All of these phenomena can be described by mixed integer programming There is a grow^ing literature on decision models of this type; for example, see Bhatia ';i975l, Dorfnan and Jacoby [1972], Goreux and Manne [1973], Kendrick and Stoutesdijk [1975], The Kuhner and Harrington [197i], .Marks [1974], Westphal [1971]. of this paper is the development of procedures for multi- main concern criteria optiniization of inte;;er progracising problems. consider is the vector maximization Max The specific model we will proMem Cx s.t. Ax 5 b I where C is a P x N matrix, A coefficients of A and b are i s (1) an M x N matrix, b is an M x integer. The columns a vector, and the 1 of A correspond to projects extending over time to be selected or rejected by P interested parties on the basfs" of the P x c 1 vector c measuring its value to these parties. may equal the present net value of project P to the pth party. j computed with a The component discount rate specific The constraints in (1) can consist of capital budget constraints for each of a number of time periods, other resource constraints, «!UVI<IJb,llfi.W!.W^ logical constraints imposing unique starting period for a project, precedence constraints on subsets a For future reference, we define the set of projects, and so on. F = (xiAx b. X. '. say F the set is finite, «» or 1); " {x ). ^i (2) • Different and more general mult icriterion optimization problems than (1) in- volving indivisible investment alternatives are clearly possible. One can envision mixed integer programming models involving continuous decision variables as well as fixed charge variables. In party may be measured by a multicri terion IP addition, the value of a feasible solution x to the pth utility function u . Such extensions to our basic model and algorithms for finding efficient solutions to it are possible and will be discussed in the conclusions section of this paper. As with convex vector maximization problems, there probably is not a vector BaxiTnum solution to problem (1), and we are led to a study of efficient solutions. Specifically, a solution x satisfying the constraints of problem (1) is said to be efficient if there does not exist a feasible solution x satisfying the P inequalities. Cx 2 Cx The characterization of efficient solutions vith strict inequality for some component. of IP problem (1) is not as simple as it is for linear or convex multicriterion ' problems, and moreover, the calculation of these solutions is radically different. The plan of this paper is as follows. IP Section 2 discusses mathematical pro- The following section contains a brief review of perties of efficient solutions. duality theory which provides the constructive means for implementing some of the results of section Section 2. 4 contains used to compute efficient solutions. description of how IF duality theory is a Section 5 contains sensitivity analysis for compiiting the convex region in which a given efficient solution is optimal in a related IP problem. Some concluding remarks and areas for future research are discussed briefly in section 2. 6. MATHEMATICAL PROPERTIES OF EFFICIENT SOLUTIONS The usual practice in generating efficient solutions to a multicriterion mathe- matical programming problem is to give positive weights to each of the P objective functions and combine them (for example, see Rjhn and Tucker Cl951J, or Karlin To this end, let 1:1959]. S-lXcR P r T , p-l • =1.\ \ i ^!gifgaP«gaE»^ | W^^ -:;ijagS^S^^f^j! \ P P i i '•- W ^M M!iti!rt^g- ip | "*«' '" ^«ft* Jg ' - J g^^ int = S (aU and S c > a for all p). P For A E IP problem int S, define the v(X) = iCx (3) s.t. X Theorem For I: c X F s any maximal solution to (3) is efficient. int S, Let x denote a maximal Proof: a coaponent. int S, Since > e solution to (3) and suppose it is not efficient; y e F such that Cy t Cx «ith strict inequaT,ity in some that is, there exists this clearly icolies (Cy) X ) P p-1 > P T p=l (Cx) X P P contradicting the optimality of x in (3).|| As is well known, the difficulty with the characterization of Theorem efficient solutions is that it The following example, due to Gabriel Bitran, illustrates solution to be efficient. The this point. IP of 1 a sufficient but not necessary condition for a is problem is 1,10,9 1 10,1,9 (4) 1-3 "2 ' 1 X. - '^3 or 1, j - 1,2,3 J There are seven feasible solutions to this problem, and three efficient = solutions X (1,1,0), x~ = (1,0,1), x (0,1,1). Let X be the weight on the top objective function in (4), and 1-X be the weight the bottom objective function. It is X easy to verify that x e[0,l/2], and x 3 is optimal in the IP is optioal for X t problem (3) derived [1/2,1], but x 1 is not optimal froni for (4) for any X e S. rnis anomaly of efficient IP solutions needs further study and will not be Instead, we focus on parametric reconciled here. Given any x e v(X) = \Cx F, for XCx comoutatlon of v(X) for x g int S. is a oiecewisc linear convex function on S. It can easily be shown that v()) X satisfying 1, P (5) p-1 -ll- l J-W Jgi^^W,» l'l/ Wijy^U JKJW^' 'M i i - '>^- t'' -'-'»'^t«W!,^^ ! m . s(x ) dennte the region of Theorem 1 if Six'') Let Cnc S r. ' S defined in (5); clearly x e int S for mul ticri tericn LP is possible by is efficient by . !< Paranetric variation of > sensitivity analysis which relies in turn on LP duality theory. The IP problem (3) can be expressed as the LP problem V (\) =» max >Cx s.t. X (6> where "[ 1" [F] e This last statement is true because all extreme denotes convex hull. points of [F] are solutions in F. fact, the zero-one structure of problem (1) In implies all solutions in F are extreme points of CF]; problem (3) to a that is, x is an extreme The difficulty with this apparent reduction of point of [F]if and only if x eF. more manageable optimization problem is that [F] is generally impossible to characterize in any real practical sense. The IP duality theory reviewed in the next section, and used in subsequent sections, can be interpreted as approximating [F] in a neighborhood of an optimal IP solution. provides insights on how to perform sensitify analysis as problem 3. As a result it varies in int in S (3) REVIEW OF In X IP DUALITY TtiEORY this section we reciew briefly the duality theory developed in Shapiro TP [1971], Fisher and Shapiro [1974], Bell !:i973a3[1973bl We consider the zero-one , Bell and Shapiro [1975]. problem ff max ex s.t Ax + . = b X. = or 1 s. »= 0.1,2, Is (7) ..., U. which is identical to problem (1) except bound en the value of the slack s.. c is 1 The integer U. is an upper x N. The slacks in (7) are integer variables because A and b are integer. A dual problem to (7) is constructed by reformulating it as follows. Let G be any finite abelian group with the representation «Z G = Z q^ e ... eZ q, "^r where the positive integers q. satisfy q.? the cyclic group of order q.. Let denote the order of G; clearly |g| - r n i"l ^ q.. I Let e 1- ^n ^"^ ]gJ 2, q. q. ,, i"l, .... r-1, and Z is ^"Y elements of this group and for any M-vector iww 'WJiwi'PBBiBiPtBMp'Wtgww^ ^^ 'ii ^ 174 f, define the element $(t) " o £ G by H (f) - y f,.e.. i=l It can easily be shown that (7) is equivalent to (has the same feasible region as) max ex (8a) (8b) (8c) It can easily be shown that surh a pair is optimal in the rtspective primal For a piven IP dual problen, there is no (Euarantee that and dual problems- the optijiality conditions can be established, but attention can be restricted to optimal dual solutions tor which we try to find a complementary optimal primal solution. V < IP problen cannot -be used the dual If V and we say there is Solution of the IP a IF dual problems analogous to The group G = 2^, = {(x,z)|x. Y the primal problem, problem (7) by dual methods is constructively achieved by generating a finite sequence of groups {G and to solve duality gap. dual problem can be shown to be ) , „, sets fY } . (9) with minimal objective function value w = or tlie groups here have the property that G k I , s = . . 0,1,2,...U.) and the corresponding linear prograiraning relaxation of (7). The k+1 implying directly that , is a subgroup of G ^k+1 supergroup of G . The critical step in this approach to solving the IP problem (7) is that if an optimal solution to the kth dual does not yield an optimal integer solution, then k+1 k+1 k we are able to construct the supergroup G c Y so that Y Moreover, the . construction eliminates the infeasible combination by the IP IP solutions £ which are used in Y dual problen to produce a fractional solution to the optimality conditions. Since the set of feasible solutions to (7) is finite, the process must converge in a finite number of an (x,s) IP dual problem constructions to ipdual problem yielding an optimal solution to (7) by the optimality conditions. The IP dual problem (9) is actually a large scale linear programming problem, Let Y = {x t T t s , w be an eniraieration of Y. ] = min The LP formulation of (9) is \> V > ub + (c-uA)x t - us = 1. (10) ..., T The linear programing dual to (10) is V = max y (ex )!i. (Ax + t=l s.t. I t=l b " b )u3 (11) ^ T as* ai)iP!W J«w iiti <» «!T! i ! ( ft i fi4* ii i mmmt The nunb«r of rows T in (10), or columns in (11), is enonaous. The solution methods given in Fisher and Shapiro ri974], generate colunms as needed (11) as a primal dual pair. algorithras for solving (10) and by descent Thu columns are generated by solving the La^raneean problem which can be solved in a matter of a few seconds for values of |g| see Glover [1969] or Gorry, Northup up to 3000; and Shapiro [1973]. The formulation (11) has a convex analysis interpretation. Specifically, the feasible region in (11) corresponds to = b, {(x,s)|Ax+Is s X. £ 1, ? s. 1 U.} [Y] n where the left hand set is the feasible region of the LP relaxation of (7). Thus, in effect, the dual approach approximates the convex hull of the set of feasible integer points [F] by the intersection of the LP feasible region with the polyhedron When the [Y]. IP dual problem (9) solves the IP Problem (7), then [Y]has cut away enough of the LP feasible region to approximate the convex hull of feasible integer solutions in a neighborhood of an optimal solution. IP CALoruaios of efficent solutions 4. f Solution methods for the mul ticri terion of A c of efficient solutions as IP IP problem \ for (3) F int S are basically methods for solving arbitrary a fixed value problems. The generat varies in int S, however, facilitates the use of the dual methods outlined solution with this value. s. t M . Ax + X. - Is = b or 1; (12) s. •= 0,1, ... ,U. q=l,....Q v(X ) is an initial lower bound on the value of an optimal solution to (12). As discussed in the previous section, a dual problem to (12) is constructed from abelian group G inducing a set Y which includes all feasible solutions to (12), M problem but which eliminates many infeasible solutions. For u t R , the Lagrangean an — " ..„ . ...u. . UN..,. " I. .. "i" '»i i Jll i «^" *^^><Wm***'MW** 177 (13) (x.sUY Y = {(x,s)| and the TP a.x. T e.x. - " e.s. y t- or 1, s. - 0, .... U.}. 1, dual problem is w(X = mill ) L(ulx s.c. u ) R e (lA) . For future reference, the maximization in (13) is written as ub + G(eiu,X ); the algorit'ons given by Shapiro [1968], Glover [1969], Gorry, Northup and Shapiro [1973], compute G(o|u,\ for all a ) e G as a natural by-product of computing G(g|u,X ). For any \c S, ve define the set J(X) = (j|x. = The IP ontiaal diial methods at = \ in ,0 1 (12) without loss of optimality} solution is found bv the ontimalitv conditions. j ) until an Alternativelv. the IP anv ooint and branch and bound nethodi can be dual methods can be abandoned at oerforued usine onlv orojects (15) can be viewed as building uo the set J(X c for c j J fX ). Without loss of ootiraalitv. the oroiect k can be eliirinated from Theoren M consideration in (12) (i.e. k c J(X )) if. for any u c R , 2: X^c*" * u(b-a"'') Proof: Let v(X t x, + G(3 - a^|u,X°) ^ denote the value of (12) = 1) N v(\°ly^ = = x"c'' + max 1) ^ j-1 N * Is « b - a j=l or 1, 0,1 j jl k U ' ''«Pi»8W:i*^-«^^!y«^.i,lii»^ •miism*''^^- (16) 0(X°) v.-itb x^ » 1 . We have ] 178 Dualizing, we obtain the following upper bound on v(X X°c^ * uCb-a"") max + - |x 11 (X°cJ - uaJ)x. I J j=l N J j=l J i-1 ' (17) ^ s.t. ^ -Oorl,jjlk X J s. = 0.1,2,. .,U. . 1 1 But the objective function in (.17) Thus, if the upper bound on v(X | simply the left hand side of inequality (16). is x^ = I) is less than v(a ), the project k can be ignored without loss of optimality 5. the best known solution, . ] SENSTTTJITY ANALYSE OF EFFICIENT SOLUTIONS For convex multicriterion optimization problems involving more than it is generally not possible to describe completely the subsets S(x is an efficient solution (see (5)). efficient solution at ^^ ... 2 criteria, . . of S in which x ) k Instead, the usual procedure is to find an and see how it changes in the direction V k Steuer [1973], Ceoffrion et al [1972]). For the (Evans and 8X multicriterion optimization IP problem (3), we are able to perform this sensitivity analysis through the dual problem. Specifically, suppose we have constructed an and let (x , s ) : Y denote the optimal ) = dual problerj which solves (3), IP solution to (3) found by the IP dual. We have v(X in ^ w(0 1 = ) Cx' and moreover. T w(x'^) " max (X°Cx')j y t=l T ^•'^- (Ax' + I IsSoi^. = b (18) t=l T u y t=l that is, = 1, u< all other u = ' ' 1; is optimal M 5 0; in this linear programming problem. Let w(\) denote the maximal objective function value in (18) as functions v{\) and v(\) are convex functions satisfying w(\) £ X varies. The v(X) by duality. Ii principle, we have only to perform linear programming sensitivity analysis on (18) ^WM *P' t-^ -i^'* ' ' W^-'L^WW .l^L.i!SMWWBtWWtUJj..i.-W>SWN"'JWU ji ^ ..^i J, r-" ^" ''''^**^^?' !a»r»)g3 W i5U«»ij ii^t)eii.i»»i i5,<(»jit f« , to compute the maximal value of with X» problem (3). 6uch that 2 and therefore that Ob) V (x ,s •" u. 1 remains optimal in (18) remains optimal in the ) IP mul ticriterion The difficulty is that the number of columns in (18) vill be enormous and column generation procedures are required. We oait further details except to state that generalized progranmiing, otherwise known as Dantzig-Wolfe decomposition (see Lasdon [1973], Magnanci, Shapiro and Wagner [1973], can be used for this purpose. 6. CONCLUS TONS In this paper, we have discussed multicriterion IP problems and how recent results in IP duality theory can be used to generate efficient solutions to these problems. The results presented here are far from complete and much research and practical experimentation remains to be done. As ve saw in Section 2, the IP dual methods are theoretically complete in the sense that an IP dual problem can always be constructed by a finite procedure which solves a given IP problem. Practical experience (Fisher, Northup and Shapiro [l97i], however, has shown that the IP dual methods sometimes, is to combine the dual methods with branch and bound The procedure of section In 5 but not always, The natural resolution of this difficulty require excessive numerical calculations. (see Fisher and Shapiro [197h]1 needs to be reconciled with this practical consideration. addition, the branch and bound procedures should be designed so that only one pass is made through the tree of enumerated solutions in computing the efficient V More generally, theoretical research is optimal in (3) as X varies in int S. x' needed into the structure of the family of IP duals (14) generated as X varies in S. There are properties of efficient solutions to problem (1) which it is important to study further and try to characterize. able to identify the set of all projects solutions. j For example, we would like to be which are included in all efficient Conversely, we would liVe to identify the set of all projects excluded from all efficient solutions. Another phenomenon to be studied which we have ignored here are the dynamics of public investment decision making. We eliminated an explicit study of dynamics by formulating the static model (1) of a T period problem. As mentioned in the introduction, some public Investment decision models in- volve distribution as well as location sub-problems implying the appropriateness of mixed P multicriterion The models. decomposition method for mixed IP IP duality theory can be combined with Benders' (Shapiro [1974] ). Thus, it appears that the approach suggested in this paper can be extended to these problems. nay wish to use max (u (x) ^^^f^VfS^- a concave utility function u (x)) instead of max Cx. ^ n£#^^iilV Ji^M J>itJ^jjejjl^!j Finally, we in the objective function of (l);that is The study of such problems is another . 180 area of future research. Bell, D.E., (19733), "The Resol'jtion of Duality Gaps in Discrete OptiTnizat ion", Tech Report So. 81, Operations Research Center, Hassachusec ts Institute of Technology. (1973b), "isproved Bounds for Integer Programs: A Supergroup Approach, IIA3A RM-73-5, International Institute for Applied Systems Analysis, Laxenburg, Austria (to appear in 5IAM J. on Applied Hath "*. and J.F. Shapiro (1975), "A Finitely Convergent Duality Theory for Zero-Or.^ Integer Prci?ramning", Operations F^esearch Center Vorking Paper No. OR 043-75, Massachusetts Institute of Technology. Bhatia, R., (19751, "Investment Planning for Petroleum and Fertilizer Industries: An Interregional Programming Model for India", Center for Population Studies Report, Harvard University. Dorfman, R., and H.D. Jacoby (1972), "As\ Illustrative Model of a Regional Water Quality Authority", in R. Dorfman, H.D. Jacoby, and H. A. Thomas, Jr. (eds.). Models f or Managing Regional Water Quality , Harvard University Press. Evans, J. P. and R.E. Steuer (1973), "A Revised Simplex Method for Linear Multiple Objective Programs", Mathematical Programming Vol. 5, 1, (August), 54-72. , Fisher, M.L., and J.F. Shapiro (1974), "Constructive Duality in Integer 27, 31-52. Programaiing", SIAH Journal on Applied Mathematics , Shapiro (1974), "Using Duality to Solve Discrete Optimization Problems: Theory and Computational F.xnerience", Operations Research Center Working Paoer No. OR 030-74, Massachusetts Institute of Technology (to appear in Mathematical Progranning ) W.D. Northup and J.F. , (1972), "An Interactive Approach Geoffrion, A.M., J.S. Dyer and A. Weinberg for Multi-Criterion Optimization, with an Application to the Operation of an Academic Department", .Management Science Vol. 19, pp. 357-368. , , Glover, (1969), "Integer Programming over SIAM Journal on Control 7, 213-231. F. , a Finite AdditTve Group", and A. S. Manne (1973), Multi-level Planning; Case St u dies in Goreux, L.M. Mexico North-Holland. , , Gorry G.A.. W.D. Northup, and J.F. Shapiro (1973), "Computational Experience with a Group Theoretic Integer Programming Algorithm", Matheijat ical Programming 4, 171-192. , ferlin, S. (1959), M.3the::iat Economics, Volume ical Methods and Theory in Games, Programming and 1, Addision-Wesley, Kendrick, D., and A. Stoutjesdijk (1975), The Planning of Industrial (Manuscript in progress). Programs; A Methodology , <« i u »iujjiB^^j».i> !i g[, w» «. i .kJLJ.J.'uuai i^ij»LL>^ Investment 13. and A. W. Tucker (1950), "Nonlinear ProgrammLng", Proc eedings the Second Berkeley Syrnposium on Mathematical Statistics and Probability , University of California Press, Berkeley, California, pp. 481-492. H.W. Kiihn, , of 16. 17. Kuhner, J., and J.J. Harrington (1974), "Large-Scale Mixed Integer Programming for Investigating Multi-Parity Public Investment Decisions: Apolication to a Regional Solid Waste Management Problem", (presented at 45th National ORSA/T IMS Meeting, Boston). Magnanti, T.L., J.F. Shapiro and M.H. Wagner (1973), "Generalized Linear Programming Solves the Dual", Operations Research Center Working Paper, No. 019-73, Massachusetts Institute of Technology, (to appear in Management Science} . 18. Lasdon, L.S., 19. Marks, D.H. ;0. Shapiro, J.F. (1970), Optin ir-at ion Theory f or Large Sys tems, Mirnillan. (1974), "Modelling for Policy Analysis in a Large Scale Underdeveloped River Basin", (presented at ORSA/T IMS Meeting, April, 1974, Boston). (\9bS) , "Dyn.imic Progranming Algorithms for the Integer Progratnni-.v Problem - I: The Integer Programming Problem Viewed as a ttiapsack T>-pe Problem", Operations Research , 16, 103-121. (1971), "Generalized Lagrange Multipliers in Integer Progranming", Operations Research 19, 68-76. , Integer Programming", "Operations Research Center Working Paper No. OR 033-74, Massachusetts nf the Congress Proceedings appear in the Technology (to Institute of on Mathematical Prograinming and Its Applications Rome, April, 1974. (1974), "A New Duality Theory for Mixed ~~ , 23. jt^t^ '*i*>*,1'l l!il>LW i Westphal, L.E. (1971), Planning Investments with Economies of Scale North-Holland ^ ^^g'g5fe3#j ;i , ^^» BMWi «|r8iS*f Lib-26-67 MAR" 4 19&' TDflO 00 3 ^^w 4"cj't,"""il' 7j^^^3 no.784- McLen nan. Ro^VThe^ 72474 M career^^of 3-7r ss mm^ R,rh,,rrt C / "^ TOaO 000 bM7 ^^2 3 T-J5 143-w no.785- 75 Hauser, John R/A normative methodology "" 724738 D»BKS D0Q2Q"' ' . t TOaO 000 bSb 2T5 3 t^^ f-;fr-W3^ no.786- 75a OPEL Kobnn, Stephe/Large scale direct n>BK^ 726385 -f^ vJ . ^ ,PQQ2(^65,4 - T06D 000 b5b 27 3 HD28.M414 \/l414 no.787- 75 Merton, 1, Robert/Option pricing when wh und "24743 . D»BKS . TOflO 3 0001985? 000 bMS bll m"i^ 3 TDflO DD^ S7b b71 HD28.M414 no.789-75 Mark/Issues Plovnick, J24742 3 . . D*BKS TOflO in the design of 00185829 002 720 Ofil ^-7S