The techniques of linear multiobjective programming

Transcription

1 REVUE FRANÇAISE D AUTOMATIQUE, D INFORMATIQUE ET DE RECHERCHE OPÉRATIONNELLE. RECHERCHE OPÉRATIONNELLE P. L. YU M. ZELENY The techniques of linear multiobjective programming Revue française d automatique, d informatique et de recherche opérationnelle. Recherche opérationnelle, tome 8, n o V3 (974), p < 8_3_5_> AFCET, 974, tous droits réservés. L accès aux archives de la revue «Revue française d automatique, d informatique et de recherche opérationnelle. Recherche opérationnelle» implique l accord avec les conditions générales d utilisation ( legal.php). Toute utilisation commerciale ou impression systématique est constitutive d une infraction pénale. Toute copie ou impression de ce fichier doit contenir la présente mention de copyright. Article numérisé dans le cadre du programme Numérisation de documents anciens mathématiques

2 R.A.I.R.O. (8 e année» novembre 974, V-3, p. 5 à 7) THE TECHNIQUES OF LINEAR MULTIOBJEGTIVE PROGRAMMING par P. L. Yu ( and M. ZELENY ( ) Abstract. In this article we dérive a generalized version of the simplex method-multicriteria Simplex Method, used to generate the set of all nondominated extreme solutions for linear programming problems with multiple objective functions. A simple nondominance subroutine is developed for testing the nondominance ofany extreme solution. We discuss an important interaction between Multicriteria Simplex Method and multiparametric linear programming. In f act we show that the décomposition of a multiparametric space (or a set of weights) into its optimal subsets can be obtained as its natural by-product. Theoretical results y numerical examples and flow diagram as well as some computer expérience are reported.. INTRODUCTION We are concèrned with linear programming problems involving multiple (possibly noncommensurable) objective functions. To résolve this type of décision problems, we could use the concept of domination structures (See Réf. -3) or linear multi-parametric programming (See Réf. 4-5). We propose a simple technique, Multicriteria Simplex Method, to generate the set of all nondominated extreme point solutions and show how the direct multiparametric approach (Réf. 6) turns out to be computationally inefficient. It is also redundant because the décomposition of the parametric space is a by-product of Multicriteria Simplex Method. Though, in gênerai, the solution to a multicriteria problem does hot have to be an extreme point, the en tire set of all nondominated solutions can be effectively generated from the set of all nondominated extreme points (see Réf. 3 or 5). Linear Multiobjective Programming represents a part of a broader field of study, Multiple Criteria Décision Making. We refer interested readers to some recent woçks summarizing up-to-date state of the arts (see for example Réf. 7, 8, 3). () Department of General Business, University of Texas, Austin, Texas 787. () Graduate School of Business, Columbia University, New York, New York 7. Revue Française d'automatique, Informatique et Recherche Opérationnelle n nov. 974, V-3.

3 5 P. L. YU AND M. ZELENY Before going further, for convenience, let us introducé the following notation. Let x = (x u..., x n ) and y = (y ls..., y n ). Then (i) x = y if and only if x i = j^ for ail y' =,... 5 TÏ. (ii) x ^ j> if and only if JCJ ^ yj for ail j =,..., n. (iii) x >.y if and only if Xj >, yj for ail y =,..., n and x = y. Usually we shall dénote a set or a matrix by a capital character. Given a matrix A, we will find it convenient to use A i and A i to dénote its Zth row and yth column respectively, and fl y its element in the ith row and the jth column. In order to simplify the présentation, let us assume that we have a compact décision space defined by X = { x R n Ax < b, x >, }, A is of order m X n. () Let C = C/ xn be a matrix with / rows (C,.. C z ) r so that C k x, k =,..., /, is the Ath objective function of our problem. Given a domination cône A (which is assumed to be convex) and x l 9 x X, we say that x is dominated by x if Cx Cx + A and Cx l # Cx. A point x JST is a N-point if it is not dominated by any other feasible point of X; otherwise it is a Z>-point. For simplicity, the sets of ail JV-points and ail Z)-points will be denoted by N and D respectively. If we dénote the set of all extreme points of X by X ex = { x,..., x r }, then let N ex.= N H X ex be the set of all nondominated extreme points. We see that N ex is finite because X is compact. Given X R\ let X (k) = {x X\ XCJC :> XCx, x Z }. () Thus, X (k) is the set of ail maximum points of \Cx over X. Note that ACx is bilinear in X and x. Given a cône A, we define its polar cône A* = { X X d 4, for ail </ A }. If A = { x Z>x ^ } is sipolyhedral cone 9 we see that A* {y D \ y ^ }. It can be shown that (Remark 5.9 of Référence ) the relative interior of A* is given by (A*) J = {yd\y > }. (It is understood that x is a column vector; y a row vector, Both represent vector s of RK) We present a theorem describing necessary and sufficient conditions for a point to be nondominated. Its proof is given in Réf. 3. Revue Française d*automatique, Informatique et Recherche Opérationnelle

4 (i) THE TECHNIQUES OF LINEAR MULTIOBJECIÎVE PROGRAMMING 53 Theorem.. Suppose that A is a polyhedral cone. Then i\rcu{x (X) A (A*)'} (ü) if Int A* ^O, then N = U { X {\) X Int A* }. The theorem holds even if X is not bounded. We may and will assume that A^{d^R l \d^} the présentation. = A- to simplify Using the results of Theorem., we shall dérive Multicriteria Simplex Method which may be regarded as a natural generalization of the simplex method. With this method we study the «connectedness» of N ex and dérive an algorithm to locate the entire set N ex.. SIMPLEX METHOD AND X (X) Recall that since we limit ourselves to the domination cone A = A=, it follows that IntA* = {d R l \d> } = A>, where Int stands for an interior. Recall from () that X (k) is the set of maximum solutions of XCx over X. Treating AC as a row vector, we see that to find X (k) is essentially a series of linear programming problems. REMARK.. Since X is compact, an optimal solution to XCx exists. We can generate the entire set of all basic feasible optimal solutions, say X^x = { x,..., x k }. Then the set of all optimal solutions to XCx is given by X (X) = H(X X ) (the convex huil generated by X ex ) (See Ref. or ). By varying X over A >, we can locate the entire set N via Theorem.. Although this method seems reasonable, it is by no means the best way to locate N, because how to vary X over A > (See Ref. 4 and 6) is still unresolved and the computational work may be quite demanding. Thus, instead of this direct approach we shall use Multicriteria Simplex Method to locate N ex. The method also indicates an efficient way to vary X over A > in order to get the set N ex. Without loss of generality we can assume that b ^> in (). (See Ref. and for the extension to other types of è.) From (), by adding slack variables, the décision space could be defined by the set of all x e R m+n, x % and n novembre 974, V-3.. (A,I mx Jx = b. (3)

5 54 P. L. YU AND M. ZELENY Our new C becomes (C, mxm ). Given a basis B which is associated with columns /= {JuJi* >»>j m }> we shall dénote the remaining submatrix and columns with respect to (3) by B' and J' respectively. Let us introducé a simplex tableau corresponding to some basic feasible solution, say x = (x B, XB) >, ), associated with B (and /) : r BASIS *. x m+l... JC,... X m +... J'lm+l m X Ô L i >W+ -y m -y mm +n... TABLEAU By the simplex method, we can systematically change /, one column at each itération, so that at each itération y = B~ l b ^ is maintained and the value of the objective function is improved until an optimal solution is obtained. For later référence, let us summarize some relevant results of the simplex method as follows (see Ref. ). Observe in Tableau that Y =s {jy }*=!,.,, = (I, B~ X B') (4) (5) (6) and (7) where XC B are criteria coefficients associated with B. Lemma.. Given a feasible basis J so that y = B l b >,, there are two possible cases which can occur in the simplex tableau : Case L Each ij >> for ail j e J\ Then x B = y = B~ x b and X' B = is a maximum solution. If each z s > forj /', then the optimal solution is unique. Otherwise there may exist infinitely many optimal solutions. Revue Française d'automatique, Informatique et Recherche Opérationnelle

6 THE TECHNIQUES OF LINEAR MULTIOBJECTIVE PROGRAMMING 55 Case. There is at least one; /' so that z j <. Let ). (8) Then by introducing theyth column into the basis, by Gaussian élimination technique we get the pth column of the identity matrix in the next tableau, (the element y pj is called the pivot element), and we obtain a new basic feasible solution with an increase of the value of the objective function by QJZJ. REMARK.. Given a simplex tableau corresponding to a basis J lt suppose we introducé theyth column, j /i, into the basis as described in Case of Lemma.. We thus produce a new basis J, J A U {j } U P } where j p J x is the column associated with I p in the simplex tableau of the basis J x. Observe that there is exactly one element in J which is not in J X9 and vice versa. Two bases such as J x and J which enjoy the above property are known as adjacent to each other. The corresponding extreme point solutions are called adjacent extreme points of X. 3, MULTTCRITERIA SIMPLEX METHOD Observe that given a basis B, the row vector z in Tableau is given by X(C B Y C). Let Z^(C B Y-C) (9) Then z = XZ. () From (9) and (), we see that given X the corresponding z can be easily computed whenever Z is known. Observe that (C B Y-^ C) = (C B Y C,..., C l BY C l ) T. Each C k BY C\ k=,...,/, can be obtained from the last row of the simplex tableau if we replace XCJC by C k x as the objective function. For a given basis B (or /), let us construct Multicriteria Simplex Tableau as Tableau (for simplicity, we have again rearranged the indices so that / appears in the first m columns). Note that { y u } is defined exactly as in (4) while V = (i?,..., v l ) = C B B~ b. Note that v k, k =,..., /is the value of the Ath objective function at the current basis. () Suppose there is no r so that y r $ >. We have an unbounded solution. Since X is assumed compact, this cannot happen. Thus 6; is well defined. n novembre 974, V-3.

7 56 P. L. YU AND M. ZELENY r BASIS x t x m+l... Xj... x m+n X *I yim+i yij yim+n y%o m x m b i ymm+ y m j y m o z )... V *J TABLEAU Then for k =,..., / we see that (*) Let us define M as (,...,, z* +,..., 4 +- ) - C ft ) = Z*. (m+ï)x(m+») Observe that M enjoys the following properties, (i) the submatrix { yj j j / }» when its rows are properly permutated» forms the identity matrix of order m x m. () (ii) The submatrix { Z y j / } is a zero matrix of order l x m. (3) For each nonbasic column ƒ e /', we shall define 6 y as in (8). By introducing the jth column into the basis we convert Mj into E p in the next tableau, where E p is the/?th column of the identity matrix of order m + l and p is such that y pj is the pivot element. At each such itération, M can enjoy the properties ()-(3) and F, Z can be easily computed. REMARK 3.. The row Z\ k =,.. /is associated with a linear programming problem with objective function C k x. In view of Lemma.. if at a basis /, Z k ^, then x(j), the basic feasible solution of /, is an optimal basic solution for C k x. If Zj >, for all j /'» then x(j) is the unique optimal solution for and clearly is an iv^-point. () Recall that Z k (or Zj) dénotes the kth row (or the jth column) of matrix Z, Similarly for matrix M in (). Revue Française d*automatique, Informatique et Recherche Opérationnelle

8 THE TECHNIQUES OF LINEAR MULTIOBJECTIVE PROGRAMMING 57 REMARK 3.. Given a basis / and j /', by introducing jth column into the basis we produee an adjacent basis J x (see Remark.). Then the values of the objective functions increase by QJZJ. That is V{J X ) V{J) QjZ p where V(J) = (v,..., v l ) at the basis /. This observation yields : Theorem 3.. Given a basis / (i) If there isj J r so that QJZJ ^, then x(/ ) e D. (ii) If there is j 6 JQ SO that QJZJ ^, then x{j x ) D 9 where J t is the new basis obtained by introducing the jth column into the basis. (iii) Let j, k /ó an( ï Jj an( ï «4 be the new bases obtained by introducing respectively the jth and kth columns into the basis. Suppose that QJZJ ^ k Z k. Then x(jj) D. Theorem 3.. and Remark 3.., although obvious, will be useful in our later computation of N ex. 4. OPTIMAL WEIGHTS AND A NONDOMINANCE SUBROUTINE Now, given a basis /, let Z be the matrix associated with /. We can then uniquely define. {X XZ^}. (4) Note that A(/) is a polyhedral cone and A(/). In view of Lemma.L, () and Theorem LI. we state Theorem 4.. (i) x(j) maximizes XCx over X for all X A(/). (ii) x(j) N ex ïï and only if A=" n A(/)#. REMARK 4.. Given /, A(/) is its associated set of optimal weights, because whenever our objectives Cx are linearly weighted as XCx for some X A(/), x(j) maximizes XCx. In the final decision-making, this is very valuable information. REMARK 4.. Given a basic feasible solution, we could use Remark 3., (i) of Theorem 3., and (ii) of Theorem 4. to detect whether it is an iv ex -pomt or not. However, although the results are useful, they cannot cover all possible cases. In the remaining part of this section, we shall dérive a simple algebraic method, called the nondominance subroutine, so that we can test whether an extreme point is an iv-point for all possible cases. Let x = x(j) represent a basic feasible solution with basis /. Let e = (e l9..., e t ) and n novembre 974, V-3. t w = max ]T e t (5)

9 58 P. L. YU AND M. ZELENY subject to : Theorem 4.. (i) * is an iv-point if and only if w =. (ii) JC is a /)-point if and only if w >. Proof. Observe that (x, ) X. Thus w =>. It suffices to show (i). However (i) is another way to define an iv-point with respect to the domination coneai. Q.E.D. x l Corollary 4.. If w >, then the corresponding maximum solution X y Cx x > Cx, is an int-point. Observe that finding whether w = or not in Theorem 4. usually does not require too much extra work. In order to see this, let B be the basis associated with x or /. The problem of (5) in a block simplex tableau can be written m x, (6) lx where = (,,..., ). *mxm /x m o Xm -lix, In the above matrix, the first and second columns are the coefficients associated respectively with the original variables and the added slack variables, the third column is the coefficients associated with the new variable e 'm (5). Note that (6) is the constraint that x X, (7) is the constraint that Cx-e >, Cx, and (8) corresponds to the objective of (5). We could rewrite (6)-(8) as follows : Cx (7) (8) B~ l A B~ l m x, B- x b (9) C B B~ l A C C B B-i /,x, ( xl () l l, l [C B B- i A C] Oix; () Note that (9) = B~ l (6), () - C B (9)-(7) (observe that C B B~ l b = Cx% and () = l lxf - () + (8). Observe that (9)-() supply a feasible simplex tableau for Problem (5) with the basic feasible solution (x, e) = (x, ). Revue Française d'automatique, Informatique et Recherche Opérationnelle

10 THE TECHNIQUES OF LINEÂR MULTIOBJECTIVE PROGRÂMMING 59 Comparing (9) and () with (4), (9) and () we see that (*). B-t C B B -i I -\Z () From (9)-() we see that to construct a simplex tableau for Problem (5) does not require much extra work. The conditions in Theorem 4. could be easily verified. In particular, we have the following sufficiency condition : Theorem 43. Given a basis J, suppose x X Z ^. Then x(j) is aniv^-point. Proof. Because the first two blocks of (), given by l X fz, X Z ^ implies that (x(j), ) is an optimal solution to (5) with value w = (see Lemma,). The assertion follows immediately from Theorëm 4.. REMARK 4.3. Suppose that the condition in Theorem 4.3 is not satisfied. Because of the special structure of (9)-() 5 the problem of (5) usually can be simply solved in a few itérations. In order to use the results of (9)-() and Theorem » one can append an extra row, corresponding to the objective function x jc, to the simplex tableau. (See the example discussed in Section 6.) 5. DECOMPOSITION OF A > AND CONNECTEDNESS OF N, ex Given a basis /, we could define its set of optimal weights A(/) as in (4). Now suppose that for some k J\ Z k ^, fe < oo. Let us introducé the k th column into the basis. Suppose that y pk is the pivot element. Then we will produce an adjacent basis K so that K' = J'U{j p }-{k} Without confusion (rearrange the indices, if necessary), let p = j r Then (3) Let W dénote Z(K) (the submatrix Z associated with K). We want to study the relation between A(/) and A(K). Toward this end. observe that by Gaussian élimination technique, if j K Z k fy pk if j = p K' (4) Zj y PJ ZJy pk if/ K f - () Note» Cj if i is an index associated with a slack variable, n novembre 974» V-3.

11 6 P. L. YU AND M. ZELENY Since it is the pivot element, y pk > (see Lemma.). Let H k = { X I XZ* = }. (5) Since y pk >, X( Z k /y pk ) ^ if and only if lz k ^. We see that A(K) C { X j XZ* ^ }. (6) But, (7) We see» from (5)-(7), that H k is a hyperplane in R\ which séparâtes the polyhedral cônes A(K) and A(/). Next» since X H k implies that XZ ft =, we have jy*nàgo = {x xz* «o, xz,.^ o,./ /' {*}} (8) and from (4) we also have J (9) However from (3), we have K' {p } = /' {k }. Thus (6>(9) imply that i/ fc n A(/) = H k n A(z) = A(/) n A(^). (3) We summarize the above results into Theorem 5.L Given a basis /? suppose that Z k^q and 8 fc < oo. Let K be the adjacent new basis obtained by introducing the k th column into the basis. Then H k as defined in (5) séparâtes A(J) and A(K). Furthermore, the equalities of (3) hold. RBMARK 5.. Given A(/) and A(K) 9 we say that A(/) and A(K) are adjacent if (6), (7) and (3) hold. Theorem 5. says that by introducing the column k with Z k = and fc < oo, into the basis, the new adjacent basis K will produce A(K) which is adjacent to A(/). However, it is possible that A(J) n A(K) = { } and A(^T) A= = { }. If this case occurs, introducing the k th column into the basis does not help solve our problem. This case can be avoided if H k n A(/) iflà^ { } (thus the intersection contains more than the zero point). A column k j f with this property will be called an effective constraint of A(/). Note from Theorem 5. that by introducing an effective constraint Z k of A(/) into the basis, we will produce A(^T) which has a nonempty intersection with A ~ { }. Now observe that for a given X, XCx will either have an unbounded or optimal solution over X. In either case, by simplex method, X will be contained by some A4 or A(/ fc ). (Observe that Ai, identifies A(/) for Q k = ). Since we have a finite number of bases and each basis has only finite number of Revue Française d'automatique, Informatique et Recherche Opérationnelle

12 THE TECHNIQUES OF LINEAR MULHOBJECTIVE PROGRAMMING 6 columns, we know that there are finite number of Ai and A(/ fc ) which form a covering of A. More precisely. Theorem 5.. There are { Ai j =,...,p } and { A(J k ) Ar =,.,., q } so that each Ai or A^) has a nonempty intersection with A- and (i) A = C U { Ai j =,.,ii } U { A(J k ) \k=h..., q } (ii) U { A(/ fc ) Ar =»..., q } fl A - is a closed convex polyhedron. In fact it is a polyhedral cone. Proof. (i) is clear from the previous discussion. In order to see (ii), observe that XCx cannot simultaneously have an optimal solution and an unbounded solution. Thus { Ai } and { A(J k ) } are mutually disjoint, and = n { Comp Ai j =,...,p } n A -, where Comp désignâtes a complement to a set. (The first equality comes from (i) and mutual disjointness which implies that X A- is in some A(J k ) if and only if it is not in some Ai.) We see that each Comp Ai is a closed half space. Our conclusion of (ii) is clear from (3). Q.E.D. REMARK 5.. Theorem 5. states that there are finite number of A i and A(J k ) that will cover A-, Such Ai and A(J k ) can be located by Multicriteria Simplex Method. Theorem 5. and Remark 5. provide a way to generate adjacent «nonoverlapping» polyhedral cônes in the parametric space. It is not reasonable to adopt a method of direct décomposition of A- to résolve our problem. The method starts with a A(/) so that A(/) H À M, then uses Theorem 5. to generate the adjacent A(J k ) or Ai. The procedure is repeated until A ï is completely covered by { A(J k ) } and { Ai }. This method was discussed in [6]. For a detailed discussion of its shortcomings and inefficiency see Réf. 4. REMARK 5.3. Once J is found to be an JV cx -basis (Theorem 4.), Eq. (4) can be used to find its related set of optimal weights A(/) at no extra work from the multi-criteria simplex tableau. Thus our remaining crucial task is to find the set N ex by Multicriteria Simplex Method. Let E = { x(i) Ï =,..., p } be a set of extreme points of X. We say that E is connected if it contains only one point or if for any two points x{ï)y x(k) in E 9 there is a séquence { x(i t ),..., x(i r ) } in E so that JC(Z'J) and x(i l+ x % =,..., r, are adjacent and x(i x ) = JCQ, x(i r ) = x(k). Following a similar proof as in [5], we have n novembre 974, V-3.

13 6 P. L. YU AND M. ZELENY Tfaeorem 5.3. The set N ex is connected. Proof. Let x(i), x(j) N ex. Suppose Iand /are the bases associated with x(i) and x(j) respectively. Then, by (ii) of Theorem., both A(7) fla > and A(7) fl A > are not empty. Let X ( A(/) n A > and X, A(J) D A >. SinceA > is convex the line segment [k it X^] CA >, From Theorem 5., we can find a finite séquence { A(J k ) \ k =,..., r } so that [X, X,] fl A{J^ ^ and In view of (ii) of Lemma. and Theorem., we see that we can find a séquence of iv ejc -points { x iu..., x ir } so that x n is adjacent to x a + u l,..., r, and x n = x(i), x ir = *(ƒ). Q.E.D. REMARK 5.4. In view of Theorem 53 we can construct a connected graph (A,, C Ü) for N exi where SJ is the set ofallverôœscorrespondingtoiv^, and yfc is the set of all arcs in the graph. Given (! ) x, x N ex the are a(x l, x ) which connects x and x is in A if and only if x l and # are adjacent. With this définition we see that the graph (&> c ü) is connected. 6. A MEIHOB TO GENERATE THE ENURE SET N ex AND AN EXAMPLE In order to generate the set N ex, we can first find a basis J x for an iv^-point, if N ex =. In view of Remark 5.4, if there is any other JV^-point, we must have an iv^-basis J adjacent to J v Thus we could use results of this section to search for such a J. If there is no such J > J\ is the unique iv ex -point. Otherwise, we consider all adjacent, but unexplored feasible bases to { J u J } to see if there is any other iv^-basis among them. If there is none, { J u J } represents the set N ex. Otherwise, we add a new iv^-basis to { J u J } and continue with the procedure until the entire set N ex is located. We shall use Flow Diagram to explain our procedure more precisely. In the diagram, we have used the following notation : (i) For each basis /, we use )(/) to dénote the set of all «obviously» dominated bases which are adjacent to /. That is, those dominated adjacent bases which can easily be checked by Theorem 3.. We also use A(J) to dénote the set of all adjacent bases to / which are not in 3)(/) and their nondominance have not been checked before. Thus A(J) dénotes the set of all adjacent bases to /of which the nondominance must be checked by nondominance subroutine. (ii) At each step i 9 N t and D t are the sets of all checked nondominated and dominated extreme points respectively, while W t is the set of all possible bases of which the nondominance must be established by the nondominance subroutine. () It is convenient, without confusion, for us to use x\ x % to represent thèir bases Ju J% and the resulting basic feasible solutions x(/i), x(j%) as well. Revue Française d'automatique, Informatique et Recherche Opérationnelle

14 THE TECHNIQUES OF LINEÀR MULTIOBJECÜVE PROGRAMMING 63 () START ( ) () Pind an N ex Basis J.,P(J.), A(J.) A l I (5) i» ( (8) n novembre 974, V-3.

15 64 P. L. YU AND M. ZELENY We briefly describe Flow Diagram as follows : Box (). From our assumption that X is compact,- we know that N ex ^. Since ± x t A >» the maximum solution to x t Cx over X is an iv^-point. We may use this iv^-point te start with. (/j) and A(J t ) are to found by Theorem 3.. Box () and (3) are clear. Box (4)-(6). Suppose W % =. Since N ex is connected (Theorem 5.3 and Remark 5.3), we know that we have already located all JV^-points. Thus we stop at Box (6). Otherwise, we go to Box (5). Observe that if N ex = N i9 then there are i ist^-points. Box (7)-(ll). In Box (7) weusenondominancesubroutine to verify whether K is an int ex -basis or not. If it is, we get one more TV^-point and go through Box (9)-(ll). Note, in Box (9), again we use Theorem 3. to find 3)(Z). To find &{K) we need to use the record of N t and D t. Once 3D(Z) and &{K) are found, Box () and () are clear. Suppose that K is not an JV^-basis. We go to Box (8). We see that D t is increased by one, while W t is decreased by one. An Example (Problem ) The objective functions : 3 Cx = 3 The constramts / x t ) / 6 \ Ax = Xj / ^ 6 / Xj >,, i = l,..., 7 We set up the initial multicriteria simplex tableau as in Tableau. Observe that the last row of the tableau is corresponding to the row of x t Cx (see Remark 4.3). Revue Française d'automatique, Informatique et Recherche Opérationnelle

16 THE TECHNIQUES OF LINEAR MULTIOBJECTIVE PROGRAMMING 65 x x * -8 x 9 x Xxl *8 x () V V v ' w i -4-5 TABLEAU 3 Observe that 6 5 Z 5 <. In view of Theorem 3., the current basis is dominated. By introducing column 5 into the basis (the circle indicates the pivot element), we get Tableau 4. x *3 x. x 6 x X% Xg X 5 x 9 * / - -/ - / - 3/ 6/) (} -3/ / / - 4 -' / - - -/ 8 8 V 3/ - / - - -/ / 3 3/ 6 4 w / 3/ -3/ 5/ 5 5/ 4 TABLEAU 4 Now, Theorem 3. could not teil whether the current basis is nondominated or not. We have to use nondominance subroutine. Observe that the last column of () and () is. Thus each,- = for the problem (5). To know whether w > or not, it may sufls.ce to check only () and (). Also from Tableau 4, we see Zj = for j = 9,,. In checking () and () these columns will never be changed. Deleting the first four rows and columns 5, 9- of Tableau 4, we get the essential part of the simplex tableau associated with ()-() for verifying whether w > or not as in Tableau 5. Observe that the n novembre 974, V-3.

17 66 P. L. YU AND M. ZELENY last row is the criterion row for the subproblem (5). Going through the usual simplex method, we find an optimal solution for (5) in Tableau 7. Observe that since w =, the basis associated with Tableau 4 is nondominated. Dénote its basic feasible solution by xk Then x l = (,,,, 8,, ). Xl x *3 x A x 6 *7 *8 e e 3 e l e 3/ - / - -/ ci> / 3 3/ W / 3/ -3/ 5/ 5 5/ TABLEAU 5 * *3 x. x 6 *7 *8 e i e *3 e l e - CD / 3/ I / w 4 3/ 5/ 3/ TABLEAU 6 x x x Xl x. *7 e e / 9/- 4 3/ 3/ 3 / 3/ w TABLEAU 7 From (4) and Tableau 4, we get A(x x ) to be the solution set of the foliowing system of linear inequalities (Xi,X,X 3 ) 3/ / ^-/ / 3 3/ ^ (3) Revue Française d'automatique, Informatique et Recherche Opérationnelle

18 THE TECHNIQUES OF LÏNEÀR MÜLTIOBJECTIVE PRÖGRAMMING 67 In order to have a two dimensionàl graphical représentation, observe that each ray of A - C- R 3 is one-to-one corresponding to à point of the simplex Thus, by setting X x = X X 3 and deleting the redundant constraints of (3) we get the set A(x l ) in S as A X 3 ^, 3X / + 3X 3 = } (3) The set of A(x l ) O 5 is given in Figure. Figure Observe that if we use y s and y 35 as pivot éléments we will get x () and x (), both of them are different bases associated with x l. However, From the basis x () A(x Uj) )ns^, 7 =,. we get another basis for xk We call it x (3>. Note, = (r i Following the procedure described ia- th&t previousr section, we get the set N ex and its related n novembre 974, V-3.

19 68 P. L. YU AND M. ZELENY CORRESPONDING BASIS VALUES OF OBJECTIVE FUNCTIONS C l x CORRESPONDING SUBSET OF A X X X 3 X 4 X 5 {5,9,,} {4,9,,} {,9,,} {,3, 9,} {3,4, 9,} {3,5, 9,} /3 6/ /3 64/ /3 6/3 M* 4 ) JC < > {,5,,} {,4,,} {3,4,,} {,3, 5, 9} / / /3 TABLEAU 8 where x = (O, O, O, O, 8, O, ), x = (O, O, O, 6,, O, Ó), x 3 ^ (6,,,,,,), x* = (8,,8,,,,), and x 6 = (o,,^,,^,, Also, we get the fouowing subsets : = { > 3X 3 3X 3 A(* ) = { X I 4X 3 ^ ; 3X + 6X3 ^ 4 ; 3X + 6X 3 ^ 4 } = { X I 3X + X 3 ^ ; 4X 3 ^ ; 3X + X 3 ^ O = i X ' 3X 3 ' 3X 3 A(x 5 ) - \ X = i X X + X 3 Revue Française d'automatique, Informatique et Recherche Opérationnelle

20 THE TECHNIQUES OF LINEAR MÜLTiOBJECTIVE PROGRAMMING 69 and -{ I = { X I 9X + 4X 3 <: ; 4X 3 ^ } A(JC () ) = { X I 9X + 4X 3 ^ ; 3X + 6X 3 ^ 4 ; 3X. + 6X 3 ^ 4 A(x 6() ) = \ X 3' J i; 7. A COMPUTER EXPERIENCE We have coded Multicriteria Simplex Program in FORTRAN according to Flow Diagram (for details see Ref. 5). Our examples are executed on IBM 74 O- It takes a total time of.88 minutes to get the set N ex for the problem described in Section 6. Observe that the problem is by no means trivial becausë of the existence of degeneracies and alternative solutions. We then try the following problem : The Objective Functions fx x \ \x 8 / The Constraints / x i\ \x 8 ) f 4\ l /? =,...,8. () The speed of IBM 74 is much slower than that of the IBM 36. We will try in the near future to exécute the same problems on the IBM 36 and report the expérience. A useful computational expérience with an alternative algorithm is available in Ref. 5. n novembre 974, V-3.

21 7 P. L. YXJ AÎNJD M. ZELENY Observe that this problem is intentionally complicated. For instance, ç»3 ^6 ( no te 3 C 3 is the third row of the objectives, while A 6, is the sixth row for the constraints), also A 3 A C. Such dependencies willcertainly make our computation more lengthy. Note that in this problem the upper limit on the number of feasible bases is I I =,87. However, we get only 3 7V ex -points. ït takes a total time of.84 minutes to exécute the problem. In the next problem we use the same constraints as in the previous one, however we have five objective functions : " 6 / Y \ \X 8 J The set N ex contains 7 points. It takes 5.65 minutes to carry out the com^, putation. For details see référence 5. Observe that our method for locating N ex indeed is a combination of a modified linear program and an enumeration technique. The time required to locate N ex consequently dépends on the size of the problem (the dimensionality of A and C) and the number of total iv^-points (the interrelation among the rows of A and C). One can easily imagine that when the dimensions of A and C get large, it might become very difficult to incorporate the method efficiently. To illustr^te how the number of iv* JC -points can effect the computation time, observe that our first problem has a lower dimensionality than the second problem but it takes more time for locating N ex because its N ex contains more éléments. Also, although the third problem has the same A as the second one, it takes much longer to locate its N ex than it does for the second problem because its N ex contains 7 éléments while the second one contains only 3 éléments. 8. CONCLUSIONS We have shown that the set of ail nondominated extreme points can be generated quite efïiciently. In complex multicriteria situations the final solution can be any iv-point, not necessarily an iv ex -point. Therefore we might be interested in finding N rather than N ex. One method using N ex to generate complete set ivis discussed in Références 3 and 5. This method produces the set of nondominated faces of a convex polyhedron. The nondominated solutions are essentially the first step toward good décision making. Actual sélection of the final solution from N remains a chair Revue Française tf Automatique, Informatique et Recherche Opérationnelle

22 THE TECHNIQUES OF LINEAR MULTIOBJECTWE PROGRAMMING 7 lenging task, see Ref. and 4. For an excellent review of possible ways of resolution see Référence 8. REFERENCES [I] Yu P. L., Cone Convexity, Cone Extreme Points and Nondominated Solutions m Décision Problems with Multiobjectives, Center for Systems Science, University of Rochester, CSS 7-, 97 (To appear in Journal of Optimization Theory and Applications). [] Yu P. L., Introduction to Domination Structures in Multicriteria Décision Problems; Systems Analysis Program, University of Rochester, No. 79 (In : Multiple Criteria Décision Making, edited by J. L. COCHRANE and M. ZELENY, USC Press, Columbia, 973). [3] Yu R L. and ZELENY M, The Set of All Nondominated Solutions in the Linear Cases and A Multicriteria Simplex Method, Center for Systems Science, University of Rochester, CSS 73-3,973 (To appear in Journal of Mathematical Analysis and Applications). [4] Yu P. L. and ZELENY M., On Some Linear Multi-Parametric Programs, Center for Systems Science, University of Rochester, CSS 73-5, 973. [5] ZELENY M., Linear Multiobjective Programming, Springer-Verlag, New York, 974, p.. [6] GAL T. and NEDOMA J., «Multiparametric Linear Programming», Management Science, Vol. 8, No. 7, Maren 97, pp [7] COCHRANE J. L. and M. ZELENY (eds.), Multiple Criteria Décision Making 9 The University of South Carolina Press, Columbia, S. C, 973, p. 86. [8] ROY B., «Problems and methods with multiple objective functions», Mathematical Programming, Vol., No., 97. [9] STOER J. and WITZGALL C, Convexity and Optimization in Finite Dimensions, Springer-Verlag, New York, 97. [] HADLEY G., Linear Programming, Addison-Wesley, Reading, Mass., 963. [II] DANTZIG G. B., Linear Programming and Extensions. Princeton University Press, Princeton, N. J., 963. [] ZELENY M., «Compromise Programming», In : Multiple Criteria Décision Making, edited by J. L. COCHRANE and M. ZELENY, USC Press, Columbia, 973. [3] ZELENY M., «A Selected Bibliography of Works Related to the Multiple Criteria Décision Making», In : Multiple Criteria Décision Making, edited by J. L. COCHRANE and M. ZELENY, USC Press, Columbia, 973. [4] ZELENY M., «A Concept of Compromise Solutions and the Method of the Displaced Ideal», (To appear in International Journal of Computers and Opérations Research). [5] EVANS J. P. and STEUER R. E., «Generating Efficient Extreme Points in Linear Multiple Objective Programming : Two Algorithms and Computing Expérience», In : Multiple Criteria Décision Making, edited by J. L. Cochrane and M. Zeleny, USC Press, Colombia, 973. n novembre 974, V-3.