• Aucun résultat trouvé

An interactive interior point algorithm for multiobjective linear programming problems

N/A
N/A
Protected

Academic year: 2021

Partager "An interactive interior point algorithm for multiobjective linear programming problems"

Copied!
8
0
0

Texte intégral

(1)

An interactive interior point algorithm for multiobjective linear programming problems

B. Aghezzaf

, T. Ouaderhman

Departement de Mathematiques et d’Informatique, Faculte des Sciences Ain chock, B.P. 5366 Marif, Casablanca, Morocco Received 19 July 1999; accepted 10 January 2001

Abstract

We propose an interactive interior point method for .nding the best compromise solution to a multiple objective linear programming problem. We construct a sequenceX0; X1; : : : ; Xk; : : :of smaller and smaller polytopes which shrink towards the compromise solution. During the kth iteration, we move from the center of polytope Xk−1 to the center of polytope Xk by performing a local optimization which consists of maximizing a linear function over an ellipsoid. c2001 Elsevier Science B.V. All rights reserved.

Keywords:Multiobjective linear programming; Interior point methods; Interactive methods; Compromise solution

1. Introduction

The area of multicriteria decision making (MCDM) in general, and that of multiobjective mathematical programming (MOMP) have been very active in recent years. As a result, a number of methods for addressing these problems have become available (see, [7,13]). The purpose of these solution methods is to optimize the decision maker’s utility (prefer- ence) function. Depending on the assumption made on the utility function, the methods can be classi.ed as utility theory, goal programming, and progressive articulation and posterior articulation approaches. The progressive articulation approach, called the interac- tive method, is becoming increasingly popular and is considered a promising approach. The interactive methods are designed to .nd one “most preferred”

Corresponding author.

E-mail address: aghezzaf@facsc-achok.ac.ma (B. Aghezzaf ).

solution for problem (MOMP), that is dictated by an explicit or implicit utility function that belongs to the DM. Every algorithm in this class approaches has to develop two basic capabilities: (1) trade-oB mecha- nism, and (2) an optimization algorithm. We discuss each of these aspects brieCy below.

Although researchers have developed numerous interactive solution methods, there are only a few multiobjective interior point methods in the literature [1–6,14]. The solution approaches that are currently available use the aEne scaling primal algorithm, or the primal–dual algorithm with various schemes for de.ning a single interior search direction [2,5,6]. Two other approaches developed by Trafalis et al. [14] use the, so-called method for centers to both linear and nonlinear multiple objective programming problems, where the decision space is a convex polyhedron de- .ned by linear inequalities. The .rst algorithm is a generalization of Renegar’s algorithm [10] to mul- tiple objective setting, that .nds an eEcient point

0167-6377/01/$ - see front matter c2001 Elsevier Science B.V. All rights reserved.

PII: S 0167-6377(01)00089-X

(2)

of the polytope. The second, is based on techniques of algebraic geometry related to the parametrization of algebraic varieties inn-dimensional spaces. It ap- proximates a portion of the set of eEcient faces by an algebraic surface using the nonnegative orthant as the ordering cone to reduce the decision space.

Since the cone depends on the inscribed ellipsoid, the best compromise solution may not be retained in the reduced space.

To remedy this situation, we present an interac- tive interior point algorithm for determining a best compromise solution to a multiobjective linear pro- gramming problem (MOLP) in situations with an im- plicitly de.ned utility function, which is illustrated by deriving the local trade-oBs (the marginal rates of substitution) by interacting with the decision maker (DM). The algorithm is based, in principle, on the same idea used by Vaidya [15] in a single objective algorithm. By employing this approach, we construct a sequenceX0; X1; : : : ; Xk; : : :of smaller and smaller polytopes which shrink towards the compromise so- lution, the algorithm generates a sequence of points x0; x1; : : : ; xk; : : : ;such thatxkis in the interior of poly- topeXk, by maximizing a linear function over an el- lipsoid.

The paper is organized as follows. In Section 2, the problem statement and some fundamental de.nitions are stated. Section 3 describes the algorithmic proce- dure of the method and the results of the computational studies with the method are included in Section 4.

2. Problem denition and preliminary discussions Multiobjective linear programming (MOLP) prob- lem is concerned with determining scalar values of ncontinuous decision variables in order to maximize l(l¿2) objectives simultaneously. Amathematical formulation of the problem is

(P) max Cx={z1(x); : : : ; zl(x)}

s:t: xX={x∈Rn|Ax¿b};

whereC andAarel×nandm×nmatrices respec- tively, andbRm. Here, the row cTi;16i6lofC are the coeEcients of 1 linear objective functionszi. In order to make the problem nontrivial, it will be as- sumed that the objectives are in conCict. Due to the conCicting nature of the objectives, an optimal solu-

tion that simultaneously maximizes all the objectives is usually not obtainable. Instead, there are several solutions, called eEcient solutions (nondominated or pareto optimum), that have the property that no im- provement in any one objective is possible without sacri.cing one or more of the other objectives. The set of all eEcient solutions (in continuous case) is known as the eEcient frontier.

Denition 2.1. Asolution LxX is said to be eEcient if there exists no pointxX such that

cTix¿cTixL ∀i;

cTix ¿ cTixL for at least onei:

With the above de.nition, the solution to a MOMP problems reduces to the problem of .nding all eE- cient solutions. Often, there are an in.nite number and they are not comparable. Hence, it is assumed that the decision maker has a real valued utility function (value function or preference function)U, de.ned on the values of objectives, which has not been accessed and is not explicitly known, with this assumption, the MOMP problem can be rede.ned as follows:

(Pu) max U(z1(x); : : : ; zl(x)) s:t: xX:

Within this mathematical framework, the primary ob- jective of the interactive approaches is to .nd a best compromise solution through interaction with the DM.

Denition 2.2. Abest compromise solution (or op- timal solution) to MOLP is an eEcient solution that maximizes the decision maker’s utility function. That is, a feasible solution Lxis the best compromise solu- tion if and only if

U(Cx)L ¿U(Cx) for allxX:

Many other technical rules were developed to ad- dress the MOLP problem (P). One class of approaches is developed in [14] by considering an ordered cone.

With the Trafalis algorithm, an approximation of a portion of the eEcient frontier is given by construct- ing a sequence of algebraic surfaces {Sk}. The gen- eral surfaceSkcan be considered the loci of analytical centers of intersection of an ordering coneand the polytopeX, where the apexofis moving onSk−1

(3)

Fig. 1.

by taking the initializing manifoldS0 to be a part of the inscribed ellipsoid ofX. But the question which arises is how to construct the conethat illustrates the DM preferences.

In using the cone of increasing direction C¿= {x∈Rn=Cx¿0} as the ordered cone, the portion of the eEcient frontier obtained by Trafalis algorithm also depends on the geometry of the polytope X as it depends on in the general case, since the al- gorithm is dependent on the inscribed ellipsoid of X. Consequently, the .nal surface S generated by the algorithm might possibly be a portion of the ef- .cient frontier that is of no further interest, as one can never be completely sure about the existence of onek which allows the desired solution in the re- stricted polytope and how to select it if that exists (see Fig. 1).

3. The proposed algorithm

In this paper, with (Pu) as the problem under con- sideration, an interactive interior point algorithm that can be used to solve MOLP will be developed and tested. For the development of the method, it is as- sumed thatX is bounded, andU is strictly increasing in each of its arguments, concave and diBerentiable on Rl. As the polytopeX is bounded, we can assume that m¿n, and the columns ofAare linearly independent.

To solve the multiobjective optimization problem, we have to .nd a way to approximate the utility function

at the current iterate and take a step in a direction that results in an improvement of the value of the objective function. The partial information on the utility func- tion of interest is the local trade-oBs (marginal rates of substitution) between the objective functions which are obtained by interacting with the decision maker.

This concept is .rst proposed in multiple criteria de- cision making by GeoBrion et al. [8].

We assume thatz1(x) is a reference criterion.

Denition 3.1. The tradeoB ratio between the objec- tivez1(x) andzi(x) at pointxkis

!1i=(@U=@zi)(xk) (@U=@z1)(xk):

Consider that the property of the utility function, the partial derivative ofU at a pointxkis

xkU= @U

@xk1; : : : ; @U

@xkn

:

= @U

@z1

l i=1

!k1i @zi

@x1k; : : : ; l

i=1

!k1i @zi

@xkn

= @U

@z1

˙xkU= @U

@z1

!kC;

where!k= [!k11; : : : ; !k1l] and ˙xkU=!kC;˙xkU is termed a positively scaled local gradient ofU atxk.

In developing our approach, the motivation has been to overcome the diEculties encountered with the Trafalis algorithm mentioned before, and gen- erating a single interior direction at current iterate that improves the utility function without combin- ing it with other directions resulting from anchoring points as in [2,6]. This algorithm also follows the Renegar approach [10], but there are three criti- cal diBerences which enable us to solve the above diEculties.

First, the reduced space is obtained by using an updating formula to update the lower bounds of objective functions instead of the ordered cone.

Second, the potential function is formulated in a manner to orient the iterates to the desired solu- tion. Third, the generate iterate is computed along a direction that could be diBerent from the direc- tion in Newton method. On the other hand, the problem is how much the lower bounds values

(4)

Fig. 2.

of the objective functions can be changed in order that the reduced space contains the compromise solution.

In the following, we propose an appropriate way to update them.

3.1. Generating a polytopes

Let xk be an interior point of X and consider in the following the updating formula used to update the objectives lower levels, that gives in iterationk, the added setCk of constraints:

Ck={x∈Rn|Cx¿(xk)};

where

(xk) = [1k+1; : : : ; k+1l ]T; ik+1=ki +ki(cTixkki);

16i6l and 06ki ¡1;

i06min

x∈X cTix; 16i6l; (1)

where the coeEcients{ki} will be changed, in gen- eral, from one iteration to another. By choosing con- veniently the values of the step-sizeski (06ki¡1) at each iteration in the updating formulas, any point in the eEcient frontier can be obtained, which is not available in the algorithm of Trafalis et al. (see Fig. 2).

And at each iteration, also we add another con- straint given by the partial information. That con- straint, which is used to eliminate a certain portion of

the decision space, is considered in such a way that the reduced decision spaceXk contains the best com- promise solution andxk in his interior. The proposed constraint is given by

˙xkUx¿k; (2)

where

k=!k(xk) +( ˙xkUxk!k(xk)) and

0¡ ¡1: (3)

Thus, the original decision space is further reduced by adding the above constraints to the existing set of constraintsX. LetXk denote the reduced decision space at thekth iteration, then

Xk={x∈Rn|Akx¿bk} ∩Ck

={x∈Rn|Ax¿b; ˙xkUx¿k;

K= 0;1; : : : ; k} ∩Ck: (4) By using the following equality:

i=l i=1

(k+1i ki) =i=l

i=1

ki(cTixkki)

=i=l

i=1

ki(cTixk−zi)+i=l

i=1

ki(ziki);

(5) where zi=cTix; xX, and if the lower bounds sat- isfy the aspirations of the decision maker, he will pre- fer not to update them, hence:i=l

i=1(k+1i ki) = 0.

The coeEcients{ki}that yieldsz=(xk) will satisfy i=l

i=1 ki(cTixk−zi) = 0, which could be guaranteed by ki =!k1iorcTixk=zi.

So, we propose that the values of the step-sizeski depend on the DM preferences and given by

k+1i =ki +!k1i mk

i=l i=1

!k1i

(cTixkik);

16i6l and mk=m+l+k:

3.2. Locating an improving direction of U in the reduced space

The purpose of this step is to locate a direction of increase for (MOLP) problem in the reduced decision

(5)

space. The scaled local gradient ofU is calculated at the current interior pointxk, that is completely speci- .ed once the local tradeoBs are obtained from the DM, and used to locate an improving direction ofUin the reduced spaceXk.

Consider the following problem:

(Pk) max ˙xkUx s:t: xXk:

The associated potential function would be Fk(x) =

mk

i=1

ln(Akixbi) + l

s=1

ln(cTsxsk+1)

+(mk) ln( ˙xkUxk):

Since the columns ofA(andm¿n), soAk, are linearly independent,Fk(x) is a strictly concave function over the interior ofXk, and the analytic center ofXkgiven by maximizingFkis indeed a unique point. Hence, the problem is converted into a new form by maximizing Fk(x) over the interior ofXk.

One more interesting concept is to introduce the in- scribed ellipsoid ofXk, denoted byEk(1), and maxi- mizingFk over it. The ellipsoidEk(1) is given by Ek(1) ={x∈Rn|(xxk)THk(xxk)61};

where Hk=mk

i=1

Aki(Aki)T

(Akixbki)2 +l

s=1

cscTs (cTsxk+1s )2

+(mk)( ˙xkU)T( ˙xkU) ( ˙xkUxk)2 :

At iterationk, the next iterate is obtained by maximiz- ing the linear approximation ofFk(x) overEk(1) and we have to solve

max F(xk)x;

s:t: xEk(1):

Letzk be the solution point of this problem. We shall now describe how to .nd the next iteratexk+1 from xk. From the theory of convex optimization, (zkxk) satis.es the system of linear equations, given by the (KKT) conditions:

Hk(zkxk) =tF(xk): (6)

For some scalar t ¿0, we .rst compute"k, a vector in the direction of (zkxk) by solving the system of linear equations

Hk"k=F(xk):

Next, we .nd a scalartk¿0, such that (tk)2("k)THk"k61:

Once the direction"kandtkare available, we take the next interior step according to the updating formula given by

xk+1=xk+tk"k; 0¡ 61:

Line search: the line search is executed along the directiontk"k by varying thevalue.

3.3. Algorithm

In this section, the major steps of the algorithm are presented

Step1:k= 0.

Initialization: x0; Ax0¿ b and select a .nite 0i6minx∈XcTix; 16i6l

Step2: Iterationk:

(a) Determine the tradeoBs vectors!katxk. (b) Evaluate

ik+1=ki + !k1i (mk)i=l

i=1 !k1i(cTixkki);

16i6l;

k=!k(xk) +( ˙xkUxk!k(xk));

06 ¡1:

(c) Solve the linear system:

Hk"k=F(xk):

(d) Determinetk: (tk)2("k)THk"k61:

(e) Find the new iteratexk+1 from:

xk+1=xk+tk"k; 0¡ 61:

Step3: If the DM acceptxk+1 as the compromise solution, stop.

Otherwise increment the iteration counter:k:=k+1, go to step 2.

(6)

4. Illustrative example

The algorithm proposed in this paper will be illus- trated using the following numerical example [4]. We have chosen this example because the approach will yield a .nal interior point better than that found in A.

Arbel approach max x1

max x2

s:t: x1+ 5x2664;

x1+x2620;

4x1+x2656;

2x1x2622;

x1x2610;

x1+ 4x2¿8;

6x1+x2¿6;

−3x1+ 2x2616;

−x1+ 2x2620;

x1¿0;

x2¿0:

For this example, an initial point and the initialization ofki are available through

x0= [2;5]; 01=02= 0:

Table 1

Solution results with= 0:1;(= 1)

k x1 x2 !k (normalized) u(x)

0 0.714286 0.285714 2., 5. 10.

1 0.687195 0.312805 2.46878, 5.4236 13.3896 2 0.662035 0.337965 2.97453, 5.82676 17.3319 3 0.638485 0.361515 3.51675, 6.21107 21.8428 4 0.616244 0.383756 4.09585, 6.57722 26.9394 5 0.595118 0.404882 4.71206, 6.92603 32.6359 6 0.575029 0.424971 5.36462, 7.25887 38.9411 7 0.556005 0.443995 6.05124, 7.57783 45.8552 8 0.538168 0.461832 6.7672, 7.88575 53.3644 9 0.521732 0.478268 7.50415, 8.18611 61.4298 10 0.507032 0.492968 8.2476, 8.48289 69.9635 11 0.494588 0.505412 8.97196, 8.7798 78.772 12 0.485236 0.514764 9.62986, 9.07746 87.4147 13 0.480261 0.519739 10.1315, 9.36195 94.8506 14 0.482106 0.517894 10.317, 9.6041 99.0859

Fig. 3.

Table 2

Solution results when varying

kmax x1 x2 u(x)

0.1 14 10.317, 9.6041 99.0859

0.2 16 10.297, 9.59045 99.7528

0.3 20 10.1945, 9.76225 99.5215

0.4 27 10.1327, 9.82124 99.5156

0.5 48 10.0269, 9.97138 99.9821

Fig. 4.

(7)

Fig. 5.

Table 3

Solution results by other testing example (= 0:1)

Example kmax x1 x2 u(x)

Arbel et al. [2] 8 7.112 6.217 80.207

Stueur [13] (p. 377) 21 5.00279 4.99688 1295.18

Arbel [3] 6 4.87026 5.06697 24.68

Sherali [11] 10 3.56098 4.50153 −12:82

To simulate the DM’s choice behavior, we will assume the following implicit preference function in order to test our algorithm:

U(x) =x1x2:

The vector optimization problem has a unique solution given through

x?= [10;10] and U?= 100:

Following the steps outlined above, we generate a se- quence of steps summarized in Table 1 and illustrated by Fig. 3.

We mentioned earlier that the value of the scalar has some eBect on the procedure, this eBect on the progress of the iterative process when varying, , is illustrated in Table 2. The utility values at the current iterate are shown in Fig. 4 below, and in Fig. 5 we il- lustrate how a sequenceX0; X1; : : : ; Xk; : : :is smaller and smaller polytopes which shrink towards the com- promise solution.

In addition to the illustrative example, the problems presented by Arbel [2,3], Stueur [13, p. 377] and Sher- ali [11] were solved for comparison purposes. The re- sults are given in Table 3.

5. Conclusion

In this paper, we have presented an interactive inte- rior point algorithm for multiobjective linear program- ming problems. The algorithm is based on the modi- .cation of the method of centers to MOLP problems, this modi.cation is based on .nding an approximation to the gradient of the implicitly-known utility func- tion by interacting with the DM. The marginal rates of substitution are obtained and used in deriving the ap- proximation gradient and to update the lower bounds of the objectives.

6. For further reading See Refs. [9,12]

References

[1] B. Aghezzaf, T. Ouaderhman, An interactive interior point algorithm for multiobjective linear programming problems, International Conference on Applied Mathematics and Engineering Sciences Proceedings, Vol. 1, Casa, Morocco, 1998, pp. 1–7.

(8)

[2] A. Arbel, An interior multiobjective primal-dual linear programming algorithm using approximated gradients and sequential generation of anchor points, Optimization 30 (1994) 137–150.

[3] A. Arbel, Anchoring points and cones opportunities in interior multiobjective linear programming, J. Oper. Res. Soc. 45 (1) (1994) 83–96.

[4] A. Arbel, P. Korhonen, Using asperation levels in an interactive interior multiobjective linear programming, J.

Multi-Criteria Decision Anal. 89 (1) (1996) 193–201.

[5] A. Arbel, S.S. Oren, Generating interior search directions for multiobjective linear programming, J. Multi-Criteria Decision Anal. 2 (1993) 73–86.

[6] A. Arbel, S.S. Oren, Using approximate gradients in developing an interactive interior primal-dual multiobjective linear programming algorithm, Eur. J. Oper. Res. 89 (1996) 202–211.

[7] V. Chankong, Y.Y. Haimes, Multiobjective Decision Making:

Theory and Methodology, North-Holland, Amsterdam, 1983.

[8] A.M. GeoBrion, J.S. Dyer, A. Freinberg, An interactive approach for multicriterion optimization with an application to the operation of an academic department, Management Sci. 19 (1972) 357–386.

[9] C.C. Gonzaga, Path following methods for linear programming, SIAM Rev. 34 (34) (1992) 167–224.

[10] J. Renegar, Apolynomial-time algorithm, based on Newton’s method for linear programming. Report 07118-86, Math.

Sciences Research Institute, University of California, Berkeley, Cal., 1986.

[11] H.D. Sherali, G.V. Loganathan, Aconvergent interactive cutting-plane algorithm for multiobjective optimization, Oper.

Res. 35 (3) (1987) 365–377.

[12] W.S. Shin, A. Ravindron, An interactive method for multiobjective mathematical programming problems, J.

Optim. Theory Appl. 68 (3) (1991).

[13] R.E. Stueur, Multiple Criteria Optimization Theory:

Computation and Application, Wiley, New York, 1989.

[14] T. Trafalis, T.L. Morin, S.S. Abhyankar, EEcient faces of polytopes: interior point algorithms, parametrization of algebraic varieties, and multiple objective optimization, Contemp. Math. 114 (1990) 319–341.

[15] P.M. Vaidya, An algorithm for linear programming which requiresO(((m+n)n2+ (m+n)1:5nL)), Math. Programming 47 (1990) 175–201.

Références

Documents relatifs

To turn the path planning problem into an optimization program amenable to interior point algorithms, it is needed to define first a usable criterion and second a mean of getting

Furthermore, some effi- cient interior-point algorithms have been proposed to CQSDO, such as Nie and Yuan in [13,14] proposed two algorithms for solving CQSDO problems, which

The duality schemes present also similarities, one main difference is that for the strong duality result in linear semidefinite programming the primal and the dual problems have to

Les résultats des études réalisées avec les AVK, et le profil pharmacocinétique des NACO semblent indiquer que, chez les patients qui présentent un faible risque thromboembo-

Plusieurs spécialistes dans les apports théoriques sont persuadés que les voyages, les excursions dans d’autres pays ou en dehors de lieux urbains occasionnant une rupture

The proposed algorithm has been applied on a real problem for locating stations of a Carsharing service for electrical vehicles.. The good results we obtained in comparison to PLS

Then, in Section (3.3) a sufficient condition on the control penalty γ u guaranteeing that any optimal solution of POCP (7) strictly satisfies the input constraints is exhibited..

Algorithm 2, with the sequence of update param- eter θ k described above, guarantees strict feasibility of the iterates and quadratic convergence of the proximity measure.. As