Chance-Constrained Approaches for Multiobjective Stochastic Linear Programming Problems

Multiple objective stochastic linear programming is a relevant topic. As a matter of fact, many practical problems ranging from portfolio selection to water resource management may be cast into this framework. Severe limitations on objectivity are encountered in this field because of the simultaneous presence of randomness and conflicting goals. In such a turbulent environment, the mainstay of rational choice cannot hold and it is virtually impossible to provide a truly scientific foundation for an optimal decision. In this paper, we resort to the bounded rationality principle to introduce satisfying solution for multiobjective stochastic linear programming problems. These solutions that are based on the chance-constrained paradigm are characterized under the assumption of normality of involved random variables. Ways for singling out such solutions are also discussed and a numerical example provided for the sake of illustration.


Introduction
Decision makers are intendedly rational.They are goal directed and they are intended to pursue those goals in conformity with the classic homo-economicus model [1].They do not always succeed because of the complexity of some environments that impose both procedural and substantive limits.This is the case when dealing with multiobjective stochastic linear programming (MSLP) problems.
In such a turbulent environment the notion of "optimum optimorum" is not clearly defined and satisficing rather than optimal search behavior seems to be the most appropriate option.A look at the literature reveals that most existing solution concepts for MSLP problems rely heavily on the expected, pessimistic and optimistic values of involved random variables [2,3].A finding from several research work [4][5][6][7] leave no doubt about the fact that these values may be useful in providing the range of possible outcomes.Nevertheless, they ignore such important factors as the size and the probability of deviations outside the likely range as well as other aspects concerning the dispersion of the probabilities.
In this paper, we discuss satisfying solution concepts for MSLP problems.These concepts are based on the chance-constrained philosophy [8].Chance-constrained applied for the purpose of limiting the probability that constraint will be violated.In this form, it adds consid-erably to both the flexibility and reality of the stochastic model under consideration.Mathematical characterization of the above mentioned solution concepts are also provided along with ways for singling them out.A numerical example is supplemented for the sake of illustration.
The remainder of the paper is organized as follows.In the next section we clearly formulate the problem under consideration.In Section 3, we present some satisfying solution concepts for this problem.In Section 4, we characterize these solution concepts.In Section 5, we discuss ways for solving resulting problems.In Section 6, we present the methods for solving this problem.An illustrative example is given in Section 7. We end up with some concluding remarks along with lines for further developments in the field of MSLP.

Multiobjective Stochastic Linear Programming Problem Formulation
A multiobjective stochastic linear programming is a problem of the following type:


 are random variables defined on the same probability space.(P1) is an ill-defined problem because of the simultaneous presence of different objective functions and of randomness surrounding data.In such a turbulent environment, neither the notion of feasibility nor that of optimality is clearly defined.One then has to resort to the Simon's bounded rationality principle [9] and seek for a satisfying solution instead of an optimal one.
Several solution concepts have been described in the literature, see for instance [10,11].Most of these solution concepts rely on the first two moments of involved random variables.As mentioned earlier, these values often offer a short-sighted view of uncertainty surrounding data under consideration.In what follows, we discuss some satisfying solution concepts based on the chanceconstrained approach [12][13][14].

Satisfying Solutions for Problem (P1)
-satisfying solution of type 1 Assume that objective functions of (P1) are aggregated in the form of .We say that x R  is an is optimal for the following optimization problem: and β are probability levels a priori fixed by the Decision Maker.
, is efficient for the following multiobjective optimization problem:  , where

Interpretation
Intuitive ways of thinking about the above solution concepts are given below.
is minimized in the Pareto sense.In the meantime, the chance of having the th i constraint of the problem be satisfied is greater than close to 1, such a solution is interesting from the stand points of feasibility and efficiency.
minimizes the weighted sum of derivations from given targets while maintaining the probability of each objective to be close to the corresponding target, to some desired levels.It also meets the requirement of being feasible with high probability.

Characterization of Satisfying Solutions
In what follows, we provide necessary and sufficient conditions for a potential action to be a satisfying one, in the above described sense.These conditions rely heavily on the hypothesis of normality of distributions of involved random variables.
We recall that random vector is said to have the multivariate normal distribution if any linear combination of its components is normally distributed.The following Lemma due to Kataoka [13] will be needed in the sequel.
are independent and normally distributed random variables with mean , the inner product is normally distributed with mean .
We are now in position to characterize an   Assume that objective functions of (P1) is aggregated are independent and normally distributed.Suppose also that for any , the random variables x y x y Indeed we have by Lemma is normally distributed with mean 0 and variance 1.
Therefore the first constraint of becomes var And we have established that the first constraint of is equivalent to relation (1).Define now the random variables  are independent and normally distributed [15].
and its variances is The chance constraints can therefore, be written as which can also be expressed as Thus the chance constraints (4) can be merely written: by their values given by ( 2) and (3) respectively in (4), we obtain 1) and ( 5) we obtain that is optimal to the following program: This problem is equivalent to (P2)' as desired.1, , 0

 
and the chance constraints: Then (P3) is equivalent to the following program: x is optimal for the following mathematical program:

Proof. Consider
As discussed in the proof of Proposition 4.2, the chance constraints As shown in the proof of Proposition 4.1, we have that can be written as follow: which is exactly (P4)'.

Singling out Satisfying Solution for Problem (P1)
Problem to be solved to single out a satisfying solution (in the sense of definitions given in, §2 for (P1) are (P2)', (P3)' and (P4)' depending on the type of solution chosen.(P2)' and (P4)' are standard mathematical programs about which a great deal is known.In the case, when these programs are convex, one may use algorithms of convex programming [16], the conjugate gradient method [17], the exterior or interior penalty functions methods [18].For the non convex case, one may use approximation type algorithms or metaheuristics [19].(P3)' is a multiobjective program.A rich array of methods for solving it may be found in [20].

Method for Yielding a Satisfying Solution of Type 1
Step 1: Fix Step 6: Write the Program ;   x is a satisfying solution of type1 for (P1)"; Step 9: Stop.

Method for Yielding a Satisfying Solution of Type 2
The same as 6.1; except that in Step 1  should be re- placed by Steps 5-7 should be as follow: Step 5: Write the Program ; Step 6: "Solve " and in;  3 P Step 7, " x is a satisfying solution of type 2"; Step 8: Stop.

Numerical Example
Consider the following multiobjective stochastic linear program: are independent normally distributed random variables, with the following parameters: solving this optimization problem using MATLAB, we obtain the following solution: 0, 0 using the global criterion method [21] and solving the resulting problem by MATLAB, we obtain the following solution: 11.95851, x x Suppose that a satisfying solution of type 3 is desirable for i 0.99   0.99 and k i    k  .Suppose further that the decision maker goals are 2,000 and 15,000 for objectives 1 and 2 respectively.Using equal weights (i.e.As can be seen, methods yielding type 2 and type 3 satisfying solutions have almost the same solution that differs from the method yielding type 1 solution satisfying solution.This is because the two first methods use the multiobjective approach while the third method uses the stochastic approach.

Concluding Remarks
In this paper, we have addressed the problems of defining and characterizing solution concepts for a multiob-jective stochastic linear program.Owing to the simultaneous presence of randomness and to the multiplicity of objective functions, we have privileged the satisfying approach rather than the optimizing one.
Solution concepts discussed in this paper have an advantage over those focusing only on the expectation and variances.As a matter of fact, these solutions take into consideration, the whole distributions of involved random variables.We have shown that when involved random variables are normally distributed, singling out solutions discussed here causes little extra computational costs.As a matter of fact, problems to be solved to obtain such solutions are standard mathematical programs about which a great deal is known.Lines for further developments in this field include:  The choice of threshold i  , 1, , i m   , and k  , 1, , K k  in an optimal way;  The consideration of other distributions rather than the normal one;  The extension of solution concepts discussed here to the case where fuzziness and randomness are in the state of affairs [22];  The generalization of methods discussed here to the bilevel case [23].

Proof
As shown in the proof of Proposition 4.1, the chance constraints in (P3) are equivalent to


And clearly, this problem is equivalent to (P3)' in the sense that they have the same efficient solutions.Finally, a characterization of an   ,   -satisfying solution of type 3 is given below.Proposition 4.3 Consider (P1) and assume that the random variables satisfying solution of type 3 for problem (P1) if and only if * satisfying solution of type 1 for   k  : while the vector