Some Explicit Results for the Distribution Problem of Stochastic Linear Programming

A technique is developed for finding a closed form expression for the cumulative distribution function of the maximum value of the objective function in a stochastic linear programming problem, where either the objective function coefficients or the right hand side coefficients are continuous random vectors with known probability distributions. This is the “wait and see” problem of stochastic linear programming. Explicit results for the distribution problem are extremely difficult to obtain; indeed, previous results are known only if the right hand side coefficients have an exponential distribution [1]. To date, no explicit results have been obtained for stochastic c, and no new results of any form have appeared since the 1970’s. In this paper, we obtain the first results for stochastic c, and new explicit results if b an c are stochastic vectors with an exponential, gamma, uniform, or triangle distribution. A transformation is utilized that greatly reduces computational time.


Introduction
Consider the linear programming problem, ( ) . : ,  z x by first observing the vector b or the vector c and then solving (1)-( 3).This paper is interested in finding explicit expressions for the distribution of ( ) max z x if either b or c is random.This is called the distribution problem of stochastic linear programming.
Early work on the distribution problem can be found in Babbar [2], Bereanu [3] [4] [5] [6] [7], Hsia [8], Prekopa [9], Sengupta, Tintner, and Millham [10], Sengupta, Tintner, and Morrison [11], and Wets [12].For additional references, see the bibliographies by Stancu-Minasian [13] and Van Der Vlerk [14].Application of the distribution problem can be found in the areas of agriculture [15] and economic planning [10], [11].Explicit results for the distribution of ( ) max z x are very difficult to obtain; in- deed, most analyses rely on approximation techniques or simulation.(See, for example, Bracken and Soland [16], Sarper [15], or Dempster [17]).Bereanu [3] discovered that under certain assumptions, the sample space of the random coefficients allows a partition into non-overlapping sets, called decision regions, such that a basis of the linear programming problem can be assigned to each of the sets, and this basis remains optimal for all of its sample points.Ewbank, et al. [1] extended this theory using a Jacobian transformation to simplify the computational analysis.To date, we believe that an explicit expression for the distribution of ( ) max z x has only been obtained for sto- chastic b [1], and no explicit results have been obtained for stochastic c.In addition, no explicit results have been obtained for non-exponential distributions.In this paper, we obtain new explicit results for exponential, uniform, gamma, and triangle distributions with b or c random.These are the first explicit results for the case in which c is random.

Theory
Following [1], consider the linear programming problem (1)-(3).Let ( ) be the vector of basic variables corresponding to the ith basis, and i B is the m m × basis matrix whose columns are the columns of ( ) be the vector of coefficients of the basic variables in th i basis and let j a be the th j column of ( ) and is feasible if For the case in which the b vector is random, let the probability space be defined by the m-tuple ( ) . Bereanu discovered that there exist non-overlapping regions where Thus,

{ }
Pr Pr Now, consider the case in which only the c vector is random.Let the probability space C be defined by the n-tuple ( ) [3] found that the space C is partitioned by the sets: where i refers to the th i basis.Further the set of points is of probability measure zero if the joint density function of i is continuous.Points in this set are such that alternate optimal basis give the same value of ( ) Thus, where ∅ is an arbitrary constant.
To evaluate the right-hand side of equation Equation (15) let where Thus, if only the c vector is random the distribution function of ( ) max z x can be found, in theory, by evaluating the integral in equation Equation (20).Given a basis Bi and sets i U and i V , the limits of the integral in Equation ( 12) and Equation (20) are the intersection of m or n hyperplanes (depending on whether the b vector or the c vector is stochastic).These limits are extremely difficult to obtain if the probability space has dimension greater than 3. Ewbank, et al., [1] developed a Jacobian transformation that greatly simplifies the computation of the integrals.
In the case of stochastic b, Let By substituting for b we have: The probability that a basis G remains feasible is where s is the set of b's defined in Equation ( 20), and by substituting Equation (21) in Equation ( 22), we have: where r J is the Jacobian Note that since  is the basis matrix, its determinant is nonzero; thus r J is also nonzero.

Computational Results
The problems were run using the Mathematica software version 8.0.1.0utilizing the supercomputer at the University of Oklahoma. CPUs: All compute nodes have dual Intel Xeon E5-2650 "Sandy Bridge" oct core 2.0 GHz CPUs; there is also one "fat node" with quad Intel Xeon E7-4830 "Westmere" oct core 2.13 GHz CPUs. RAM: Most of the compute nodes have 32 GB of 1333 MHz RAM and 23 with 64 GB of 1333 MHz RAM; the one "fat node" has 1 TB of 1066 MHz RAM, which is called large memory. Accelerators: There are 18 NVIDIA Tesla M2075 cards, for an aggregate of an additional approximately 9 TFLOPs double precision.
In order to compare the run times, four types of distributions were considered as shown in Table 1.The coefficients were randomly generated in small interval, because large intervals led to computational results that had results with coefficients of the orders of 1020 or larger.

Case
Case I. Stochastic Resource Vector.Ewbank's model ) )

Problem 1
Case Case I. Stochastic Resource Vector.Ewbank's model

Computational Time Comparisons
The different distributions were solved using both Bereanu's method and the Ewbank, Foote and Kumin transformation method to compare the two.Table 2 and Table 3 compare the run times for both methods for case I and case II.The results show that the EFK method substantially reduces the computational time.In addition, Bereanu's method is not able to solve some larger sizes of the problem.All times are measured in seconds.
How to cite this paper: Ansaripour, A., Mata, A., Nourazari, S. and Kumin, H. (2016) Some Explicit Results for the Distribution Problem of Stochastic Linear Programming.where c is an vector.Further assume that b and c are random vectors with joint density functions ( ) f b and ( ) g c re- spectively.Next, consider the value of ( )

Table 1 .
Equations of distribution.

. Results for Stochastic c with Triangle Distribution 3.8.1. Problem 1
5

Table 2 .
Comparison between Bereanu and EFK method for case I.

Table 3 .
Comparison between Bereanu and EFK method for case II.