No label defined (Q4035542)

[For the entire collection see Zbl 0753.00017.] The paper demonstrates how a common class of stochastic models, namely the large-scale stochastic linear programs with recourse can be efficiently solved by using a blending of the classical Benders decomposition with a relatively new technique called importance sampling. A stochastic linear program breaks down into two parts: a deterministic dynamic part (master-problem) and a stochastic part. The master-problem is a dynamic linear program with \(T\) stages. Each stage is associated with a stochastic sub-problem. The sub-problems in each stage are assumed to be stochastically independent. The sub-problem costs \(f_ t\) and the technology matrix \(D_ t\) (\(t =1,\dots,T\)) are assumed to be deterministic parameters. The master-problem is optimized to obtain an approximate optimal feasible solution that is used as input to the sub- problems. Lower and upper confidence bounds to the problem are computed. Monte Carlo methods are recommended to compute multiple integrals or multiple sums arising in large-scale stochastic problems. The code of the subroutine MINOS has been adapted for solving both the master-problem and the sub-problems. Numerical results of large-scale problems in the area of expansion planning of power systems and financial planning are presented. These numerical results of large problems with numerous stochastic parameters indicate that very accurate solutions of the problems can be obtained using only small sample sizes.