Problem-based optimal scenario generation and reduction in stochastic programming (Q2118075): Difference between revisions

In this paper the authors study a problem-based approach to scenario generation and reduction for stochastic programming models without information constraints. The generation of scenarios is an important issue for solving applied stochastic programming models. Presently Monte Carlo sampling methods are the preferred approach, but besides Quasi-Monte Carlo and sparse grid methods also best approximation methods are in use. The authors show that the optimal scenario generation problem can be formulated as generalized semi-infinite program (Theorem 1) which is convex in some cases (Theorem 2),enjoys stability (Theorem 3) and allows a transformation into a standard semi-infinite program in a number of cases. Also, the authors revisit the problem of optimal scenario reduction for two-stage models and provide a new formulation as mixed-integer linear semi-infinite program. In the last part, the authors present a mixed-integer linear semi-infinite program for optimal scenario generation in chance constrained programming. In the ``Appendix'' the authors provide a short description of the discretization method due to \textit{R. Reemtsen} [J. Optim. Theory Appl. 71, No. 1, 85--103 (1991; Zbl 0793.90088)] Finally, the authors illustrate the approach to scenario generation for the classical newsvendor problems with random demand.