Solving Stochastic Programming Problems by
Successive Regression Approximations
Numerical Results*
Istvan Deak
1
Operations Research Group, Dept. of Differential Equations, Budapest University
of Technology and Economics, H-llll Budapest, XI. Muegyetem rkp. 3.
deak@math.bme.hu
Summary. Recently a heuristic procedure has been developed for solving two-
stage stochastic programming problems. Using the same ideas solution algorithms
are presented here for solving probabilistic constrained problems and a model fam-
ily, combining probabilistic constraint and recourse, proposed by Pnlkopa. The
common feature of these algorithms is that all rely on replacing the "hard" part of
the problem by an easy-to-compute regression function. Numerical examples and
test results are also given.
1 Introduction.
Regression (or in other words least squares approximation) is a two-hundred
years old tool for approximating problems, effectively reducing measurement
errors [1], [12], [14], [25], [28]. In a series of recent papers [4], [6] we proposed to
use regression functions for approximation in a one-dimensional subproblem
of probability constrained, STABIL-type models [23], [24]. It turned out, that
successive regression approximations can be used for solving one-dimensional
nonlinear equations, with deterministic function values [7], [8].
Based on these results, anticipating the n-dimensional analogies of the
one-dimensional techniques proved, a heuristic method has been developed
for solving two-stage problems [10], [11]. Here two additional algorithms are
presented for solving stochastic programming problems of probabilistic con-
strained and a combined model type. The main contribution of this work is
thought to be the demonstration of the fact, that the same algorithm can be
used to solve these problems.
Two large model-families dominate the area of stochastic programming;
one of them is called the probabilistic constrained models, the other one is the
two-stage or recourse type models. There exists several solution techniques
for both type, but a method designed for one family can not be used to solve
a problem in the other one. The proposed heuristic method - according to
* Submitted for publication to Proceedings, IIASA, Laxenburg conference on
Stochastic Programming, based on lectures at EURO-2000 Conference on Oper-
ations Research, Budapest and IIASA conference on Stochastic Programming,
2002, Laxenburg
K. Marti et al. (eds.), Dynamic Stochastic Optimization
© Springer-Verlag Berlin Heidelberg 2004