Content:

In the standard formulation of stochastic programs with recourse one assumes that the distribution of the random parameters doesn't depend on the decisions. In other words, the probability distribution is exogenous to the decision process. This allows for the design of solution procedures that are (relatively) efficient, often making it possible to exploit the convexity of the (overall) problem.

But for certain classes of stochastic programs this `independence' assumption is far from being satisfied. Think of an oil exploration problem where geological information will be gained from the drilling location(s), i.e., the probability of finding oil in a certain region is strongly affected by the location of the wells being drilled. Or, a production problem where `costs' can be much better assessed after production is started.

This paper identifies a class of problems that are `manageable' and proposes an algorithmic procedure for solving problems of this type. The proposed algorithm is closely related to the branch and bound techniques commonly used to solve integer programming problems. It is assumed that the variables controlling the information (= changes in the probability distributions of the random parameters) are discrete. Computational experience is reported.

`Directions of descent' identifying (the information controlling) variables that should be considered for inclusion in a potentially optimal solution are obtained by means of a variational analysis argument.

April 3, 1997