Cost Minimization in Networks with Discrete Stochastic Requirements

Multistage minimum-cost network-flow analysis solves many practical problems in production-inventory-distribution, marketing, personnel, and finance. Unlike previous network papers, which generally restricted themselves to a deterministic situation, this paper investigates the stochastic environment. Starting from the standard multistage network-flow problem, we create a stochastic network by permitting the node requirements to be discrete random variables with known conditional probability distributions. Our goal is to determine the minimum-expected-cost flow and thereby solve the problem. Although linear programming under uncertainty can determine this flow, it would ignore the special structure of network-flow problems that allows development of computationally efficient algorithms. In this paper, we instead exploit the underlying network structure to produce both a new structure that is not a network but maintains many of the properties of a network, and a new node that replicates flows instead of conserving them. The new nodes, called replication nodes, together with the new structure, allow the development of an efficient computational algorithm that is capable of solving problems much larger than those solvable by linear programming under uncertainty.