Approximations to stochastic dynamic programs via information relaxation duality
© 2019 INFORMS. In the analysis of complex stochastic dynamic programs, we often seek strong theoretical guarantees on the suboptimality of heuristic policies. One technique for obtaining performance bounds is perfect information analysis: this approach provides bounds on the performance of an optimal policy by considering a decision maker who has access to the outcomes of all future uncertainties before making decisions, that is, fully relaxed nonanticipativity constraints. A limitation of this approach is that in many problems perfect information about uncertainties is quite valuable, and thus, the resulting bound is weak. In this paper, we use an information relaxation duality approach, which includes a penalty that punishes violations of the nonanticipativity constraints, to derive stronger analytical bounds on the suboptimality of heuristic policies in stochastic dynamic programs that are too difficult to solve. The general framework we develop ties the heuristic policy and the performance bound together explicitly through the use of an approximate value function: heuristic policies are greedy with respect to this approximation, and penalties are also generated in a specific way using this approximation. We apply this approach to three challenging problems: stochastic knapsack problems, stochastic scheduling on parallel machines, and sequential search problems. In each of these problems, we consider a greedy heuristic policy generated by an approximate value function and a corresponding penalized perfect information bound. We then characterize the gap between the performance of the policy and the information relaxation bound in each problem; the results imply asymptotic optimality of the heuristic policy for specific “large” regimes of interest.
Volume / Issue
Start / End Page
Electronic International Standard Serial Number (EISSN)
International Standard Serial Number (ISSN)
Digital Object Identifier (DOI)