Balancing Optimal Large Deviations in Sequential Selection

Abstract

In the ranking and selection problem, a sampling budget is allocated among a finite number of designs with the goal of efficiently identifying the best. Allocations of this budget may be static (with no dependence on the random values of the samples) or adaptive (decisions are made based on the results of previous decisions). A popular methodological strategy in the simulation literature is to first characterize optimal static allocations by using large deviations theory to derive a set of optimality conditions, and then to use these conditions to guide the design of adaptive allocations. We propose a new methodology that can be guaranteed to adaptively learn the solution to these optimality conditions in a computationally efficient manner, without any tunable parameters, and under a wide variety of parametric sampling distributions. This paper was accepted by Baris Ata, stochastic models and simulation.