Predicting the simulation budget in ranking and selection procedures

Abstract

The goal of ranking and selection (R&S) procedures is to identify the best among a finite set of alternative systems evaluated by stochastic simulation, providing a probability guarantee on the quality of the solution. To solve large-scale R&S problems, especially in parallel computing platforms where variable numbers of cores might be used, it is helpful to be able to predict the simulation budget, which is almost always the dominant portion of the running time of a given procedure for a given problem. Non-trivial issues arise due to the need to estimate the system configuration. We propose a set of methods for predicting the simulation budget. Numerical results compare our predictions for several leading R&S procedures.