General-purpose ranking and selection for computer simulation

Abstract

ABSTRACT Many indifference-zone Ranking-and-Selection (R&S) procedures have been invented for choosing the best simulated system. To obtain the desired Probability of Correct Selection (PCS), existing procedures exploit knowledge about the particular combination of system performance measure (e.g., mean, probability, variance, quantile) and assumed output distribution (e.g., normal, exponential, Poisson). In this article, we take a step toward general-purpose R&S procedures that work for many types of performance measures and output distributions, including situations where different simulated alternatives have entirely different output distribution families. There are only two versions of our procedure: with and without the use of common random numbers. To obtain the required PCS we exploit intense computation via bootstrapping, and to mitigate the computational effort we create an adaptive sample-allocation scheme that guides the procedure to quickly reach the necessary sample size. We establish the asymptotic PCS of these procedures under very mild conditions and provide a finite-sample empirical evaluation of them as well.