Combining simulation allocation and optimal splitting for rare-event simulation optimization

This paper presents research toward generalizing the optimization of the allocation of simulation replications to an arbitrary number of designs, when the problem is to maximize the Probability of Correct Selection among designs, the best design being the one with the smallest probability of a rare event. The simulation technique within each design is an optimized version of the splitting method. An earlier work solved this problem for the special case of two designs. In this paper an alternative two-stage approach is examined in which, at the first stage, allocations are made to the designs by a modified version of the Optimal Computing Budget Allocation. At the second stage the allocation among the splitting levels within each design is optimized. Our approach is shown to work well on a two-tandem queuing model.