Balancing Optimal Large Deviations in Ranking and Selection

The ranking and selection problem deals with the optimal allocation of a simulation budget to efficiently identify the best among a finite set of unknown values. The large deviations approach to this problem provides very strong performance guarantees for static (non-adaptive) budget allocations. Using this approach, one can describe the optimal static allocation with a set of highly nonlinear, distribution-dependent optimality conditions whose solution depends on the unknown parameters of the output distribution. We propose a new methodology that provably learns this solution (asymptotically) and is very computationally efficient, has no tunable parameters, and works for a wide variety of output distributions.