Efficient Dynamic Simulation Allocation in Ordinal Optimization

Abstract

Ordinal optimization has emerged as an efficient technique for simulation optimization. A good allocation of simulation samples across designs can further dramatically improve the efficiency of ordinal optimization. We investigate the efficiency gains of using dynamic simulation allocation for ordinal optimization by comparing the sequential version of the optimal computing budget allocation (OCBA) method with optimal static and one-step look-ahead dynamic allocation schemes with "perfect information" on the sampling distribution. Computational results indicate that this sequential version of OCBA, which is based on estimated performance, can easily outperform the optimal static allocation derived using the true sampling distribution. These results imply that the advantage of sequential allocation often outweighs having accurate estimates of the means and variances in determining a good simulation budget allocation. Furthermore, the performance of the perfect information dynamic scheme can be viewed as an approximate upper bound on the performance of different sequential schemes, thus providing a target for further achievable efficiency improvements using dynamic allocations