Efficient computing budget allocation for stochastic simulation optimization

The growing interest in simulation has resulted in an imperative need and considerable attention for simulation optimization which targets to find the optimal solution based on the outputs from stochastic simulation. Simulation optimization is much more complicated than deterministic optimization because of randomness incurred from stochastic simulation. How to improve the efficiency is a central issue in simulation optimization. This dissertation attempts to improve the efficiency in simulation optimization by optimally allocating computing budget. Simulation optimization can be categorized in two settings: (1) fixed number of alternative systems, (2) unbounded/uncountable number of alternatives. In the setting of fixed number of alternatives, we propose a computing budget allocation method to efficiently select the single best with respect to the expected opportunity cost, whereas most of existing selection methods focuses on maximizing probability of correct selection. Furthermore, we propose an efficient easy-to-implement heuristic sequential allocation procedure to select the best subset instead of the single best. In the setting of unbounded/uncountable number of alternatives, we enhance the efficiency of population-based global optimization methods by more accurately selecting the elite samples with our best subset selection method. Moreover, we propose a customized allocation algorithm which improves the update of the sampling distribution in the Cross-Entropy (CE) method, so that the efficiency of both standard and extended version of CE method is significantly improved. Numerical experiments show the proposed algorithms significantly improve the efficiency of simulation optimization in both settings.