In this thesis, we introduce a new approach to rare event simulation. Because of the extensive simulation required for precise estimation of performance criteria dependent on rare event occurrences, obstacles such as computing budget/time constraints can become prohibitive, particularly if comparative study of different system designs is involved. Existing methods for rare events simulation have focused on simulation budget reduction while attempting to generate accurate performance estimates. In this thesis, we propose a new approach for rare event system analysis in which we soften the simulation goal to the isolation of a set of good enough designs with high probability. Given this relaxation, referred to as ordinal optimization and advanced by Ho et. al. (1992), this thesis' methodology calls instead for the consideration of an appropriate surrogate design problem. This surrogate problem is characterized by its approximate design rank equivalence to the original problem and its performance criterion's dependence not on rare events, but on more frequent events. By evaluating this surrogate problem, a subset of designs is selected. We characterize the quality of this selected subset by alignment probability, the probability that at least some good enough designs are chosen. In order to generate alignment probability estimates, we introduce a strictly rank-based model and combine its implementation with universal alignment probability curves (Lau-Ho (1995)). Experimental results indicate that the use of an appropriate less rare surrogate problem together with this estimation procedure forms a practical solution approach for the stochastic optimization of rare event probability problems.
展开▼
