A trust region-based algorithm for continuous optimization via simulation

Abstract

Continuous Optimization via Simulation (COvS) involves the search for specific continuous input parameters to a stochastic simulation that yield optimal performance measures. Typically, these performance measures can only be evaluated through simulation. We introduce a new algorithm for solving COvS problems. The main idea is to use a nonparametric regression model that uses few samples, and embed it in an iterative trust-region framework. We name the proposed algorithm Simulation Optimization--Learning Via Trust Regions (SO-LViT). We discuss the algorithmic elements of this implementation, and hypothesize that this approach is especially suitable for situations where samples are expensive to obtain and the dimensionality of the problem is fairly large. We demonstrate promising results through computational experience, wherein we compare SO-LViT against several other approaches over a large test set under Gaussian noise conditions.