Bayesian Simulation Optimization Using Augmented Probability Simulation

Under the subject expected utility paradigm, decisions are made by finding the alternative with the maximum expected utility. In Bayesian simulation, the probability distribution used is the distribution of a simulation output. While methods have been developed under the Bayesian paradigm for choosing the best simulated system from a discrete, finite set of alternatives, the only methods for optimization with Bayesian simulations are borrowed from classical simulation and deterministic optimization. This paper presents an algorithm for finding the decision with the maximum expected utility, based on a sampling technique, called augmented probability sampling. The algorithm is particularly appealing as both the decision and the simulation outputs are handled using sampling techniques derived from the Metropolis sampler, which is commonly used in Bayesian estimation. (Simulation optimization; Bayesian statistics; Markov chain Monte Carlo)