Accounting for input uncertainty in discrete-event simulation

The primary objectives of this research are formulation and evaluation of a Bayesian approach for selecting input models in discrete-event stochastic simulation. This approach takes into account the model, parameter, and stochastic uncertainties that are inherent in most simulation experiments in order to yield valid predictive inferences about the output quantities of interest. We use prior information to specify the prior plausibility of each candidate input model that adequately fits the data, and to construct prior distributions on the parameters of each model. We combine prior information with the likelihood function of the data to compute the posterior model probabilities and the posterior parameter distributions using Bayes' rule. This leads to a Bayesian Simulation Replication Algorithm in which: (a) we estimate the parameter uncertainty by sampling from the posterior distribution of each model's parameters on selected simulation runs; (b) we estimate the stochastic uncertainty by multiple independent replications of those selected runs; and (c) we estimate model uncertainty by weighting the results of (a) and (b) using the corresponding posterior model probabilities. We also construct a confidence interval on the posterior mean response from the output of the algorithm, and we develop a replication allocation procedure that optimally allocates simulation runs to input models so as to minimize the variance of the mean estimator subject to a budget constraint on computer time. To assess the performance of the algorithm, we propose some evaluation criteria that are reasonable within both the Bayesian and frequentist paradigms. An experimental performance evaluation demonstrates the advantages of the Bayesian approach versus conventional frequentist techniques.