Simulation Optimization Under Input Model Uncertainty

Abstract

Stochastic simulation is driven by the input model, which is a collection of distributions that model the randomness in the system. Since the input model is often estimated from data of past observations, simulation is subject to the so-called input model uncertainty due to the finiteness of the data. As a result, optimizing the corresponding simulation model may lead to solutions that perform poorly under the true model. In the past, simulation optimization has been mostly studied under the assumption that the input model is accurate, thus only accounting for stochastic uncertainty in the system but ignoring input model uncertainty. In this chapter, we aim to answer two questions: (1) how to quantify the impact of input model uncertainty on the simulation optimization problem; (2) how to “optimize” a simulation model when taking into account input model uncertainty. To address the first question, we provide asymptotic results and confidence intervals on the optimality gap and performance of solutions. For the second question, we review a recently proposed framework of Bayesian risk optimization that captures trade-off between the expected performance and the variability of the actual performance. Many research questions still remain for simulation optimization under input model uncertainty, and will be discussed briefly at the end of this chapter.