Reusing Simulation Outputs of Repeated Experiments Via Likelihood Ratio Regression

Abstract

Simulation experiments are sometimes conducted periodically, with updated parameters of the stochastic system being modeled. Storing and reusing the past simulation experiment data may be helpful for the current simulation experiment. In this paper, we consider reusing simulation data in repeated experiments to develop high-quality metamodels. Specifically, we propose a generalized least square regression metamodel whose input data include simulation outputs from the current and the past experiments. Moreover, the past simulation outputs are reused via the likelihood ratio method. Asymptotic variance analysis is provided to show the benefits of reusing past simulation data in prediction accuracy, and the numerical results show the effectiveness of the proposed method.