METAMODEL-ASSISTED RISK ANALYSIS FOR STOCHASTIC SIMULATION WITH INPUT UNCERTAINTY

For complex stochastic systems, simulation can be used to study the system inherent risk behaviors characterized by a sequence of percentiles. In this paper, we develop a Bayesian framework to quantify the overall estimation uncertainty of percentile responses. Suppose that the input parametric families are known. The input model estimation uncertainty is quantified by posterior samples of input parameters. Then, a distributional metamodel is introduced to simultaneously model the percentile response surfaces, which can efficiently propagate the input uncertainty to outputs. Our Bayesian framework can deliver credible intervals for percentiles, and a variance decomposition is further derived to estimate the contributions of input and simulation uncertainties. The empirical studies indicate that our approach has promising performance for system risk analysis.