A sequential experiment design for input uncertainty quantification in stochastic simulation

When we use simulations to estimate the performance of a stochastic system, simulations are often driven by input distributions that are estimated from real-world data. There is both input and simulation uncertainty in the performance estimates. Non-parametric sampling approaches, e.g., the bootstrap, could be used to generate samples of input distributions quantifying both input model and parameter uncertainty. In this paper, a sequential experiment design is proposed to efficiently propagate the input uncertainty to output mean and deliver a percentile confidence interval to quantify the impact of input uncertainty on the system performance. Compared to the classical equal allocation, it could assign more computational budget to samples of input distributions that contribute most to the percentile confidence interval estimation. Our approach is supported by rigorous theoretical and empirical study.