Metamodel-Assisted Sensitivity Analysis for Controlling the Impact of Input Uncertainty

Given finite real-world data, input models are estimated with error. Thus, the system performance estimation uncertainty includes both input and simulation uncertainties. Built on the global sensitivity analysis proposed by Oakley and O’Hagan, we develop a metamodel-assisted Bayesian framework to quantify the contributions from simulation and input uncertainties. It further estimates the impact from each source of input uncertainty and predicts the value of collecting additional input data, which could guide the data collection to efficiently improve the system response estimation accuracy. The empirical study demonstrates that our approach has promising performance.