부산대학교 도서관

This tutorial reviews methodologies for quantifying statistical uncertainty in computationally expensive black-box models, which arise frequently in data-driven simulation analyses under input uncertainty. When facing these models, it can be difficult to run repeated evaluations due to computation cost, and also to obtain auxiliary information such as gradients due to analytical intractability, thus rendering many traditional statistical approaches challenging to apply. We describe several lines of approaches to resolve these challenges, including data-splitting methods based on batching variants, a recent so-called cheap bootstrap approach, and subsampling schemes. We discuss the applications of these approaches to simulation, including problems suffering from both aleatory error exhibited via Monte Carlo noises and epistemic error stemming from the input uncertainty.