Robust simulation optimization with stratification

Stratification has been widely used as a variance reduction technique when estimating a simulation output, whereby the input variates are generated following a stratified sampling rule from previously determined strata. This study shows that an adaptive sampling class of simulation optimization solvers called ASTRO-DF could become more robust with stratification, S-ASTRO-DF. For a simulation optimization algorithm, we discuss how to monitor the robustness in terms of bias and variance of the outcome and introduce several metrics to compute and compare the robustness of solvers. We find that while stratified sampling improves the algorithm's performance, its robustness is sensitive to the stratification structure. In particular, as the number of strata increases, the stratified sampling-based algorithms may become less effective.