Parametric and distribution-free bootstrapping in robust simulation-optimization

Most methods in simulation-optimization assume known environments, whereas this research accounts for uncertain environments combining Taguchi’s world view with either regression or Kriging (also called Gaussian Process) metamodels (emulators, response surfaces, surrogates). These metamodels are combined with Non-Linear Mathematical Programming (NLMP) to find robust solutions. Varying the constraint values in this NLMP gives an estimated Pareto frontier. To account for the variability of this estimated Pareto frontier, this contribution considers different bootstrap methods to obtain confidence regions for a given solution. This methodology is illustrated through some case studies selected from the literature.