Sensitivity analysis of simulation models: an overview

Abstract

This Invited Lecture covers classic and modern designs, and their metamodels. Classic resolution-III designs including fractional-factorial two-level designs assume first-order polynomial metamodels. Resolution-IV and resolution-V designs assume such polynomials augmented with two-factor interactions. Central Composite Designs (CCDs) assume second-degree polynomials. Simulations with many factors require modern factor-screening designs; e.g., sequential bifurcation and Morris's designs. Globally fitted Kriging metamodels require space-filling designs; e.g., Latin Hypercube Sampling (LHS). Sensitivity analysis also serves optimization of the simulated system. Classic Response Surface Methodology (RSM) is popular. Novel methods select one of the multiple simulation outputs as goal variable, and satisfy given constraints on the other outputs and the inputs. Taguchian robust optimization allows for uncertainties in some (environmental) inputs.