Regression and Kriging metamodels with their experimental designs in simulation: A review

This article reviews the design and analysis of simulation experiments. It focusses on analysis via two types of metamodel (surrogate. emulator); namely, low-order polynomial regression, and Kriging (or Gaussian process). The metamodel type determines the design of the simulation experiment, which determines the input combinations of the simulation model. For example, a first-order polynomial regression metamodel should use a “resolution-III”design, whereas Kriging may use “Latin hypercube sampling”. More generally, polynomials of first or second order may use resolution III, IV, V, or “central composite” designs. Before applying either regression or Kriging metamodeling, the many inputs of a realistic simulation model can be screened via “sequential bifurcation”. Optimization of the simulated system may use either a sequence of low-order polynomials—known as “response surface methodology” (RSM)—or Kriging models fitted through sequential designs—including “efficient global optimization” (EGO). Finally, “robust”optimization accounts for uncertainty in some simulation inputs.