Kriging interpolation in simulation: a survey

Many simulation experiments require much computer time, so they necessitate interpolation for sensitivity analysis and optimization. The interpolating functions are ´metamodels´ (or ´response surfaces´) of the underlying simulation models. Classic methods combine low-order polynomial regression analysis with fractional factorial designs. Modern Kriging provides ´exact´ interpolation, i.e., predicted output values at inputs already observed equal the simulated output values. Such interpolation is attractive in deterministic simulation, and is often applied in computer aided engineering. In discrete-event simulation, however, Kriging has just started. Methodologically, a Kriging metamodel covers the whole experimental area; i.e., it is global (not local). Kriging often gives better global predictions than regression analysis. Technically, Kriging gives more weight to ´neighboring´ observations. To estimate the Kriging metamodel, space filling designs are used; for example, latin hypercube sampling (LHS). This paper also presents novel, customized (application driven) sequential designs based on cross-validation and bootstrapping.