Kriging metamodeling for simulation

Many scientific disciplines use mathematical models to describe complicated real systems. Often, analytical methods are inadequate, so simulation is applied. This thesis focuses on computer intensive simulation experiments in Operations Research/Management Science. For such experiments it is necessary to apply interpolation. In this thesis, Kriging interpolation for random simulation is proposed and a novel type of Kriging - called Detrended Kriging - is developed. Kriging turns out to give better predictions in random simulation than classic low-order polynomial regression. Kriging is not sensitive to variance heterogeneity: i.e. Kriging is a robust method. Moreover, the thesis develops a novel method to select experimental designs for expensive simulation. This method is sequential, and accounts for the specific input/output function implied by the underlying simulation model. For deterministic simulation the designs are constructed through cross-validation and jackknifing, whereas for random simulation the customization is achieved through bootstrapping. The novel method simulates relatively more input combinations in the interesting parts of the input/output function, and gives better predictions than traditional Latin Hypercube Sample designs with prefixed sample sizes.