|
Abstract:
|
Computer experiments using deterministic simulators are often used to replace or supplement physical system experiments. As computational power and mathematical model complexity have increased, so has the sophistication of computer simulators. This talk compares several design criteria for initial computer experiments using their empirical prediction accuracy. One focus is design construction using the integrated mean squared prediction error (IMSPE), computed for a Gaussian process model with Gaussian correlation function. Specifically, the IMSPE-design criteria are (1) minimizing IMSPE with respect to a given set of correlation parameter values, and (2) minimizing a weighted IMSPE over a prior distribution of correlation values. The IMSPE-based designs are compared with widely-used space-filling designs. Additionally, prediction accuracy is studied for the design of an experiment that combines responses from a physical system and outputs from a simulator of the physical system. Empirical prediction accuracy is used to choose designs for both experiments when the goal is accurate calibrated Bayesian predictions of the mean of the physical system. Design recommendations are made.
|
