Analyzing Computer Experiments: What Matters, What Doesn't
We compare strategies for the analysis of an initial computer experiment. A prior treating the real-valued output as a realization of a Gaussian stochastic process (GaSP) is fairly common now, but the user is faced with some operational decisions. First, the mean of the GaSP may be just a constant or could be a regression model in the input variables. Secondly, there are several possible families for the (assumed stationary) covariance function of the process. We take an evidence-based approach to evaluating such modelling options. While the main thrust of the talk is analysis, it turns out that design of an experiment has some important and perhaps surprising implications. Our findings and recommendations will be based on results from simple illustrative functions, real codes, and simulations.
Joint work with Jason Loeppky (UBC, Okanagan) and Jerome Sacks (NISS)