Computer models of physical processes have become important tools in all areas of science. Two properties shared by many large-scale models are the requirement of considerable computer time for each run and the dependence on a large number of input variables. Computational experiments are often performed using such models, with the aim of creating an approximation of the model, or simply to discover which inputs have the greatest influence on outputs. In many cases, most inputs are unimportant; this phenomenon is called "effect sparsity" in experiments of smaller scale on physical systems.
The subject of this talk is a sequential design and analysis procedure, motivated by ideas from the statistical literature on computer experiments and group screening, to (1) identify the important inputs, and (2) produce a good experimental design for model approximation. The technique is demonstrated using a computer model of flow of a contaminant through an ecosystem.
KEY WORDS: Computer Experiments, Experimental Design, Group Screening, Stochastic Process