Models, Experimental Designs, and Input Screening

• Activity Number: 242
• Type: Topic Contributed
• Date/Time: Tuesday, August 13, 2002 : 2:00 PM to 3:50 PM
• Sponsor: Section on Physical & Engineering Sciences*
• Author(s): Max Morris*+ and Leslie Moore
• Affiliation(s): Iowa State University and Los Alamos National Laboratory
• Keywords: Experimental Design ; Computer models ; Sensitivity Analysis ; Computer experiments ; Input screening

Abstract

Computer models of complicated processes often require the specification of values for a large number of input variables. For many realistic models, the specific values selected for many of these inputs make very little difference in the values of the computed model outputs. The early stages of examination of a computer model often include "sensitivity analyses" constructed to identify the (often few) inputs which are most important or influential in determining the output values. In this talk, we will examine a specific form of sensitivity analysis based on the use of randomized factorial designs and an ANOVA-like decomposition of output variance. In particular, we shall consider the costs and benefits associated with using combinations of fractional factorial designs in place of full factorial plans.