Factor Screening in Computer Simulation

Abstract

In many simulation models, only a few input variables have a significant effect on the output. By identi fying those variables in some reasonable way, we could undoubtedly make the model simpler, more effi cient, cheaper to run, and easier to analyze.

Factor screening methods attempt to identify the more important variables. The most effective methods, however, require more computer runs than are normally reasonable or affordable. Thus screening must usually be based on nonstandard designs not customarily dis cussed in standard statistical references.

In this paper, we review and discuss the major classes of factor screening designs. We pay partic ular attention to two strategies that are useful when there are more factors than available runs. The first of these is group-screening, in which fac tors are grouped and tested in a multistage proce dure ; the other is based on a combination of random balance and Plackett-Burman designs. Unfortunately, the performance of a screening strategy is difficult to measure; the modeler must consider both how much it costs and how accurately it classifies factors. An artificial example illustrates the tradeoffs that the modeler must make.