Using unreplicated 2^k-p designs for characterizing moderately dimensioned deterministic computer models

Deterministic computer simulation models (DCSMs) have been used extensively by two very different schools of researchers. One has produced highly dimensioned models (hundreds of parameters) of very complex physical phenomena and has explored various means of sensitivity analysis. The other school typically produces low-dimensioned models (dozens of parameters) and has very limited experience with sensitivity analysis. However, the sophisticated techniques used in the large models are not readily accessible to users of the smaller models. The approaches of both schools are reviewed before describing how classical DOE plans can be applied to sensitivity analysis of moderately dimensioned (a dozen to a couple hundred parameters) DCSMs. Because the usual analysis of classical DOE requires randomness, alternative means of experimental analysis are discussed. Four techniques for DCSM experimental analysis, including per cent contribution to variation, are described and the results are statistically compared.