Statistical fitting and validation of non-linear simulation metamodels: A case study

Linear regression metamodels have been widely used to explain the behavior of computer simulation models, although they do not always provide a good global fit to smooth response functions of arbitrary shape. In the case study discussed in this paper, the use of several linear regression polynomials results in a poor fit. The use of a non-linear regression metamodeling methodology provides simple functions that adequately approximate the behavior of the target simulation model. The importance of metamodel validation is emphasized by using the generalization of Rao's test to non-linear metamodels and double cross-validation.