Generic Sequential Sampling for Metamodel Approximations

Abstract

Metamodels approximate complex multivariate data sets from simulations and experiments. These data sets often are not based on an explicitly defined function. The resulting metamodel represents a complex system’s behavior for subsequent analysis or optimization. Often an exhaustive data search to obtain the data for the metamodel is impossible, so an intelligent sampling strategy is necessary. While multiple approaches have been advocated, the majority of these approaches were developed in support of a particular class of metamodel, known as a Kriging. A more generic, commonsense approach to this problem allows sequential sampling techniques to be applied to other types of metamodels. This research compares recent search techniques for Kriging metamodels with a generic, multi-criteria approach combined with a new type of B-spline metamodel. This Bspline metamodel is competitive with prior results obtained with a Kriging metamodel. Furthermore, the results of this research highlight several important features necessary for these techniques to be extended to more complex domains.