Gaussian Process meta-modeling and comparison of GP training methods

Abstract

The ability of Gaussian Process to flexibly and accurately fit arbitrary, even highly nonlinear data sets has lead to considerable interest in their application to many areas. Firstly, the usefulness of Gaussian Process models for application to complex systems metamodeling is proposed. Secondly, several approaches for training Gaussian Process models are examined, which include local optimization algorithm, Genetic Algorithms and Estimation of Distribution Algorithms. The results of these training methods are compared for several example problems, and guidance is provided in GP training methods.