Investigation of RBF Hyperparameter Variation on Evolutionary Search

Abstract

Computationally-intensive problems across engineering and science are often solved by using a metamodel-assisted evolutionary algorithm. A commonly used metamodel is the radial basis functions (RBF) which depends on a hyperparameter. The latter needs to determined by a numerical procedure which by itself could be an intensive task, and this raises the question in which cases it is justified, namely when do variations of the hyperparameter yield significant performance gains. To address this issue this study presents an extensive set of numerical experiments and their analysis to identify when is the hyperparameter variation justified.