Surrogate-assisted coevolutionary search
Abstract
This paper is concerned with an experimental evaluation of coevolutionary optimization techniques, which are integrated with surrogate models of the fitness function. The motivation for this study arises from the fact that since coevolutionary search is based on the divide-and-conquer paradigm, it may be possible to circumvent the curse of dimensionality inherent in surrogate modeling techniques such as radial basis networks. We investigate the applicability of the algorithms presented in this paper to solve computationally expensive optimization problems on a limited computational budget via studies on a benchmark test function and a real world two-dimensional cantilevered space structure design problem. We show that by employing approximate models for the fitness, it becomes possible to converge to good solutions even for functions with a high degree of epistasis.
Authors: Yew-Soon Ong, Andy J. Keane, Prasanth B. Nair
Published in: International Conference on Neural Information Processing (ICONIP) (2004)