Distribution of evolutionary algorithms in heterogeneous networks

Abstract

While evolutionary algorithms (EAs) have many advantages, they have to evaluate a relatively large number of candidate solutions before producing good results, which directly translates into a substantial demand for computing power. This disadvantage is somewhat compensated by the ease of parallelizing EAs. While only few people have access to a dedicated parallel computer, recently, it also became possible to distribute an algorithm over any bunch of networked computers, using a paradigm called "grid computing". However, unlike dedicated parallel computers with a number of identical processors, the computers forming a grid are usually quite heterogeneous. In this paper, we look at the effect of this heterogeneity, and show that standard parallel variants of evolutionary algorithms are significantly less efficient when run on a heterogeneous rather than on a homogeneous set of computers. Based on that observation, we propose and compare a number of new migration schemes specifically for heterogeneous computer clusters. The best found migration schemes for heterogeneous computer clusters are shown to be at least competitive with the usual migration scheme on homogeneous clusters. Furthermore, one of the proposed migration schemes also significantly improves performance on homogeneous clusters. © Springer-Verlag Berlin Heidelberg 2004.