Migration of probability models instead of individuals: An alternative when applying the island model to EDAs

Abstract

In this work we experiment with the application of island models to Estimation of Distribution Algorithms (EDAs) in the field of combinatorial optimization. This study is motivated by the success obtained by these models when applied to other meta-heuristics (such as genetic algorithms, simulated annealing or VNS) and by the use of a compact representation of the population that make EDAs through probability distributions. This fact can be exploited during information interchange among islands. In this work we experiment with two types of island-based EDAs: (1) migration of individuals, and (2) migration of probability models. Also, two alternatives are studied for the phase of model combinations: assigning constant weights to inner and incoming models or assigning adaptive weights based-on their fitness. The proposed algorithms are tested over a suite of four combinatorial optimization problems. © Springer-Verlag 2004.