A Study on Multimemetic Estimation of Distribution Algorithms

Abstract

Multimemetic algorithms (MMAs) are memetic algorithms in which memes (interpreted as non-genetic expressions of problem solving strategies) are explicitly represented and evolved alongside genotypes. This process is commonly approached using the standard genetic procedures of recombination and mutation to manipulate directly information at the memetic level. We consider an alternative approach based on the use of estimation of distribution algorithms to carry on this self-adaptive memetic optimization process. We study the application of different EDAs to this end, and provide an extensive experimental evaluation. It is shown that elitism is essential to achieve top performance, and that elitist versions of multimemetic EDAs using bivariate probabilistic models are capable of outperforming genetic MMAs.