Extending cellular evolutionary algorithms with message passing

Abstract

Cellular evolutionary algorithms (cEAs) use structured populations whose evolutionary cycle is governed by local interactions among individuals. This helps to prevent the premature convergence to local optima that usually takes place in panmictic populations. The present work extends cEAs by means of a message passing phase whose main effect is a more effective exploration of the search space. The mutated offspring that potentially replaces the original individual under cEAs is considered under message passing cellular evolutionary algorithms (MPcEAs) as a message sent from the original individual to itself. In MPcEAs, unlike in cEAs, a new message is sent from the original individual to each of its neighbors, representing a neighbor’s mutated offspring whose second parent is selected from the neighborhood of the original individual. Thus, every individual in the population ultimately receives one additional candidate for replacement from each of its neighbors rather than having a unique candidate. Experimental tests conducted in the domain of real function optimization for continuous search spaces show that, in general, MPcEAs significantly outperform cEAs in terms of effectiveness. Specifically, the best solution obtained through MPcEAs has an importantly improved fitness quality in comparison with that obtained by cEAs.