A Memory-Efficient Elitist Genetic Algorithm

Abstract

This paper proposes a memory-efficient elitist genetic algorithm (me²GA) for solving hard optimization problems quickly and effectively. The idea is to properly reconcile multiple probability (distribution) vectors (PVs) with elitism. Multiple PVs (rather than a single PV as in compact GA (cGA)) provide an effective framework for representing the population as a probability distribution over the set of solutions. A coordinated interplay amongst multiple PVs maintains genetic diversity, thereby recovery from decision errors is possible. On the other hand, reconciling with elitism allows a potentially optimal (elitist) solution to be kept current as long as other (competing) solutions generated from PVs are no better. This is because it exerts a selection pressure that is high enough to offset the disruptive effects of uniform crossover. It also attempts to adaptively alter the selection pressure in accordance with the degree of problem difficulty through pair-wise tournament selection strategy. Experimental results show that the proposed algorithm generally exhibits a superior quality of solution. Moreover, the proposed algorithm deploys memory more efficiently than extant sGA and cGA, especially when the problem is difficult.