Reproductive bias, linkage learning and diversity preservation in bi-objective evolutionary optimization

Abstract

Abstract Diversity preservation is a crucial component for any multiobjective evolutionary algorithm, and its effectiveness defines how well an algorithm can find solutions to cover the whole extension of the Pareto-optimal front. In this paper, we show that traditional reproduction operators such as p-uniform and n-point crossover may sabotage the functioning of diversity preservation mechanisms by producing more solutions in some areas of the objective-space than in others, i.e., they are biased. We argue that such reproductive bias is due to their high degree of disruptiveness which favors the generation of average quality offspring. Additionally, we demonstrated the impact of linkage learning in decreasing disruptiveness and reproductive bias. Such a result helps to understand the benefits of estimation of distribution algorithms in bi-objective optimization. We performed experiments on instances of the ρMNK-model, in which the use of unbiased reproduction operators was shown to work in synergy with diversity preservation mechanisms favoring the diversity of the Pareto-fronts obtained.