An operator pre-selection strategy for multiobjective evolutionary algorithm based on decomposition

Abstract

Listen

Translate article

Evolutionary algorithms (EAs) are a population-based optimization method that adopts survival-of-the-fittest rules. The performance of EAs can be greatly affected by offspring quality. Many researchers hope to adjust the parameters of evolutionary operators to improve offspring quality, but most current methods are regulated by the feedback model, which does not recognize offspring quality. Instead, it adjusts an operator’s search structure dynamically based on its historical performance. In fact, there is a co-evolutionary effect between different operators. An operator will not always produce offspring with the best quality, so the feedback model only selects the best operator in history to generate offspring, which ignores the capacity for co-evolution. In this paper, we propose an operator pre-selection strategy (OPS). First, to ensure operators’ capacity for co-evolution, we construct an operator pool using five operators. Second, we select the best offspring based on the classification and feedback models. Finally, we evaluate the offspring and include it in the evolution. Compared with the traditional feedback model, OPS directly identifies higher-quality offspring. Even if the historical performance of an operator is poor, as long as it can generate excellent offspring, it is selected. Experimental results demonstrate the effectiveness of OPS.