A multi-objective decomposition-based evolutionary algorithm with enhanced variable space diversity control

Most Multi-objective Evolutionary Algorithms (MOEAs) operate without explicitly promoting the diversity of the variable space. Nevertheless, in the single-objective domain it has been shown that properly managing this kind of diversity might lead to higher-quality solutions. In this paper the diversity of the variable space is analyzed for several state-of-the-art MOEAs with well-known benchmarks, showing that in the long term, the diversity is lost in a subset of variables. This loss implies an important degradation of the performance. In order to show that increasing the diversity can solve these issues, MOEA/D with Enhanced Variable-Space Diversity (MOEA/D-EVSD) is proposed. This variant induces a gradual loss of diversity by altering the mating selection process. In addition, a final phase to properly intensify is included. The experimental validation was carried out with the Walking Fish Group (WFG) benchmark and several state-of-the-art MOEAs showing the benefits of the proposal. Particularly, WFG1 and WFG8, which are not properly solved by most state-of-the-art approaches, are readily solved by our proposal.