A multistage evolutionary algorithm for many-objective optimization

Convergence and diversity are of high significance to many-objective optimization, which are considered by most state-of-the-art many-objective evolutionary algorithms (MaOEAs) simultaneously. However, it is not easy to balance them during the optimization process due to their conflicting nature. This study proposes a multistage MaOEA to address this issue, where convergence and diversity are processed respectively at different optimization stages. At the first stage, the population approaches Pareto front rapidly and the diversity is ignored. After the population is converged, the diversity will be emphasized by applying the decision variable clustering method at the second stage. When the population achieves high convergence and diversity, the algorithm will enter the last stage, where the quality of the solution set is fine-tuned by substituting those solutions with worse convergence and diversity degrees. As demonstrated by the experimental results with peer competitors on common benchmark problems, that the proposed algorithm is promising.