Dominance relation selection and angle-based distribution evaluation for many-objective evolutionary algorithm

Abstract

In many-objective optimization, most existing evolutionary algorithms struggle to effectively balance the convergence and diversity of population. Given the failure of Pareto-dominance-based selection mechanism in high-dimensional spaces and the growing recognition of the advantages of angle-based similarity evaluation, we propose a Many-objective Evolutionary Algorithm using Dominance Relation Selection and Angle-based Distribution Evaluation (DSAE). Firstly, based on both traditional Pareto dominance and a newly proposed Angle-dominance with higher selection pressure, a Dominance Relation Selection (DRS) strategy is designed, which dynamically adjusts the selection pressure by assessing the overall proximity between parent and offspring populations. Then, an Angle-based Distribution Evaluation is proposed to determine the distribution levels of solutions, so as to adaptively maintain promising distribution performance of population in objective space. Furthermore, a Pareto-dominance-based global optimal solution set is introduced to save solutions with the best diversity throughout the iteration as the final output. Finally, DSAE is compared with 7 state-of-the-art algorithms by using benchmark test suites DTLZ and WFG with 5, 8 and 10 objectives, and the experimental results demonstrate that DSAE has the most advanced performance.