Three-stage multi-population constrained multi-objective evolutionary algorithm with role dynamic adjustment

Abstract

Multi-population co-evolutionary algorithms have been widely recognized as an effective means for solving constrained multi-objective optimization problems. However, designing effective cooperation mechanisms to simultaneously balance the trade-offs between objectives and constraints, and between diversity and convergence, remains a challenging issue. First, this study proposes a three-stage cooperation mechanism that dynamically adjusts the roles of subpopulations to control the evolutionary trajectory of the population. Second, we develop an evolutionary state-driven multi-strategy constraint-handling technique. Third, we construct an adaptive search strategy based on an external archive and deep reinforcement learning. Finally, we propose a three-stage multi-population constrained multi-objective evolutionary algorithm with dynamic role adjustment. The performance of the proposed algorithm is validated on 47 benchmark test problems and 21 real-world problems. Experimental results demonstrate that the proposed algorithm significantly outperforms ten compared algorithms. Furthermore, ablation studies and parameter analyses are conducted to verify the effectiveness of each component.