A two-stage coevolutionary algorithm based on adaptive weights for complex constrained multiobjective optimization

In the constrained multiobjective optimization problems (CMOPs), various complex constraints need to be satisfied simultaneously, which further challenges evolutionary algorithms in balancing feasibility, convergence and diversity. Recent advances in evolutionary computation have led to the development of multi-stage and multi-population strategies for handling CMOPs. However, most algorithms have shown poor performance when dealing with problems with low feasible ratio or discontinuous feasible regions. To address this issue, we propose a two-stage coevolutionary algorithm based on adaptive weights (AW-TCEA), aiming to balance convergence, diversity and feasibility to handle CMOPs with complex Pareto fronts (PFs). Specifically, the first stage uses two populations to explore the objective space and feasible regions respectively, one driven by objective information and the other by feasible information. The second stage adopts a set of weight vectors to search for unexplored feasible regions to enhance diversity. In addition, for handling complex constrained PFs, a novel adaptive weight adjustment strategy is proposed to explore ineffective directions and develop potential regions. Experimental comparisons with multiple state-of-the-art algorithms are performed on 50 test problems and 5 real-world problems. The results show that the proposed algorithm exhibits better performance on various CMOPs.