Growing Neural Gas Network for Offspring Generation in Evolutionary Constrained Multi-Objective Optimization

When using evolutionary algorithms to solve constrained multi-objective optimization problems, both the constraint handling technique and the search operator play a crucial role. However, existing algorithms always focus on the former for the promising solution selection while neglecting the importance of the latter for the feasible solution generation. In this article, we emphasize the role of the search operator and propose a model-driven offspring generation method for evolutionary constrained multi-objective optimization. To be specific, the growing neural gas network is used to collect the feasibility information during the evolutionary search process and gradually model the distribution of feasible regions in the decision space. In order to make use of infeasible solutions to assist in the modeling of feasible regions, both feasible and infeasible solutions are taken as two classes of samples to train the network in the supervised learning manner. The proposed method can utilize the learning network to generate the promising offspring and thus make the population converge faster toward the optimal feasible regions. Experimental results on three public test suites demonstrate the competitiveness of the proposed approach, especially on the test problems with complicated feasible regions.