An Efficient Sampling Approach to Offspring Generation for Evolutionary Large-Scale Constrained Multi-Objective Optimization

Constrained multi-objective evolutionary algorithms have been extensively used for solving real-world problems. However, most algorithms struggle to efficiently find feasible solutions when the problem involves massive decision variables and decision space constraints. To tackle this issue, an efficient sampling approach is suggested to guide the offspring generation, where three types of directions are utilized according to the status of the current population, including the feasibility-preferred direction, the convergence-preferred direction, and the diversity-preferred direction. Specifically, the proposed approach adopts the feasibility-preferred direction to guide solutions towards constraint satisfaction when most solutions are infeasible, whereas the convergence-preferred direction is utilized to guide solutions to approach the optimal set when most solutions are dominated, and the diversity-preferred direction is employed to spread solutions to cover the optimal set when most solutions are non-dominated. Besides, a reinforcement learning approach is proposed to automatically determine the constraint handling technique in each iteration. With the proposed approaches, a large-scale constrained multi-objective evolutionary algorithm is also developed. The experiment is conducted on 31 benchmark problems with 1000 dimensions and five real-world problems with dimensions varying from 1170 to 2610, and experimental results reveal the competitive effectiveness of the proposed algorithm.