Expensive Constrained Multi-objective Evolutionary Algorithm with Pareto Set Learning
Ruitao Mai, Yifeng Qiu, Wenji Li, Zhaojun Wang, Biao Xu, Wei Chen, Jiafan Zhuang, Yun Li, Zhun Fan · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2024
Expensive constrained multi-objective optimization problems (ECMOPs) necessitates the acquisition of feasible optimal solutions within a limited number of evaluations. To address these challenges, we introduce a neural network-based Pareto set learning method. This method enhances search efficiency by learning the relationship between weight vectors in the objective space and their corresponding Pareto optimal solutions, thus enabling direct prediction of solutions within the Pareto set. Additionally, we propose a push and pull search algorithm based on Pareto set learning (PPS-PSL) for ECMOPs. This algorithm operates in two stages: learning unconstrained Pareto sets (UPS) in the push stage and constrained Pareto sets (CPS) in the pull stage, while employing evolutionary algorithms to guide the population away from local optima. Comparative experimental results, conducted on fourteen benchmark problems against six comparison algorithms, demonstrate the superiority of our proposed approach.