A Pareto Front searching algorithm based on reinforcement learning for constrained multiobjective optimization

Constraint multiobjective algorithms are the most widely applied direction in intelligent optimization, with excellent research value. Currently, most multiobjective multi-constraints algorithms are designed based on the relationship between feasible and infeasible solutions, but they ignore the complex associations between constraints. At the same time, the most significant difficulty in constraint problems lies in the irregularity of constrained Pareto front (CPF) and the lack of a powerful strategy for exploration. The paper proposes a Pareto front searching based on reinforcement learning (PFRL) for multi-constraints multiobjective optimization problems. The algorithm employs reinforcement learning to guide the evolution process through interaction with the environment and adaptively learns the shape and characteristics of CPF to cover the structure of CPF effectively. The environment and CPF information gained by reinforcement learning are utilized for CPF translation and extension to deal with various irregular feasible regions. In addition, the paper also designs a constraint priority evaluation mechanism based on the correlation distance (CD) metric to process constraint relationships. It allows the algorithm to effectively cross over Pareto front (PF) of a single constraint that is unrelated to CPF, improving algorithm efficiency. The introduced algorithm implemented the above strategy using only one population. The effectiveness of the introduced algorithm was verified and compared with nine state-of-the-art algorithms and four real-world constrained multiobjective optimization problems (CMOPs). Experimental results show that the algorithm provides a low-resource and efficient method for solving CMOPs.