Feasibility-guided search and prediction for dynamic constrained multiobjective evolutionary optimization

Dynamic constrained multiobjective optimization problems (DCMOPs) are characterized by the variations of both objectives and constraints over time, posing two main challenges: (1) balancing feasibility, convergence, and diversity in the evolutionary search and (2) generating an effective initial population for new environments. To address these problems, this paper proposes a dynamic constrained multiobjective evolutionary algorithm with feasibility-guided search and prediction (called FGSP), which integrates a feasibility-guided evolutionary search (FGES) and a feasible information guidance prediction (FIGP). Specifically, FGES adaptively adjusts evolutionary strategies by monitoring the proportion of infeasible solutions and a time-dependent tolerance threshold for infeasibility, such that it can perform exploration without constraints to navigate through large infeasible regions and conduct feasibility-driven exploitation to refine solutions near the constrained Pareto front, thereby balancing convergence, feasibility, and diversity. Concurrently, FIGP utilizes an artificial neural network trained on historically feasible solutions to predict a high-quality initial population for new environments, significantly accelerating adaptation to dynamic changes via pattern learned from past environments. After comparing the proposed FGSP with five state-of-the-art algorithms on the latest benchmark problems and one real-world problem, the experimental results validate the effectiveness of FGSP in obtaining feasible non-dominated solutions. • FGES is proposed to help the population balance feasibility, convergence, and diversity. • FIGP predicts changes in feasible solutions using past data to train a neural network. • FGSP combines FGES and FIGP to solve DCMOPs, validated on benchmarks and a real case.