A subspace-knowledge transfer based dynamic constrained multiobjective evolutionary algorithm

Dynamic constrained multiobjective optimization problems (DCMOPs) have gained increasing attention in the evolutionary computation field during the past years. Among the existing studies, it is a significant challenge to rationally utilize historical knowledge to track the changing Pareto optima in DCMOPs. To address this issue, a subspace-knowledge transfer based dynamic constrained multiobjective evolutionary algorithm is proposed in this article, termed SKTEA. Once a new environment appears, objective space is partitioned into a series of subspaces by a set of uniformly-distributed reference points. Following that, a subspace that has complete time series under certain number of historical environments is regarded as the feasible subspace by the subspace classification method. Otherwise, it is the infeasible one. Based on the classification results, a subspace-driven initialization strategy is designed. In each feasible subspace, Kalman filter is introduced to predict an individual in terms of historical solutions preserved in external storage. The predicted individuals of feasible neighbors are transferred into the infeasible subspace to generate the one, and then an initial population at the new time is formed by integrating predicted and transferred individuals. Intensive experiments on 10 test benchmarks verify that SKTEA outperforms several state-of-the-art DCMOEAs, achieving good performance in solving DCMOPs.