Twin-population multiple knowledge-guided transfer prediction framework for evolutionary dynamic multi-objective optimization

Dynamic multi-objective evolutionary algorithms (DMOEAs) have been widely studied, and one of the main tasks is the need for algorithms to track Pareto optimal front (POF) under dynamic environmental changes. Existing methods integrate transfer learning (TL) techniques to predict the initial population for the new environment. However, the lack of transferred individual diversity and inaccurate moving directions lead to poor performance of DMOEAs. Therefore, this work proposes a Twin-population Multiple Knowledge-guided Transfer prediction (TMKT) framework to form an initial population for the new environment. Three strategies, i.e., Twin Populations Guided prediction (TPG), SVM-based Multi-knowledge prediction (SVM-M) and Kernel Subspace Alignment for Transfer prediction (KSA-T), are designed to mine and transfer positive historical knowledge for accurately predicting changing POFs. First, TPG is used to obtain new approximate individuals and provide potential directions of subsequent transfer, which splits the population into two twin populations based on upper and lower quartile points of the first objective and their angles. Subpopulations transmit information by different similarity methods to obtain their new positions. Secondly, to obtain solutions with better diversity and convergence, SVM-M trains a certain classifier that can discriminate between positive and negative solutions and predicts labels of noisy solutions based on useful knowledge from the first two environments. Third, KSA-T is proposed to further enhance the accuracy of the new population prediction. The kernel trick and second-order feature alignment are introduced in subspace alignment to develop a new TL technique called Kernel Subspace Alignment (KSA) for adaptively achieving homotypic distributions of the source domain and target domain. Solutions predicted by TPG as the target domain are employed to guide the evolution, and obtainedSVM-M positive solutions are transferred to the new environment via KSA. TMKT is integrated with two baseline algorithms MOEA/D and NSGA-II to construct DMOEAs. Numerical results on 14 functions of different variation types and a real parameter optimization problem of control system validate the superior dynamic optimization performance of TMKT compared with five state-of-the-art algorithms.