Solving Dynamic Multiobjective Optimization Problems via Feedback-Guided Transfer and Trend Manifold Prediction

Solving dynamic multiobjective optimization problems (DMOPs) is very challenging due to the requirements to respond rapidly and precisely to changes in an environment. Many prediction- and memory-based algorithms have been recently proposed for meeting these requirements. However, much useful knowledge has been ignored during the historical search process, and prediction deviations could occur, thus limiting the applicability of these methods to a variety of problems. Facing these concerns, this article proposes an evolutionary algorithm named FGTTMP based on feedback-guided transfer (FGT) and trend manifold prediction (TMP) for solving DMOPs. The FGT employs an information feedback model to extract valuable knowledge using all historical environments and then identifies excellent individuals using cluster-transfer learning. This can both accelerate convergence and introduce diversity for future environments. The TMP applies the probability-based trend prediction method to estimate the mass center of the whole population and the corresponding manifold, relying on the two previous moments. Thus, the FGT and TMP strategies combine historical knowledge with prediction techniques, synergistically leading the population in promising directions. The performance of the proposed algorithm is fully investigated and compared with eight state-of-the-art algorithms. Experimental results demonstrate that the proposed FGTTMP method can achieve better convergence and diversity on 19 various benchmark problems than the state-of-the-art algorithms.