A Joint Prediction Strategy Based on Multiple Feature Points for Dynamic Multi-objective Optimization

This paper proposes a joint prediction strategy based on multiple feature points (JP-MFP) to improve the prediction accuracy and balance the convergence and diversity. On the one hand, a multi-step prediction based on center points is designed to generate some predicted individuals around the potential area of the new Pareto optimal Set (POS), which expands the coverage area of the predicted population and enhances its diversity. On the other hand, a multi-directional prediction based on extreme and knee points is proposed in which the nondominated solutions are firstly clustered based on the extreme points and knee points, and then different movement directions are assigned to each cluster. Resultantly, it can predict the new POS from multiple directions and improve the distribution of the population. Further, the convergence can be accelerated. Finally, the predicted sub-populations obtained by the two strategies are selected according to non-dominated sorting and crowding distance to generate a joint predicted population for the new environment. Experimental results demonstrate that the proposed JP-MFP can effectively solve DMOPs by comparing with several advanced algorithms on the DF test problems.