A dynamic multi-objective optimization based on knowledge prediction and density clustering strategy

Abstract

Dynamic multi-objective evolutionary algorithms (DMOEAs) that extract historical knowledge from the past environment to predict new solutions are known to be effective for solving dynamic multi-objective optimization problems (DMOPs). However, most of the existing methods simply reuse historical solutions without further extracting the knowledge between different historical environment solutions, which may make the algorithm ignore some important historical knowledge and limit its performance. In this paper, we propose a knowledge prediction strategy and a density clustering strategy for DMOEA, called KPDCS-DMOEA, which aim to extract historical knowledge from the past environment to build a more accurate prediction model. Firstly, the trend of change in the initial environment is obtained by predicting previous environmental changes through linear prediction methods based on knee point clusters. Secondly, a strategy was proposed to pair the solutions between adjacent environments and construct each dimensional motion vector as historical knowledge. The training set is constructed according to the motion step of the motion vector and the motion direction of each dimension, and the neural network is trained to predict the initial population in the new environment. Finally, a guided evolution strategy based on a density clustering algorithm is developed to speed up population convergence and ensure that the population is well distributed. KPDCS-DMOEA is compared with several state-of-the-art DMOEAs. Experimental results show that the performance of KPDCS-DMOEA is better than the selected comparison algorithms.