Guided prediction strategy based on regional multi-directional information fusion for dynamic multi-objective optimization

Region partitioning is effective for solving dynamic multi-objective optimization problems (DMOPs). However, most region partitioning approaches use only specific individual information to predict directions within each region. Their efficiency degrades when the distribution of individuals is irregular, and the use of several methods to obtain high-quality areas incurs high computational costs. To address these problems, this study develops a guided prediction strategy based on regional multi-directional information fusion for dynamic multi-objective optimization (RMDIF). Firstly, quantiles are used in the subregional segmentation, whose computational cost is small. Secondly, to increase the prediction accuracy and adaptability of the algorithm for individuals with irregular distributions, information from the center and boundary points of each subregion is fused to construct a new direction for generating initial individuals in new environments. Similar to the quantile-guided dual-prediction strategy, a dual-space prediction strategy is used to generate individuals in new environments to increase the population diversity. Finally, a "maintain-decline-maintain" strategy is used to determine the proportion of new individuals from two prediction spaces. Compared with the fixed proportion method, the proposed method better balances convergence and diversity. RMDIF and six other algorithms are tested on 27 DMOPs, the proposed algorithm outperformed the others in most cases.