A priori injection strategy for dynamic multi-objective optimization via dynamic routing evolution method
Fei Wu, Jiacheng Chen, Wanliang Wang, Guoqing Li · Journal of Computational Design and Engineering · 2025
Abstract The dynamic multi-objective problem aims to swiftly track the Pareto frontier, necessitating a delicate equilibrium between convergence speed and population diversity in the evolutionary process. Numerous strategies have been implemented, aiming to expedite convergence through prediction or diversification via the generation of new populations. Yet, these strategies often lead to unintended and counterproductive outcomes. Additionally, the incorporation of prior knowledge has been largely overlooked in this context. In this paper, we propose an a priori injection strategy for dynamic multi-objective optimization via dynamic routing evolution approach (DR-MOEA). The strategy refines the partitioning of the degree of change of the dynamic environment, aggregates the advantages of multiple dynamic optimization strategies, and responds flexibly to dynamic demands. The dynamic routing threshold can adaptively segment the dynamic environment into three levels. The reverse checking strategy is used in the case of slightly changing environment. On the contrary, the population reconfiguration strategy is used in the case of a drastically changing environment. The population reconfiguration strategy is more helpful for the population to adapt to the massively changing environment than the single strategy that cannot cope with the massively changing environment and the strategy that blindly initializes the population. In the case of moderate change, the strategy will be dynamically routed to the a priori knowledge injection strategy, which is more instructive for population prediction. The proposed DR-MOEA is compared with five state-of-the-art dynamic strategies on 14 benchmarks, and the experimental results show that the DR-MOEA method proposed in this paper is effective, robust, and competitive.