Dynamic Multi-Objective Evolutionary Algorithm Based on Learning-Evolved Knowledge

Biao Xu

Keywords: Dynamic multi-objective， Evolutionary algorithms，Transfer learning，Multilayer perceptron

Abstract: Dynamic multi-objective optimization problems (DMOPs) pose significant challenges for traditional evolutionary algorithms due to the continuous evolution of their Pareto-optimal sets (PSs) and Pareto frontiers (PFs). Learning-based approaches demonstrate great potential for rapidly tracking changing Pareto optimal solution sets while maintaining population diversity. However, existing methods for learning evolutionary knowledge fail to adequately account for problem-specific variations. Directly applying historical evolutionary knowledge to new problems may yield suboptimal results when problems undergo significant changes. To overcome this limitation, this study proposes a novel dynamic multi-objective evolutionary optimization algorithm comprising a transfer learning model and a multi-layer perceptron (MLP) model. First, individuals are improved based on a selection mechanism. If the transfer learning model is selected, historical evolutionary knowledge is transferred to the new environment, mitigating knowledge obsolescence due to problem differences and creating more promising individuals. If the multilayer perceptron model is chosen, it enhances individual quality by learning evolutionary process knowledge specific to the current problem. Population iteration is then completed through differential evolution operators and environmental selection. Our proposed algorithm underwent rigorous testing using benchmark functions, with results demonstrating superior performance compared to existing state-of-the-art algorithms.

Submission Number: 45