A multi-task optimization algorithm via reinforcement learning for multimodal multi-objective optimization

Abstract

Solving multimodal multi-objective optimization problems (MMOPs) via evolutionary algorithms has recently garnered increasing attention. Maintaining diversity in both decision and objective spaces is crucial for effectively handling MMOPs. However, most traditional multimodal multi-objective evolutionary algorithms (MMEAs) prioritize convergence in the objective space, often eliminating poorly converged solutions which could enhance diversity in the decision space. To address this issue, this paper proposes a novel MMEA, named QLMTMMEA. Specifically, a multi-task optimization framework comprises a main task and three auxiliary tasks based on different strategies for MMOPs is designed. Then, Q-Learning (QL) is utilized to the adaptively selects optimal auxiliary tasks in the evolution process. In addition, a new diversity enhancement technique is proposed for objective space and decision space by dynamically adjusting the relaxation factor to maintain high quality solutions. Seven state-of-the-art MMEAs are adopted to make comparisons for demonstrating the performance of QLMTMMEA, experimental results show that QLMTMMEA is competitive compared to others MMEAs on 34 complex MMOPs.