Integrating Actor–Critic Reinforcement Learning With Evolutionary Algorithm for Multimodal Multiobjective Optimization

Effectively solving multimodal multiobjective optimization problems (MMOPs) requires maintaining an optimal balance between the diversity and the convergence. Traditional algorithms often struggle with environmental selection adaptability, leading to suboptimal performance across diverse MMOPs. This article innovatively integrates the actor-critic reinforcement learning (RL) with the evolutionary algorithm, significantly improving environmental selection adaptability through synergistic online learning between its actor and critic components. An RL process that dynamically optimizes the niche size is established, which critically determines the trade-off between the diversity and convergence preferences. The process is formulated by defining: 1) convergence and diversity measures as the state; 2) niche size adjustment as the continuous action; and 3) state improvement as the reward. Two specialized fully connected neural networks are employed as the actor network for action generation and the critic network for value estimation, which collaboratively adapt to population states through real-time online learning. This adaptive niching technology, integrated with local convergence quality assessment, enables comprehensive evaluation of both diversity and potential convergence. Extensive experimental validation demonstrates the superior performance of the proposed algorithm against ten state-of-the-art algorithms across 48 benchmark problems and a real-world application. The results consistently show significant improvements in balance maintenance and overall optimization effectiveness compared to existing algorithms.