Deep Reinforcement Learning Based Environmental and Mating Selection for Evolutionary Multi-objective Optimization
Real-world optimization problems often involve multiple and conflicting objectives, giving rise to multi-objective optimization problems (MOPs). Over the past decades, evolutionary algorithms have been widely recognized as an effective approach for solving MOPs. In multi-objective evolutionary algorithms (MOEAs), mating selection and environmental selection play a crucial role in guiding the evolutionary search. Numerous selection mechanisms have been proposed, yet no single mechanism is universally optimal for all problems or throughout all stages of the evolutionary process. To address this challenge, this study proposes a deep reinforcement learning (DRL)-based adaptive selection framework for MOEAs. Specifically, a deep Q-network (DQN) is employed to dynamically choose suitable environmental and mating selection mechanisms from a predefined pool during the evolutionary process. We design several state representations to capture the convergence and distribution characteristics of the search progress. Furthermore, an informative reward function is developed to effectively guide the DRL agent. The proposed algorithm is extensively evaluated on 31 standard benchmark functions (DTLZ, WFG, and MaF) and compared against several state-of-the-art DRL-based adaptive MOEAs. Experimental results demonstrate that the proposed approach consistently achieves superior performance.