Reinforcement learning-assisted parameter control for constrained multi-objective evolutionary optimization

Reinforcement learning (RL)-assisted parameter control has become an important route for constrained multi-objective evolutionary optimization, but many existing studies still follow an in-situ learning paradigm in which policy learning, feedback acquisition, and controller updating remain coupled with the solving process of the current instance. This coupling increases runtime cost, limits policy learning under sparse single-instance interaction data, and prevents cross-instance reuse of learned control regularities. This paper develops RL-IMTCMO-BS, a reinforcement learning-assisted extension of IMTCMO-BS. The proposed method learns a reusable parameter-control policy over a set of training instances before formal solving. The controller is trained using search-state representation, continuous action design, and the dual-track lead-delta reward. During formal solving, the trained controller is frozen and embedded into the original solver, where the learned policy configures parameters dynamically according to the search state. An instance-level training, validation, and test protocol is established to evaluate the learned control policy on held-out instances independently of the training data. Experimental results under two cross-instance settings, within-problem and multi-problem, show that RL-IMTCMO-BS achieves positive lead relative to the original IMTCMO-BS in most tested cases. Under both settings, RL-IMTCMO-BS achieves statistically significant rank advantages over multiple competitors. These results support generalization of the learned control policy to held-out instances of a given SDC test function and across multiple SDC test functions.