A Learning-Assisted Tri-Network-Based Parameter Control Framework: From Small-Scale to Large-Scale Global Optimization

Parameter control improves metaheuristics by automated parameter generation. However, it often relies on handcrafted formulas derived from expert knowledge. The design rationale and decision-making are opaque and hard to generalize. In large-scale scenarios, the cost of the design becomes prohibitively high. This work proposes an Explainable Evolutionary Parameter Learning (EEPL) framework, which constructs a learning-assisted evolutionary system for parameter adaptation. The evolutionary system performs the original parameter setting. To build the learning system, a Long Short-Term Memory (LSTM) network is regarded as a parameter controller that captures an optimal parameter sequence with temporal dependencies. To alleviate the curse of dimensionality in training, a Gene Targeting Differential Evolution (GTDE) evolves LSTMs and targets bottleneck dimensions of the optimal LSTM parameters repeatedly. Subsequently, the optimal parameter controller is embedded into a learned algorithm to learn control parameters offline. A Deep Q-Network (DQN) replaces the human expert by autonomously deciding which system to execute, ensuring efficient collaboration between the two systems. A Search Trajectory Network (STN) provides explainability for parameter behaviors. EEPL has learned Genetic Algorithm (GA), Particle Swarm Optimization (PSO), DE, GTDE, and EBOwithCMAR, and has been validated on CEC benchmarks ranging from 10D to 1000D. Notably, a controller trained on 10D problems generalizes directly to 1000D problems. The results demonstrate that the learned algorithms enhance their original versions and achieve performance comparable to the 13 purely handcrafted CEC winners. EEPL is an explainable, efficient, and fast parameter control method that combines online and offline learning, achieving zero-shot generalization and broad applicability.