Optimization and Prediction of Crashworthiness in Thin-Walled Structures Using a Multi-Strategy Multi-Objective Algorithm Based on Reinforcement Learning

This paper presents a reinforcement learning-based multi-strategy multi-objective evolutionary algorithm (RLMMOEA) aimed at optimizing the energy absorption (EA) performance of thin-walled structures. By integrating reinforcement learning with the NSGA-III algorithm, the system adaptively selects the most effective strategy in real-time through three distinct methodologies. An adaptive mutation strategy is introduced to circumvent the algorithm’s convergence to local optima, while a sliding window mechanism ensures that the solution archive reflects diverse solutions and promotes population diversity throughout the evolutionary process. Additionally, a neural network with Bayesian regularization is employed to enhance model generalization and robustness. The findings indicate that RLMMOEA surpasses the traditional NSGA-III in structural optimization, achieving a shape factor of 0.9, a mass of 261[Formula: see text]g, a wall thickness of 1.48[Formula: see text]mm, an IPCF of 113.18[Formula: see text]kN, and an EA of 6523.46[Formula: see text]J, thereby successfully fulfilling lightweighting objectives. Validation through finite element analysis corroborates the enhanced EA and reduced peak force in the optimized design. Furthermore, RLMMOEA exhibits superior performance across 22 standard test functions, underscoring its competitive advantage. This methodology presents a robust approach for enhancing the crashworthiness of thin-walled structures, with substantial implications for engineering applications.