CHSAC: reinforcement learning-based lane-changing decision-making integrating cognitive hierarchy and risk awareness

In mixed-traffic environments, autonomous vehicles must interact with human drivers exhibiting strategic but bounded-rational behaviors, which introduces significant uncertainty for decision-making. To address this challenge, we propose CHSAC (Cognition-aware Hierarchical Soft Actor-Critic), a novel framework that integrates cognitive modeling with deep reinforcement learning for safe highway lane-change planning. CHSAC unifies risk assessment, real-time inference of surrounding vehicles’ cognitive levels, and belief-augmented policy optimization, enabling the vehicle to anticipate human intentions and adjust its policy foresightedly. Extensive simulations demonstrate that CHSAC significantly outperforms competitive reinforcement learning and game-theoretic baselines in safety and efficiency, achieving near-zero collisions while maintaining smooth and efficient driving. These results highlight the effectiveness of incorporating cognitive belief inference into deep policy learning, offering an adaptive and robust decision-making paradigm for autonomous driving in complex mixed-traffic scenarios.