A Highway Vehicle Lane-Changing Decision Optimization Framework Based on Two-stage GAIL-PPO Training

This paper addresses challenges in intelligent vehicle lane-changing decision-making on highways, including the high dependency of imitation learning on data quality and its limited generalization capability, as well as the low training efficiency of reinforcement learning and its difficulty in balancing multiple objectives. To this end, we propose a two-stage collaborative optimization framework based on Generative Adversarial Imitation Learning (GAIL) and Proximal Policy Optimization (PPO). A Wasserstein distance-based adversarial mechanism with gradient penalty is introduced into the GAIL discriminator to enhance training stability. PPO is integrated into the generator update process of GAIL through an actor–critic architecture, improving the robustness of policy learning. The pre-trained policy is then fine-tuned using PPO-based multi-objective reinforcement learning with a multi-objective reward function that balances traffic efficiency and safety constraints. This enables the transition from expert imitation to policy optimization under complex scenario constraints. Experimental results in the highway-env simulation environment demonstrate that the proposed approach improves average travel speed by approximately 4% and 8% compared with the PPO baseline and DQN, respectively, while effectively reducing unnecessary lane changes. Further analysis of longitudinal acceleration time series and robustness tests validates the method's stability and generalization capabilities under different driving durations, traffic densities, lane numbers, and vehicle dynamics.