A potential game car-following model for autonomous vehicles with reinforcement learning-guided genetic algorithm calibration

Abstract

The car-following behaviour of autonomous vehicles (AVs) directly influences road capacity and infrastructure design. In this study, a potential game-based car-following model (PG-CFM) was developed, and its parameters were calibrated using a reinforcement learning-guided genetic algorithm. A potential game cost function was employed to represent multi-objective decision making under spacing, speed-difference and acceleration constraints. The model was calibrated and validated using real-world trajectory data from the Waymo data set and compared with the intelligent driver model and the Gipps model. The results showed that the PG-CFM outperformed both benchmark models in predictive accuracy and stability. Under open-loop validation, the model achieved a mean absolute error of 1.00 m/s2 and a root mean square error of 1.31 m/s2, representing a 4.2% reduction relative to the Gipps model. The error distribution was also more concentrated, indicating improved robustness. The calibrated parameters suggested that AVs operate with an expected time headway of approximately 1.18 s while maintaining smoother acceleration profiles. This behaviour reflects a stable balance between safety and traffic efficiency. These findings provide quantitative support for headway setting, road-capacity assessment and infrastructure planning for autonomous mobility systems.