P-DQN integrates lane-changing and car-following for safer CAVs

A new study from researchers Hao Zhang, Zihao Li, and Yang Zhou tackles a critical bottleneck in autonomous driving: unifying lane-changing and car-following control. Traditionally treated as independent tasks, their interplay is key to understanding how traffic disturbances propagate. The team proposes a framework based on the Parametrized Deep Q-Network (P-DQN), which uses a hybrid action space—discrete for lane-change decisions and continuous for lateral/longitudinal acceleration. This interdependence mimics real-world driving, where a lane change is not a separate action but a coordinated maneuver.

In numerical experiments, the P-DQN-based strategy outperformed separated methods like the MOBIL lane-change model combined with the Intelligent Driver Model (IDM). It scored higher on both safety metrics (reducing sudden braking events) and passenger comfort. The work, published on arXiv (2607.06771), shows promise for real-world CAV deployments where smooth, efficient traffic flow is critical. The framework's ability to learn optimal policies from cumulative rewards could eventually be fine-tuned using real traffic data, bringing us closer to fully integrated autonomous driving systems.