Collaborative Decision Making With Hybrid Deep Reinforcement Learning for Connected and Autonomous Vehicles in Mixed Traffic

Collaborative decision-making for connected and autonomous vehicles plays a critical role in improving traffic efficiency and safety. However, some challenges emerge, such as the stochastic nature of driving behaviors of human-driven vehicles in mixed traffic, the complexity of the longitudinal and lateral decision-making processes for connected and autonomous vehicles (CAVs). To address these challenges, this study proposes a hybrid deep reinforcement learning framework that optimizes both longitudinal acceleration and lateral lane-changing decisions for CAVs. A graph-based information interaction mechanism is developed to capture dynamic information on interdependencies in mixed traffic flows, enabling real-time communication and coordination. Additionally, a spatial-temporal encoder is integrated to effectively process heterogeneous environmental data and enhance the model’s capacity to learn complex traffic patterns. Finally, the hybrid deep reinforcement learning framework, combining an actor-critic architecture with Q-learning, is established to derive optimal policies for collaborative decision-making. Simulation results demonstrate that the proposed model achieves an 82% destination arrival rate and reduces 72% average traffic delay. Furthermore, the robustness evaluations under noisy communication conditions reveal that the model maintains stable decision-making.