MV-STGHAT: Multi-View Spatial-Temporal Graph Hybrid Attention Network for Decision-Making of Connected and Autonomous Vehicles

Cooperative decision-making technology for connected and autonomous vehicles (CAVs) plays a crucial role in the advancement of autonomous driving. Recently, graph reinforcement learning (GRL)-based methods have shown impressive results in addressing decision-making challenges by utilizing graph-based technologies. However, existing GRL research still faces difficulties in fully modeling mixed traffic scenarios and effectively generating driving feature embeddings. To address these issues, a multi-view spatial-temporal graph hybrid attention network (MV-STGHAT) is proposed to improve the feature extraction ability of the neural network model in the GRL-based framework, thereby improving the decision-making performance of CAVs. Initially, a method for constructing a multi-view spatial-temporal graph is introduced to effectively represent the interactions amongst vehicles. Then, an MV-STGHAT model is proposed, which integrates a dynamic gated graph attention network (DG-GAT) and channel attention temporal convolutional network (TCN) to efficiently generate the informative feature embedding. Furthermore, the double deep Q-learning (DDQN) algorithm is employed to train the proposed MV-STGHAT model. Finally, extensive experiments are conducted across three representative traffic scenarios to validate the proposed approach. Results show that our proposed method achieves zero collision and outperforms the baselines in multiple critical evaluation metrics, with improvements of up to 55.61% in overall reward, 33.68% in average speed, 33.49% in traveling efficiency, and 88.48% in model stability.