Multi-View Graph Convolution Network Reinforcement Learning for CAVs Cooperative Control in Highway Mixed Traffic.

The control of connected autonomous vehicles (CAVs) for cooperative sensing and driving in mixed traffic flows is critical for the development of intelligent transportation systems. However, the challenge lies in effectively representing and learning from the interactions between heterogeneous traffic agents, as well as the complex traffic states they produce. To address this, we propose the Multi-View Graph Convolution Network Reinforcement Learning (MVGRL) algorithm for the decision-making and control of CAVs in mixed traffic scenarios. First, multi-view graphs are constructed based on multiple logical correlations and the perception-communication relationship of heterogeneous vehicles in mixed traffic. This enables us to represent the dynamic vehicle networks and gain insights into vehicle behavior. Then, the features of multi-view vehicle network graphs are derived by multi-view graph convolutional network reinforcement learning. Finally, the lane change decisions of CAVs are obtained based on complex traffic state learning. In the highway scenario, the model is able to control the CAVs to complete the ramp exit task and maintain an overall efficient traffic flow. Various test scenarios were formed by adjusting the traffic conditions under a highway off-ramp scenario for comparison experiments. The results show that the proposed MVGRL algorithm model outperforms in decision-making and robustness than the baseline models.