Group-benefit control strategy for connected automated vehicles in mixed traffic at intersections

• A MADDPG-based model with dynamic state encoding manages uncertainty in heterogeneous traffic flows. • A group-benefit strategy jointly optimizes traffic efficiency and emissions at signalized intersections. • The policy network uses attention, regularization , and dropout to improve adaptability to dynamic traffic. The unpredictable driving behavior of Human-driven Vehicles (HVs) exhibits significant complexity when mixed with Connected Autonomous Vehicles (CAVs) and the periodic variations of the intersection signal phase impact the efficiency and stability of traffic flows. In response to these challenges, this study proposed a group-benefit driving behavior control strategy based on a Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm to provide comprehensive and overall efficient driving guidance for CAVs in multi-agent scenarios and reduce the impact of heterogeneous traffic flow on the signalized intersections. The dynamically encoded state enables the model to effectively adapt to dynamically changing environments, and each agent independently makes combined action decisions to achieve group benefits. Regularization methods and the attention mechanism are integrated to prevent overfitting and enhance stability, with the attention mechanism having a more significant impact. The results indicate that the model effectively reduces delays and fuel consumption even at low CAV Market Penetration Rates (MPRs), with the optimization effects significantly improving as MPR increases. Compared to the HV traffic flow, the optimized CAV traffic flow reduces average vehicle delay by 10.11% and CO 2 emissions by 20.93%. The comparison with test models, which separately adjust traffic flow and signal phases, and interpolation experiments, which modify the MPR, demonstrates the model’s strong generalization ability and adaptability. Furthermore, compared with the Multi-Agent Deep Q-Network (MADQN) model, the proposed model demonstrates greater robustness with significantly better optimization results.