Adaptive hybrid control strategy for vehicle platoon combining model predictive control and deep reinforcement learning

Traditional control methods extensively applied to longitudinal control for vehicle platoon suffer from performance degradation when faced with complex and dynamic traffic environment, and deep reinforcement learning (DRL)-based approaches are generally constrained by lower training and sampling efficiency. Therefore, this paper proposes an adaptive hybrid control strategy (AHCS) for vehicle platoon utilizing model predictive control (MPC) and DRL. Specifically, MPC is incorporated into the training process of DRL to tackle the issues of slow convergence and inefficient exploration for DRL agents. Subsequently, we develop an adaptive soft switching mechanism (AS2M) module that integrates a kinematic stopping distance algorithm and a metric called sensitivity factor (SF) to combine the outputs from MPC and DRL modules. Moreover, a rule is formulated to guarantee that the combined output adheres to constraints in real-world driving scenarios. The proposed AHCS not only ensures fundamental control performance through MPC but also utilizes the exploration capability of DRL to address the limitations of MPC. The hybrid deep deterministic policy gradient (H-DDPG) is proposed under the AHCS. The training and testing of DRL agents utilize vehicle trajectories from real-world driving scenarios and simulation data under extremely hazardous conditions. Simulation results demonstrate that H-DDPG achieves superior control performance compared to cooperative adaptive cruise control (CACC), MPC, and DDPG.