该文提出了基于分层柔性演员-评论家（SAC）强化学习的交叉口信号配时-车辆轨迹联合优化方法。模型包括信号配时优化层和车辆轨迹优化层，两者的状态空间均包含车辆位置、速度和信号配时状态，奖励函数均为交通效率、安全指标和油耗的加权和。信号配时优化层的动作为信号相位持续时间，车辆轨迹优化层的动作为车辆加速度。两个优化层分别具有独立的价值网络和策略网络。价值网络根据当前状态和动作输出当前状态-动作价值，评估策略网络性能；策略网络基于当前状态生成高斯分布的均值和标准差，并从参数化的高斯分布中采样动作。在策略网络损失函数中引入熵系数和温度系数，自动调节策略探索的广度和深度，降低模型训练性能对超参数变化的敏感度。针对信号配时优化和车辆轨迹优化间隔不一致的问题，设计信号配时层-车辆轨迹优化层异步训练算法，通过反向传播算法，同时对同一层的价值网络和策略网络进行训练。此外，利用SUMO对模型进行训练和评估，实验结果表明，与数学规划模型、仅优化信号配时和仅优化车辆轨迹的模型相比，提出的模型可使车辆油耗分别平均降低24.24%、5.39%和22.23%，在基本不降低平均速度的情况下实现能耗优化，且在状态观测扰动下仍能保持5%以内的性能偏差，具有鲁棒性。

This study proposes a joint optimization method for intersection signal timing and vehicle trajectory based on the Soft Actor-Critic (SAC) reinforcement learning framework. The model consists of two layers: signal ti-ming optimization and vehicle trajectory optimization. The state space for both layers includes vehicle position, speed, and traffic signal status, while the reward function is a weighted sum of traffic efficiency, safety, and fuel consumption. In the signal timing optimization layer, the action is the duration of the signal phase, and in the vehicle trajectory optimization layer, the action is the vehicle acceleration. Each optimization layer has independent value networks and policy networks. The value network outputs the state-action value and assesses the policy network’s performance according to the current state and action. The policy network generates the mean and standard deviation of a Gaussian distribution based on the current state and samples actions from this parameterized Gaussian distribution. The loss function of the policy network includes entropy and temperature coefficients to automatically adjust the breadth and depth of policy exploration, reducing the model’s sensitivity to hyperparameter variations. To address the inconsistency in the intervals between signal timing optimization and vehicle trajectory optimization, this study designed an asynchronous training algorithm for the signal timing layer and vehicle trajectory optimization layer. Both the value network and the policy network of the same layer were trained simultaneously using backpropagation. The model was trained and evaluated with SUMO, and experimental results indicate that the proposed method reduces vehicle fuel consumption by an average of 24.24%, 5.39%, and 22.23%, compared to mathematical programming methods, signal-timing-only optimization methods, and trajectory-only optimization methods, respectively. It can achieve energy optimization without significantly reducing average speed, while maintaining performance deviations within 5% under state observation disturbances, demonstrating good robustness.