为解决无信号交叉口的智能车决策控制问题，以双向单车道交叉口下两车合流工况为对象，采用强化学习方法开展研究，建立车辆状态空间到动作空间的映射关系。针对目前研究中环境车辆车速设置过于简单的问题，以实际场景下采集的数据作为环境车辆的轨迹信息以构建场景模型。基于自回归滑动平均模型对环境车辆的车速进行预测。结合智能车以及所预测的环境车辆车速时序信息建立先行让行决策模型以计算本车参考车速，引入参考车速构建强化学习的奖励函数加速训练收敛速度。结果表明：所提出的强化学习模型具有较快的收敛速度，训练得到的智能体在与不同驾驶风格的环境车辆博弈时能够安全通过交叉口，为无信号交叉口智能车安全通行决策控制问题提供了参考依据。

In order to solve the problem of autonomous vehicle decision and control at an unsignalized intersection, this paper studies the confluence condition of two vehicles at a two-way single-lane intersection. The method of reinforcement learning is used for the decision and control of autonomous vehicle to establish the mapping relationship between vehicle state space and action space. Aiming at the problem that the vehicle speed setting is too simple in the present study, the data collected in the actual scene was used as the trajectory information of the environmental vehicle to construct the scene model. Based on the autoregressive moving average model, the vehicle speed of environmental vehicle is predicted. Combining with the autonomous vehicle and the predicted environmental vehicle speed timing information, the forward decision-making model is established to calculate the reference speed, and the reference speed is introduced to construct the reinforcement learning reward function to accelerate the training convergence speed. The results show that the enhanced learning model has a fast convergence speed, and the trained agent can safely pass through the intersection when gambling with environmental vehicles with different driving styles, this paper provides a reference for autonomous vehicle safety decision-making and control at no-signal intersection.

重庆市基础研究与前沿探索项目（CSTC2018JCYJAX0409）；四川省科技计划资助项目（2019 YFG0528）