基于深度强化学习的自主换道控制模型

Autonomous lane change model with integrated control based on deep reinforcement learning

为解决自动驾驶汽车快速安全换道问题,提出并改进了一种基于深度强化学习的自主换道控制模型。首先建立车辆动力学运动模型,其次使用深度确定性策略梯度(DDPG)算法更新模型,最后通过MATLAB/CarSim对学习到的控制策略进行联合仿真验证。为了使模型更真实可靠,提出将CarSim融入智能体的训练,同时为解决传统模型在换道后期控制效果不理想问题,提出一种基于采样时间的方向盘转角输出模型。结果表明:在60、80 km/h车速下,提出的模型从换道开始到稳定行驶的过程相比于改进前更平顺、快速,验证了模型能够实现一般车速下的自主换道控制,为车辆的自主换道研究提供一定的参考。

In order to solve the problem of fast and safe lane change of autonomous vehicle,an autonomous lane change control model based on deep reinforcement learning was proposed and improved.Firstly,the vehicle dynamic motion model was established,secondly,the depth deterministic strategy gradient(DDPG)algorithm was used to update the model,and finally,the learned control strategy was co-simulated by MATLAB/CarSim software.In order to make the model more realistic and reliable,CarSim software was integrated into the training of the agent.Meanwhile,a steering wheel angle output model based on sampling time was proposed to solve the problem that the control effect of the traditional model was not ideal in the later period of lane change.The results showed that under the speed of 60 km/h and 80 km/h,the control process from lane change to stable running of the model was smoother and faster than before the improvement,which verified that the model could realize the autonomous lane change control under the general speed,and had certain significance for the research of autonomous lane change of vehicles.