基于深度强化学习的端到端无人驾驶决策

端到端的驾驶决策是无人驾驶领域的研究热点.本文基于DDPG（Deep Deterministic Policy Gradient）的深度强化学习算法对连续型动作输出的端到端驾驶决策展开研究.首先建立基于DDPG算法的端到端决策控制模型，模型根据连续获取的感知信息（如车辆转角，车辆速度，道路距离等）作为输入状态，输出车辆驾驶动作（加速，刹车，转向）的连续型控制量.然后在TORCS（The Open Racing Car Simulator）平台下不同的行驶环境中进行训练并验证，结果表明该模型可以实现端到端的无人驾驶决策.最后与离散型动作输出的DQN（Deep Q-learning Network）模型进行对比分析，实验结果表明DDPG决策模型具有更优越的决策控制效果.

End-to-End Autonomous Driving Decision Based on Deep Reinforcement Learning

The end-to-end driving decision making is a research hotspot in the field of autonomous driving.This paper studies the end-to-end driving decision of continuous action output based on DDPG (Deep Deterministic Policy Gradient) deep reinforcement learning algorithm.First,an end-to-end decision-making control model based on DDPG algorithm is established.The model outputs the continuous control quantity of vehicle driving action (acceleration,braking,steering) according to the continuously acquired perception information (such as vehicle angle,vehicle speed,road distance,etc.) as the input state.Then,the model is trained and verified in different driving environments on the platform of TORCS (The Open Racing Car Simulator).The results show that the model can realize the end-to-end decision-making of autonomous driving.At last,it is compared with DQN(Deep Q-Learning Network) model of discrete action output.The experimental results show that DDPG model has better decision control effect.