仿驾驶员DDPG汽车纵向自动驾驶决策方法  

A Driver-like Decision-making Method for Longitudinal Autonomous Driving Based on DDPG

作　　者：高振海[1] 闫相同 高菲[1] 孙天骏 Gao Zhenhai;Yan Xiangtong;GaoFei;Sun Tianjun(Jilin University,State Key Laboratory of Automotive Simulation and Control,Changchun 130022)

机构地区：[1]吉林大学,汽车仿真与控制国家重点实验室,长春130022

出　　处：《汽车工程》2021年第12期1737-1744,共8页Automotive Engineering

基　　金：国家重点研发计划(2017YFB0102601);国家自然科学基金(51775236,U1564214);纵侧向运动控制软件开发国内技术采购项目(3R2210469415)资助。

摘　　要：汽车纵向自动驾驶的决策层根据车辆当前运动状态与环境信息,决策出理想的动作指令。目前如何在自动驾驶决策策略中考虑人类驾驶员的行为成为研究热点。在纵向自动驾驶决策策略中传统的基于规则的决策策略难以运用到复杂的场景中,而当前使用强化学习和深度强化学习的决策方法大多通过设计安全性、舒适性、经济性相关公式构建奖励函数,得到的决策策略与人类驾驶员相比仍然存在较大差距。针对以上问题,本文使用驾驶员数据通过BP神经网络拟合设计奖励函数,使用深度强化学习DDPG算法,建立了一种仿驾驶员的纵向自动驾驶决策方法。最终通过仿真测试验证了该方法的有效性和与驾驶员行为的一致性。The decision-making layer of vehicle longitudinal autonomous driving decides the ideal action instruction according to the current motion state of the vehicle and environmental information.At present,how to consider the behavior of human drivers in autonomous driving decision-making strategies has become a hotspot.In longitudinal autonomous driving decision-making strategies,traditional rule-based decision-making strategies are difficult to be applied to complex scenarios.Current decision-making methods use reinforcement learning and deep reinforcement learning to construct reward functions designed with safety,comfort,and economy formulas.The obtained decision-making strategy still has a big gap compared with that of the human driver.To solve the above problems,this paper uses driver data to design a reward function by BP neural network,and uses DDPG algorithm to establish a driver-like longitudinal autonomous driving decision-making method.Finally,the effectiveness of the method and the consistency with the driver's behavior are verified by simulation tests.

关 键 词：自动驾驶 决策算法 深度强化学习 深度确定性策略梯度 

分 类 号：U463.6[机械工程—车辆工程]