摘要
目前虚拟环境中的无人驾驶车辆运动行为的智能训练主要通过无人驾驶车辆和环境的不断交互完成强化学习决策任务,存在无人驾驶车辆训练速度较慢且效率低下的问题.针对此问题,本文将车辆运动模拟技术和CMA-ES算法结合,提出了一种虚拟环境中基于改进协方差自适应算法(Covariance Matrix Adaptive Evolutionary Strategy,CMA-ES)算法的无人驾驶车辆运动行为智能训练方法,首先模拟车辆的运动行为,得到特殊路段的运动行为和环境信息数据;然后对模拟得到的环境信息数据进行编码预处理、分组;最后拟合无人驾驶车辆运动行为策略神经网络的权重参数,实现对场景的预学习,利用车辆运动模拟技术模拟车辆在复杂路段的运动行为作为相关经验,指导无人驾驶车辆运动行为强化学习的过程,实验表明该方法加快了无人驾驶车辆运动行为的训练速度.
At present,the intelligent training of driverless vehicle motion behavior in the virtual environment mainly completes the reinforcement learning decision-making task through the continuous interaction between the driverless vehicle and the environment,and there is a problem that the training speed of driverless vehicles is slow and the efficiency is low.To solve this problem,this paper proposes an intelligent training method of driverless vehicle motion behavior based on improved CMA-ES algorithm in virtual environment.Firstly,the motion behavior of vehicle is simulated to obtain the motion behavior and environmental information data of special road sections.Then,the simulated environmental information data are preprocessed and grouped.Finally,the weight parameters of the neural network of driverless vehicle motion behavior strategy are fitted to realize the pre-learning of the scene.The vehicle motion simulation technology is used to simulate the vehicle motion behavior in complex road sections as relevant experience to guide the process of reinforcement learning of driverless vehicle motion behavior.Experiments show that this method speeds up the training speed of driverless vehicle motion behavior.
作者
马亚丹
邹佳丽
MA Yadan;ZOU Jiali(School of Electronic Information Engineering,Zhengzhou Sias University,Zhengzhou 451100,China)
出处
《湖北民族大学学报(自然科学版)》
CAS
2021年第3期338-345,共8页
Journal of Hubei Minzu University:Natural Science Edition
基金
国家自然科学基金项目(62072415).
关键词
车辆运动模拟
无人驾驶车辆运动行为训练
神经进化
虚拟仿真
强化学习
vehicle motion simulation
driverless vehicle sports behavior training
neuroevolution
virtual simulation
reinforcement learning