摘要
常规自动驾驶仪中使用的人工网络结构的弯道预判决策性能差,导致行驶过程中误差较大,因此,设计一种基于深度强化学习的自动驾驶仪控制软件。首先在Simulink平台上建立自动驾驶仪控制律模型,在直路行驶模块中控制输出指令,保证方向盘控制系统处于静止状态,在转弯行驶模块输入预设里程与行驶速度,经过大量行驶轨迹训练中进化自身在转弯处的行驶速度规划;深度强化学习神经算法采用的是NEAT算法,其中包含的进化思想需要搭配算法库的配置信息才能实现控制精准化。在仿真实验中对常规控制软件和设计的软件进行控制误差对比,结果表明设计的软件在轨道偏差与角度偏差方面的误差均有减少,验证了设计软件的控制精准度有一定的改进。
The artificial network structure used in conventional autopilot has poor curve prediction and decision performance, which leads to large error in the driving process. Therefore, a kind of autopilot control software based on deep reinforcement learning is designed. Firstly, the autopilot control law model was established on Simulink platform, and the output instructions were controlled in the straight road driving module to ensure that the steering wheel control system was in a static state. The preset mileage and driving speed were input in the turning driving module. After a lot of driving track training, the driving speed planning of the autopilot at the turn was evolved. The neural algorithm of deep reinforcement learning adopts NEATalgorithm, which contains the evolutionary idea that requires the configuration information of algorithm library to realize the precision control. In the simulation experiment, the control errors of the conventional control software and the designed software are compared. The results show that the errors of the designed software in orbit deviation and Angle deviation are reduced, which verifies that the control accuracy of the designed software has been improved to a certain extent.
作者
张震
ZHANG Zhen(Dalian Ocean Unwersity,Dalian Liaoning 116300,China)
出处
《自动化与仪器仪表》
2021年第10期53-56,共4页
Automation & Instrumentation
基金
辽宁省职业技术教育学会科研规划项目:供给侧改革背景下大学生创新创业能力培养的路径研究(No.LZY17573)。
关键词
深度强化学习
自动驾驶仪
控制软件
Deep reinforcement learning
Autopilot
Control software