| 摘 要: 针对自动驾驶中复杂跟驰场景,提出了一种基于优先经验回放的双延迟深度确定性策略梯度(Twin Delayed Deep Deterministic Policy Gradient with Prioritized Experience Replay,TD3-PER)算法的车辆跟驰控制策略。该策略结合安全性、舒适性和效率设计奖励函数,并引入容忍范围机制和优先经验回放算法,以提升适应性和训练效率。通过PreScan/Simulink仿真验证,结果表明:相比模型预测控制(Model Predictive Control,MPC)、深度Q网络(DeepQ-Network,DQN)和深度确定性策略梯度(Deep Deterministic Policy Gradient,DDPG),该策略在安全隐患方面分别降低65.64%、31.14%和48.78%;行驶效率提升15.85%、17.24%和12.24%;舒适性改善16.67%、61.66%和7.77%。紧急制动场景下,在安全隐患方面分别降低68.06%、43.96%和22.96%,展现出优越的控制性能和适应性。 |
| 关键词: 智能交通 自动驾驶 跟驰模型 深度强化学习 优先经验回放 |
|
中图分类号:
文献标识码: A
|
| 基金项目: 江苏省自然科学基金项目(BK20171303) |
|
| Research on Intelligent Vehicle Car-Following Strategy Based on TD3-PER |
|
LIU Zhimin1,2, LU Dang1, GAO Xiong1
|
(1. New Energy Vehicle Motion Control Research Institute, Fujian University of Technology, Fuzhou 350118, China;
2. School of Electronic, Electrical Engineering and Physics, Fujian University of Technology, Fuzhou 350118, China)
1009439938@qq.com; ludang@jlu.edu.cn; 62202303@fjut.edu.cn
|
| Abstract: To address complex ca-r following scenarios in autonomous driving, this study proposes a vehicle ca-r following control strategy using the Twin Delayed Deep Deterministic policy gradient with Prioritized Experience Replay (TD3-PER) algorithm. The strategy designs a reward function integrating safety, comfort, and efficiency, while introducing a tolerance range mechanism and prioritized experience replay to enhance adaptability and training efficiency. Simulations via PreScan/Simulink demonstrate that compared to MPC, DQN, and DDPG, the proposed strategy achieves 65.64% , 31.14% , and 48.78% reduction in safety risks; 15.85% , 17.24% , and 12.24% improvement in driving efficiency; and 16.67% , 61.66% , and 7.77% enhancement in comfort respectively. In emergency braking
scenarios, it further reduces safety risks by 68.06% , 43.96% , and 22.96% , showcasing superior control performance and adaptability. |
| Keywords: intelligent transportation autonomous driving car following model deep reinforcement learning prioritized experience replay |
