自动驾驶车辆多风格自适应跟驰决策  

Multi-Style Adaptive Following Decision Making for Automated Vehicles

作　　者：虢力源 潘为刚 高嵩 

机构地区：[1]山东交通学院信息科学与电气工程学院,山东 济南 [2]山东交通学院轨道交通学院,山东 济南

出　　处：《交通技术》2024年第6期403-411,共9页Open Journal of Transportation Technologies

摘　　要：本研究提出了一种适用于自动驾驶车辆的多风格自适应跟驰决策框架,结合深度强化学习(DRL)与改进的粒子群优化算法,实现了对不同驾驶风格的精确控制。该框架通过DRL算法设计自适应的跟驰速度控制策略,并基于驾驶安全性、效率、舒适性和紧急制动等因素构建了复杂的奖励机制。为进一步优化跟驰性能,本文在传统粒子群优化算法中引入了莱维扰动,以精确计算奖励函数中的最优权重组合,确保算法在不同驾驶场景下均能灵活应对。实验结果显示,该算法在多种复杂交通场景中具有鲁棒性,提供了更安全、高效且舒适的驾驶体验。In this study, a multi-style adaptive following decision framework for self-driving vehicles is proposed, which combines deep reinforcement learning (DRL) with an improved particle swarm optimization algorithm to achieve accurate control of different driving styles. The framework designs an adaptive following speed control strategy through the DRL algorithm. In order to further optimize the following speed performance, this paper introduces the Lévy perturbation into the tradi-tional particle swarm optimization algorithm to accurately calculate the optimal weight combinations in the reward function, ensuring that the algorithm can flexibly cope with different driving scenarios. Experimental results show that the algorithm is robust in multiple complex traffic scenarios, providing a safer, more efficient and comfortable driving experience.

关 键 词：跟车驾驶 速度控制 强化学习 深度确定性策略梯度 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]