基于区间分块Q学习的智能车辆安全舒适刹车算法  

Interval-block-based Q-learning algorithm for safe andcomfortable braking of intelligent vehicles

作　　者：余欣磊 周贤文[1] 张依恋 顾伟[1] Yu Xinlei;Zhou Xianwen;Zhang Yilian;Gu Wei(Key Laboratory of Transport Industry of Marine Technology&Control Engineering,Shanghai Maritime University,Shanghai 201306,China)

机构地区：[1]上海海事大学航运技术与控制工程交通行业重点实验室,上海201306

出　　处：《计算机应用研究》2024年第1期183-187,共5页Application Research of Computers

基　　金：国家自然科学基金面上项目(62176150);上海市地方院校能力建设项目(20040501400)。

摘　　要：针对当前智能汽车刹车场景下的安全与舒适性问题,提出一种基于区间分块的Q学习算法。首先在Q表中将前车加速度以一定间隔划分入等长区间,用区间中值做间隔来划分后车加速度。其次通过在安全条件下与加速度呈负相关的奖励设置,使智能体在保证安全的前提下尽量降低刹车加速度。最后在智能体训练的过程中遵循ε-贪心策略以减少随机性,在训练完毕后遵循贪心策略以最大程度利用智能体。将提出的算法与传统Q学习算法在三种常见道路场景上进行仿真测试。实验结果显示使用提出算法的智能车辆在刹车场景中安全率100%、平均刹车加速度小于2 m/s~2且能处理连续刹车加速度,表明提出的算法能够在确保智能汽车安全刹车的同时实现较低的刹车加速度。同时在连续刹车加速度与离线环境等复杂情况下,算法均能正常使用。To address the safety and comfort issues in intelligent vehicle braking scenarios,this paper proposed a Q-learning algorithm based on interval partitioning.Firstly,the algorithm divided the acceleration of the preceding vehicle into equal-length intervals with a certain interval in the Q-table,and used the interval median to partition the acceleration of the following vehicle.Secondly,the algorithm used a reward function that was negatively correlated with acceleration under safe conditions to encourage the agent to minimize braking acceleration while ensuring safety.Finally,the algorithm followed theε-greedy strategy during the training of the agent to reduce randomness,and followed the greedy strategy after training to maximize the utilization of the agent.This paper simulated the proposed algorithm and the traditional Q-learning algorithm on three common road scenarios.The experimental results show that the intelligent vehicle used the proposed algorithm has a 100%safety rate in braking scenarios,with an average braking acceleration of less than 2 m/s 2,and can handle continuous braking acceleration,which indicates that the proposed algorithm can achieve lower braking deceleration to improve passengers’comfort while ensuring safe braking of the intelligent vehicles.In addition,the algorithm is effective in complex scenarios including continuous braking deceleration and offline environments.

关 键 词：智能汽车 智能刹车 Q学习 区间分块 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]