基于换道意图和强化学习的车辆匝道汇入决策

Vehicle On-ramp Decision-making Based on Lane-change Intention and Reinforcement Learning

对高速公路匝道汇入决策开展了基于换道意图识别和深度强化学习的研究。首先,根据目标车辆及其周围车辆的特征参数,识别车辆换道意图并设计环境状态空间。其次,根据车辆的离散横向换道行为与连续纵向加减速行为,设计了混合的动作空间。然后,基于车辆的乘坐舒适度、通行效率、行车安全和汇入成功四方面,设计了环境奖励函数;依据匝道汇入任务特点,设计了环境终止条件。最后,基于实际道路和城市交通仿真软件搭建匝道汇入仿真平台来验证模型的可行性,多组对比分析实验结果表明,所提出方法在碰撞率和成功率上均表现最优,并且模型在不同道路车流密度条件下均取得90%以上的成功率。今后将对不同场景开展更为鲁棒的决策。

A study based on lane change intention recognition and deep reinforcement learning was conducted for the decision-making of highway on-ramps.Firstly,the study identified the lane-changing intentions of vehicles based on the characteristic parameters of the target vehicle and its surrounding vehicles,while also designing the environmental state space.Secondly,a mixed action space was developed that incorporates both discrete lane-changing behaviors and continuous longitudinal acceleration and deceleration behaviors of the vehicle.Subsequently,an environmental reward function was formulated based on four key aspects:vehicle ride comfort,traffic efficiency,driving safety,and successful merging.In consideration of the characteristics inherent in ramp merging and lane-changing tasks,environmental termination conditions were established.Finally,a ramp merging simulation platform was constructed,utilizing actual road data and the Simulation of Urban MObility framework,to validate the feasibility of the proposed model.The results of multiple comparative analysis experiments indicate that the proposed method outperforms others in terms of collision rate and success rate,and the model demonstrates a superior success rate across varying road traffic density conditions.In the future,more robust decision-making will be carried out for different scenarios.