Scholarlink

<p>Modeling lane-changing maneuvers constitutes one fundamental component of microscopic traffic simulation. However, a few studies focus on truck lane changing, and existing models oversimplify execution paths and do not capture true lane-changing behavior. This paper seeks to improve accuracy in describing truck lane-changing behavior. It divides the process into lane change behavior planning and trajectory planning phases according to human cognition and subsequently proposes a hierarchical deep reinforcement learning model. Then, a lane-reward mechanism for trucks is designed, and a Dueling Double Deep Q-network with Prioritized Experience Replay (P-D3QN) is introduced to make lane-changing decisions. Next, a polynomial trajectory prediction model pre-trained on real data using supervised learning is combined with SAC to dynamically plan and update the truck’s longitudinal acceleration and lateral displacement. The proposed model is then simulated and validated in a typical three-lane straight freeway scenario. Simulation results indicate that, with the introduction of a lane-reward mechanism and a trajectory planning layer, the model not only ensures safety but also effectively promotes right-lane driving behavior for trucks. The resulting trajectories are smoother and more continuous, enabling more reasonable lane changes that better reflect real-world truck driving behavior.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Modeling Truck Lane Change Behavior on Highways Using a Hierarchical Reinforcement Learning Framework

  • 期刊论文

  • 发表日期：

    2025年8月18日

查看全文
  • Zhenxin Hu,
  • Dalin Qian,
  • Wenlong Xu,
  • Xinwu Yuan,
  • Sixian Li

摘要

Modeling lane-changing maneuvers constitutes one fundamental component of microscopic traffic simulation. However, a few studies focus on truck lane changing, and existing models oversimplify execution paths and do not capture true lane-changing behavior. This paper seeks to improve accuracy in describing truck lane-changing behavior. It divides the process into lane change behavior planning and trajectory planning phases according to human cognition and subsequently proposes a hierarchical deep reinforcement learning model. Then, a lane-reward mechanism for trucks is designed, and a Dueling Double Deep Q-network with Prioritized Experience Replay (P-D3QN) is introduced to make lane-changing decisions. Next, a polynomial trajectory prediction model pre-trained on real data using supervised learning is combined with SAC to dynamically plan and update the truck’s longitudinal acceleration and lateral displacement. The proposed model is then simulated and validated in a typical three-lane straight freeway scenario. Simulation results indicate that, with the introduction of a lane-reward mechanism and a trajectory planning layer, the model not only ensures safety but also effectively promotes right-lane driving behavior for trucks. The resulting trajectories are smoother and more continuous, enabling more reasonable lane changes that better reflect real-world truck driving behavior.