Proceedings of the 4th International Conference on Public Management and Intelligent Society, PMIS 2024, 15–17 March 2024, Changsha, China

Research Article

Transit Signal Priority Control Method Based on Deep Reinforcement Learning

Download26 downloads
  • @INPROCEEDINGS{10.4108/eai.15-3-2024.2346425,
        author={Junnan  Chen and Xufei  Zhuang and Heng  Li and Chenxi  Yang},
        title={Transit Signal Priority Control Method Based on Deep Reinforcement Learning},
        proceedings={Proceedings of the 4th International Conference on Public Management and Intelligent Society, PMIS 2024, 15--17 March 2024, Changsha, China},
        publisher={EAI},
        proceedings_a={PMIS},
        year={2024},
        month={6},
        keywords={intelligent transportation; transit signal priority control; deep reinforcement learning; sumo simulation},
        doi={10.4108/eai.15-3-2024.2346425}
    }
    
  • Junnan Chen
    Xufei Zhuang
    Heng Li
    Chenxi Yang
    Year: 2024
    Transit Signal Priority Control Method Based on Deep Reinforcement Learning
    PMIS
    EAI
    DOI: 10.4108/eai.15-3-2024.2346425
Junnan Chen1, Xufei Zhuang1,*, Heng Li1, Chenxi Yang1
  • 1: Inner Mongolia University of Technology
*Contact email: zxf@imut.edu.cn

Abstract

Transit signal priority control plays a pivotal role in enhancing the efficiency of public transportation and mitigating traffic congestion. Despite advancements in current research on transit signal priority control using deep reinforcement learning, the field is still nascent, encountering challenges like limited model generalization and slow training convergence, particularly in intricate traffic environments. This paper proposes a transit signal priority control model based on deep reinforcement learning. The model utilizes a deep neural network framework, integrating crucial optimizations like Dueling DQN, Distributional DQN and the PER mechanism, alongside enhancements to the loss function, to accommodate diverse traffic scenarios. The model incorporates essential information at intersections, encompassing vehicle positions, velocities, and lane conditions as input states to comprehensively depict the traffic situation. Concurrently, by delineating a set of signal phase configurations at intersections, a reward function is devised that considers various factors, such as intersection passage efficiency, bus delays, and passenger experience. This guides the agent to optimize the entire system throughout the learning process, duly considering signal control strategies to ensure effective decision-making.