Scholarlink

<p>As the number of motor vehicles in cities continues to increase, the probability of traffic congestion is increasing, and the pressure of management department is growing. Artificial intelligence technology has promoted intelligent transportation construction. The traffic signal control method based on deep reinforcement learning has recently become a mainstream solution to ease traffic congestion. In this paper, we propose an action value weighted QMIX (AVW-QMIX) based cooperative traffic signal control method, which models a multi-intersection traffic network in a region as a multi-agent reinforcement learning system, and uses each local signal agent to sense the traffic flow information of its incoming lane. The algorithm uses a QMIX fusion network supplemented with global vehicle state information to fit the value function of the intra-regional agents to the overall joint action value function. The weights generated by the independent agent guidance network and the mixing network are also used to guide the updating of the QMIX network parameters to mitigate the negative impact of the convergence of the original QMIX algorithm to suboptimal results. AVW-QMIX also adds the distribution of action values and signal phase states of neighboring agents at the previous time step in the process of policy learning to facilitate communication and coordination among intersections. Finally, simulation experiments in Vissim demonstrate the superior performance of the AVW-QMIX-based signal control method. Compared to the suboptimal baseline, it reduces the overall average delay by 1.54–4.32%, the average queue length by 2.24–3.36%, and the number of stops by 1.25–4.21%, with particularly significant gains under high-traffic peak conditions.</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

Area traffic signal coordination control method based on action value weighted QMIX

  • 期刊论文

  • 发表日期：

    2026年7月23日

查看全文
  • Huizhen Zhang,
  • Youqing Chen,
  • Minglei Liu,
  • Zhixian Ling,
  • Jiatian Xu,
  • Zhenwei Fang,
  • Wenjuan Feng

摘要

As the number of motor vehicles in cities continues to increase, the probability of traffic congestion is increasing, and the pressure of management department is growing. Artificial intelligence technology has promoted intelligent transportation construction. The traffic signal control method based on deep reinforcement learning has recently become a mainstream solution to ease traffic congestion. In this paper, we propose an action value weighted QMIX (AVW-QMIX) based cooperative traffic signal control method, which models a multi-intersection traffic network in a region as a multi-agent reinforcement learning system, and uses each local signal agent to sense the traffic flow information of its incoming lane. The algorithm uses a QMIX fusion network supplemented with global vehicle state information to fit the value function of the intra-regional agents to the overall joint action value function. The weights generated by the independent agent guidance network and the mixing network are also used to guide the updating of the QMIX network parameters to mitigate the negative impact of the convergence of the original QMIX algorithm to suboptimal results. AVW-QMIX also adds the distribution of action values and signal phase states of neighboring agents at the previous time step in the process of policy learning to facilitate communication and coordination among intersections. Finally, simulation experiments in Vissim demonstrate the superior performance of the AVW-QMIX-based signal control method. Compared to the suboptimal baseline, it reduces the overall average delay by 1.54–4.32%, the average queue length by 2.24–3.36%, and the number of stops by 1.25–4.21%, with particularly significant gains under high-traffic peak conditions.