Scholarlink

While existing research has made progress in optimizing the route planning performance for a small number of vehicles, it often falls short of adequately considering the dynamic interactions and mutual influence among vehicles in large-scale vehicular networks. This deficiency limits its effectiveness in addressing urban traffic congestion and enhancing the overall efficiency of the transportation system. Therefore, in this paper, we propose a multi-agent mean field reinforcement learning (MAMFRL) framework for large-scale vehicle path planning problems, aiming to improve the efficiency of individual vehicles and the entire transportation system. Specifically, we first utilize mean field (MF) theory to simplify the interactions between agents. Second, MAMFRL employs a convolutional neural network (CNN) layer to extract road information features and obtain spatial correlations of urban traffic. Finally, MAMFRL constrains the rewards to improve the performance of the whole transportation system. Experimental results show that the proposed method can reduce average vehicle travel time by up to 9 \(\%\) and average intersection queue lengths by up to 27.8 \(\%\) .

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

Efficient Path Planning for Large-Scale Vehicular Networks via Multi-agent Mean Field Reinforcement Learning

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Tian Chen,
  • Wenbin Chen,
  • Huifei Gao

摘要

While existing research has made progress in optimizing the route planning performance for a small number of vehicles, it often falls short of adequately considering the dynamic interactions and mutual influence among vehicles in large-scale vehicular networks. This deficiency limits its effectiveness in addressing urban traffic congestion and enhancing the overall efficiency of the transportation system. Therefore, in this paper, we propose a multi-agent mean field reinforcement learning (MAMFRL) framework for large-scale vehicle path planning problems, aiming to improve the efficiency of individual vehicles and the entire transportation system. Specifically, we first utilize mean field (MF) theory to simplify the interactions between agents. Second, MAMFRL employs a convolutional neural network (CNN) layer to extract road information features and obtain spatial correlations of urban traffic. Finally, MAMFRL constrains the rewards to improve the performance of the whole transportation system. Experimental results show that the proposed method can reduce average vehicle travel time by up to 9 \(\%\) and average intersection queue lengths by up to 27.8 \(\%\) .