Scholarlink

<p>Achieving effective and intelligent dispatching has become a pressing task due to the ride-hailing industry’s rapid development. Traditional dispatching algorithms frequently only evaluate the distance between present passengers and idle drivers, ignoring future supply and demand conditions, which might increase idle miles and uneven resource allocation. To address this, paper proposes a multi-agent reinforcement learning-based dispatching optimization model (FAMD-MARL), which considers potential future orders and idle vehicles within a certain time frame. The model introduces a “future expected Q-value” mechanism to assist multi-agent decision-making by comparing immediate and future Q-values. The model’s effectiveness was validated using ride-hailing order data from Haikou City. Compared to traditional greedy matching, single-agent PPO, and random strategies, the proposed model reduced average passenger waiting time by approximately 25%–40%, decreased driver idle mileage by about 20%, improved overall system service rate by 5%–15%, and increased driver earnings by approximately 40%–80%.</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

Ride-Hailing order Dispatching Based on Multi-Agent Reinforcement Learning

  • 期刊论文

  • 发表日期：

    2026年2月6日

查看全文
  • Li Yang,
  • Wei Zhang

摘要

Achieving effective and intelligent dispatching has become a pressing task due to the ride-hailing industry’s rapid development. Traditional dispatching algorithms frequently only evaluate the distance between present passengers and idle drivers, ignoring future supply and demand conditions, which might increase idle miles and uneven resource allocation. To address this, paper proposes a multi-agent reinforcement learning-based dispatching optimization model (FAMD-MARL), which considers potential future orders and idle vehicles within a certain time frame. The model introduces a “future expected Q-value” mechanism to assist multi-agent decision-making by comparing immediate and future Q-values. The model’s effectiveness was validated using ride-hailing order data from Haikou City. Compared to traditional greedy matching, single-agent PPO, and random strategies, the proposed model reduced average passenger waiting time by approximately 25%–40%, decreased driver idle mileage by about 20%, improved overall system service rate by 5%–15%, and increased driver earnings by approximately 40%–80%.