Scholarlink

<p>In flexible manufacturing systems, multi-load automated guided vehicles (AGVs) are widely used for material transfer and job fetching to improve transportation efficiency. Compared with unit-load AGVs, multi-load AGV scheduling should consider not only task assignment but also path planning to optimize the visiting sequence of multiple task points. Moreover, high-mix low-volume manufacturing and changing production plans bring great challenges to the real-time scheduling of multi-load AGVs. A two-stage dynamic scheduling method based on reinforcement learning (RL) is proposed to reduce average task delay and AGV travel costs. In the task assignment stage, a two-layer dueling double deep Q-network (D3QN) is utilized to select the optimal AGV dispatching rule and task selection rule according to the dynamic states of AGVs and tasks. In the path planning stage, an integer programming based multi-load AGV path planning method is applied to re-plan the visiting sequence of multiple task points with optimal task delay and travel costs. The effectiveness of the proposed dynamic scheduling method in terms of travel costs and task delay is illustrated by test examples based on randomly inserted tasks and different environment settings.</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

Dynamic Scheduling Based on Two-Layer Deep Reinforcement Learning for Multi-load AGVs

  • 期刊论文

  • 发表日期：

    2025年5月3日

查看全文
  • Gaoshang Wang,
  • Yuanyuan Zou,
  • Yaru Yang,
  • Shaoyuan Li

摘要

In flexible manufacturing systems, multi-load automated guided vehicles (AGVs) are widely used for material transfer and job fetching to improve transportation efficiency. Compared with unit-load AGVs, multi-load AGV scheduling should consider not only task assignment but also path planning to optimize the visiting sequence of multiple task points. Moreover, high-mix low-volume manufacturing and changing production plans bring great challenges to the real-time scheduling of multi-load AGVs. A two-stage dynamic scheduling method based on reinforcement learning (RL) is proposed to reduce average task delay and AGV travel costs. In the task assignment stage, a two-layer dueling double deep Q-network (D3QN) is utilized to select the optimal AGV dispatching rule and task selection rule according to the dynamic states of AGVs and tasks. In the path planning stage, an integer programming based multi-load AGV path planning method is applied to re-plan the visiting sequence of multiple task points with optimal task delay and travel costs. The effectiveness of the proposed dynamic scheduling method in terms of travel costs and task delay is illustrated by test examples based on randomly inserted tasks and different environment settings.