Scholarlink

This study proposes a deep reinforcement learning approach to enhance the efficiency of multi-agent pickup and delivery with immediate tasks (MAPD-I). MAPD-I is an extension of conventional MAPD to handle immediate and emergent tasks in real time that are inserted into a previously planned sequence of delivery tasks. To address this challenge, as the first step, we devise straightforward rule-based methods that prioritize immediate tasks in fixed-ways for rapid processing. Finally, we propose ProcrastiNet, a deep reinforcement learning model that determines the optimal degree of procrastination for immediate tasks execution. Although the rule-based methods tend to fail to account for the delays imposed on regular tasks, our approach can enhance overall efficiency by strategically delaying the handling of immediate tasks in a flexible manner. Through experimental evaluations, we demonstrated that ProcrastiNet successfully reduced the overall task-completion time while minimizing the urgency of immediate tasks, outperforming both the typical interrupt handling method and the rule-based approach.

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 Assignment of Immediate Tasks Using Deep Reinforcement Learning in Multi-Agent Pickup and Delivery

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Taisei Hirayama,
  • Kohei Yoshida,
  • Hiroki Sakaji,
  • Itsuki Noda

摘要

This study proposes a deep reinforcement learning approach to enhance the efficiency of multi-agent pickup and delivery with immediate tasks (MAPD-I). MAPD-I is an extension of conventional MAPD to handle immediate and emergent tasks in real time that are inserted into a previously planned sequence of delivery tasks. To address this challenge, as the first step, we devise straightforward rule-based methods that prioritize immediate tasks in fixed-ways for rapid processing. Finally, we propose ProcrastiNet, a deep reinforcement learning model that determines the optimal degree of procrastination for immediate tasks execution. Although the rule-based methods tend to fail to account for the delays imposed on regular tasks, our approach can enhance overall efficiency by strategically delaying the handling of immediate tasks in a flexible manner. Through experimental evaluations, we demonstrated that ProcrastiNet successfully reduced the overall task-completion time while minimizing the urgency of immediate tasks, outperforming both the typical interrupt handling method and the rule-based approach.