Scholarlink

<p>Recently, various methods have been explored to address the challenges of solving the large-scale flexible Job-shop scheduling problem (FJSP), where operations can be scheduled on multiple machines, posing state representation and decision-making difficulties. However, existing approaches struggle to handle complex scenarios effectively. To tackle these challenges, this paper introduces an innovative end-to-end MARLSIO framework based on multi-agent reinforcement learning (MARL) and structural information optimization. The proposed strategy decomposes the FJSP into two sub-issues: job operation selection and machine selection. To facilitate the allocation of operations to machines, the presented framework employs the multi-agent proximal policy optimization algorithm to train the job agents. Every agent generates actions according to global state and its observed respective state. The decision-making process for each job agent involves combining the machine selection and operation assignment into a composite decision. Furthermore, a novel structural information representation of scheduling states is introduced, enabling the architecture to describe detailed connections between machines and operations. This enhances the agents’ learning efficiency. Experimental results demonstrate that the proposed method outperforms the traditional approaches and exhibits computational efficiency, even when applied to instances of larger scales and distinct characteristics not encountered during training.</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

Structural entropy-based scheduler for job planning problems using multi-agent reinforcement learning

  • 期刊论文

  • 发表日期：

    2025年1月12日

查看全文
  • Lixin Liang,
  • Shuo Sun,
  • Zhifeng Hao,
  • Yong Yang

摘要

Recently, various methods have been explored to address the challenges of solving the large-scale flexible Job-shop scheduling problem (FJSP), where operations can be scheduled on multiple machines, posing state representation and decision-making difficulties. However, existing approaches struggle to handle complex scenarios effectively. To tackle these challenges, this paper introduces an innovative end-to-end MARLSIO framework based on multi-agent reinforcement learning (MARL) and structural information optimization. The proposed strategy decomposes the FJSP into two sub-issues: job operation selection and machine selection. To facilitate the allocation of operations to machines, the presented framework employs the multi-agent proximal policy optimization algorithm to train the job agents. Every agent generates actions according to global state and its observed respective state. The decision-making process for each job agent involves combining the machine selection and operation assignment into a composite decision. Furthermore, a novel structural information representation of scheduling states is introduced, enabling the architecture to describe detailed connections between machines and operations. This enhances the agents’ learning efficiency. Experimental results demonstrate that the proposed method outperforms the traditional approaches and exhibits computational efficiency, even when applied to instances of larger scales and distinct characteristics not encountered during training.