Scholarlink

<p>Generalization is a critical requirement for reinforcement learning (RL) models applied in production planning tasks such as facility layout optimization. This study investigates whether a single trained RL agent can generate efficient layouts across a diverse set of facility layout problems without retraining. To achieve this, we develop a scalable modeling approach using fixed-size image-based state representation and a static, masked action space, enabling consistent input–output dimensions regardless of layout size. The agent is trained using Proximal Policy Optimization (PPO) on a subset of layout instances and evaluated on unseen problem configurations that differ in machine count and flow matrix. Material handling cost (MHC) is used as the primary performance metric, calculated as the weighted sum of Euclidean distances between machine pairs. Our results show that the trained agent significantly outperforms initial random layouts across most test cases, demonstrating meaningful generalization. This approach offers practical benefits for scalable deployment in industrial environments where retraining for each new layout instance is often infeasible. This work establishes a foundation for adaptive, data-driven facility planning using deep reinforcement learning.</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

From theory to application: investigating the generalizability of facility layout problems using a deep reinforcement learning approach

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年5月28日

查看全文
  • Benjamin Heinbach,
  • Peter Burggräf,
  • Fabian Steinberg

摘要

Generalization is a critical requirement for reinforcement learning (RL) models applied in production planning tasks such as facility layout optimization. This study investigates whether a single trained RL agent can generate efficient layouts across a diverse set of facility layout problems without retraining. To achieve this, we develop a scalable modeling approach using fixed-size image-based state representation and a static, masked action space, enabling consistent input–output dimensions regardless of layout size. The agent is trained using Proximal Policy Optimization (PPO) on a subset of layout instances and evaluated on unseen problem configurations that differ in machine count and flow matrix. Material handling cost (MHC) is used as the primary performance metric, calculated as the weighted sum of Euclidean distances between machine pairs. Our results show that the trained agent significantly outperforms initial random layouts across most test cases, demonstrating meaningful generalization. This approach offers practical benefits for scalable deployment in industrial environments where retraining for each new layout instance is often infeasible. This work establishes a foundation for adaptive, data-driven facility planning using deep reinforcement learning.