An Adaptive Meta-Reinforcement Learning Framework for Dynamic Flexible Job Shop Scheduling

作者：Lincong Wu, Xiaoxia Li, Xin Lu, Zaiwen Feng, Yanguo Jing · 发表于：IEEE Transactions on Automation Science and Engineering · 年份：2025 · DOI：10.1109/tase.2025.3627282 · 被引用次数：4 · 研究领域：Scheduling and Optimization Algorithms、Digital Transformation in Industry、Flexible and Reconfigurable Manufacturing Systems

The optimization of flexible job shop scheduling is essential for improving manufacturing efficiency and performance in dynamic production environments. However, existing scheduling methods face challenges in scalability and adaptability, which limits their effectiveness in such environments. To address these limitations, this paper proposes a generalized and modular DFJSP framework that systematically decomposes shop-floor elements into key modules, enabling flexible and resilient scheduling under dynamic conditions. Building on this framework, an Adaptive Markov Decision Process (AMDP) is formulated to capture real-time shop-floor states and guide optimal action selection. Leveraging Meta-Reinforcement Learning (MRL), the proposed approach integrates Model-Agnostic Meta-Learning (MAML) with Proximal Policy Optimization (PPO) to facilitate rapid adaptation to new scheduling tasks while enhancing policy generalization. Numerical experiments demonstrate that the framework effectively balances multiple dynamic objectives, including makespan and energy consumption, and adapts efficiently to real-time variations in job priorities, machine availability, and processing times. The results highlight the potential of combining modular problem formulation, AMDP modeling, and MRL for scalable, efficient, and robust DFJSP solutions in modern manufacturing environments.