Extended Invalid Action Mask: Training-Time Size-Agnostic Safe Reinforcement Learning for Flexible Job Shop Scheduling Problems

The flexible job shop scheduling problem (FJSP) is an NP-hard problem in which machine assignments are additionally considered compared to the classical job shop scheduling problem. It has received increased attention in recent years, as its global hard constraints highlight the importance of policy safety, especially in real production contexts. Deep reinforcement learning (DRL) has become a popular method for solving FJSP; however, ensuring both size-agnostic and safety-guaranteed properties during training remains challenging. Penalty-based approaches allow infeasible actions, degrading policy quality, while standard invalid action masking prevents training across different problem sizes. To address this, we propose an extended action-masking mechanism that maintains semantic consistency across varying scales while strictly enforcing action feasibility. Our framework integrates state and action embeddings to support adaptive scheduling and multi-scale training. Empirical results demonstrate improved performance over the latest DRL-based methods, indicating that our approach is scalable and robust for production scheduling.