Graph-Augmented Deep Reinforcement Learning for Key Feature Extraction in FJSP
Jiahui Du, Fangyu Li · 2025
The Flexible Job Shop Scheduling Problem (FJSP) represents a significant optimization challenge in modern manufacturing, requiring efficient job allocation and machine selection under complex constraints. Traditional methods struggle to handle high-dimensional state spaces and identify critical features necessary for optimal decision-making, resulting in limitations in performance and stability. We propose a scheduling method based on deep reinforcement learning to extract key features from the state space using Graph Isomorphism Networks (GIN) and the information bottleneck (IB) principle to provide efficient scheduling decisions for FJSP. First, we leverage GIN to model the scheduling environment and capture complex relationships between jobs and machines. Second, we integrate IB to eliminate redundant state information and retain decision-relevant features. Third, we formulate the scheduling process as a Markov Decision Process (MDP) and apply Proximal Strategy Optimization (PPO) for robust policy learning. The experiment results show that the proposed method performs excellently minimizing makespan and ensures stability across different datasets at the same time.