Probing a novel dual-phase adaptive deep learning approach for dynamic flexible job shop scheduling

In the research area of the Dynamic Flexible Job-shop Scheduling Problem (DFJSP), real-time responsiveness to disturbances and solution quality are the key factors of dynamic scheduling performance. Traditional methods can not offer ideal dynamic scheduling performance as they either lack real-time responsiveness to disturbances or sacrifice solution quality. In contrast, scheduling methods based on Reinforcement Learning (RL) allow the agent to interact with a dynamic environment and update the scheduling strategy in real-time, resulting in high real-time responsiveness and strong adaptability. However, relying solely on dynamic decision-making experience will limit the quality of the solution. Therefore, this paper proposed a novel Dual-Phase Adaptive Learning Framework (DPALF) that introduces a pretraining phase with Imitation Learning (IL) to equip the decision-making agent with the static scheduling capability before RL in the disturbed scheduling environment. Experimental results show that introducing the pretraining phase improves learning performance and enhances the stability of the formal training process. Additionally, the proposed method outperforms three representative RL methods in solution quality, proving more effective for real-time dynamic scenarios.