Flexible Job Shop Scheduling via Dual Attention Network-Based Reinforcement Learning

Area of Science:

• Operations Research
• Artificial Intelligence
• Manufacturing Systems Engineering

Background:

• Flexible job shop scheduling problems (FJSP) present significant complexity due to operations being processable on multiple machines.
• Existing deep reinforcement learning (DRL) methods for FJSP, while promising, often yield solutions with suboptimal quality compared to exact methods.
• There is a need for advanced methods to effectively capture intricate operation-machine relationships in FJSP.

Purpose of the Study:

• To develop a novel end-to-end learning framework for solving the flexible job shop scheduling problem (FJSP).
• To enhance decision-making in FJSP by integrating self-attention mechanisms with deep reinforcement learning.
• To improve the quality and scalability of solutions for complex manufacturing scheduling.

Main Methods:

• A dual-attention network (DAN) is proposed, incorporating interconnected operation and machine attention blocks for deep feature extraction.
• The DAN leverages self-attention to precisely model complex relationships between operations and machines.
• Deep reinforcement learning (DRL) is employed for scalable decision-making, guided by features extracted from the DAN.

Main Results:

• The proposed DAN-DRL framework significantly outperforms traditional priority dispatching rules (PDRs) and existing state-of-the-art DRL methods on synthetic and benchmark FJSP datasets.
• The approach achieves solution quality comparable to exact methods (e.g., OR-Tools) in specific scenarios.
• Demonstrated favorable generalization capabilities on large-scale and real-world unseen FJSP instances.

Conclusions:

• The novel end-to-end learning framework effectively addresses the limitations of current DRL approaches for FJSP.
• The dual-attention network successfully captures complex production dynamics, leading to high-quality scheduling decisions.
• This method offers a scalable and effective alternative for complex manufacturing scheduling, approaching the performance of exact methods.