Scalable Scheduling of Semiconductor Packaging Facilities Using Deep Reinforcement Learning.

Area of Science:

• Artificial Intelligence
• Operations Research
• Manufacturing Engineering

Background:

• Scheduling semiconductor operations is complex due to large scale and diverse job types.
• Traditional methods face challenges with increasing learning complexity and function approximation in reinforcement learning (RL).

Purpose of the Study:

• To develop an efficient deep reinforcement learning (RL) method for scheduling semiconductor packaging facilities.
• To address the challenges of large-scale scheduling problems with diverse job types and dynamic shop floor conditions.

Main Methods:

• A centralized deep RL agent is proposed for job allocation to machines.
• A novel state representation is introduced to handle variable machine availability and production requirements.
• A continuous action representation is utilized to manage a dynamic action space.

Main Results:

• The proposed RL method demonstrates superior performance compared to metaheuristics, rule-based methods, and other RL approaches.
• The method significantly reduces makespan in semiconductor scheduling tasks.
• The approach requires substantially less computation time than traditional metaheuristics.

Conclusions:

• The developed deep RL method offers an effective solution for complex semiconductor scheduling problems.
• The novel state and action representations enhance the scalability and adaptability of RL for manufacturing.
• This approach provides a computationally efficient and high-performing alternative for optimizing semiconductor operations.