EKG-AC: A New Paradigm for Process Industrial Optimization Based on Offline Reinforcement Learning With Expert...

Area of Science:

• Process Control and Optimization
• Artificial Intelligence
• Machine Learning

Background:

• Operation optimization is vital for industrial product quality and profitability.
• Reinforcement learning (RL) offers autonomous learning but faces real-world constraints due to high costs and risks.
• Offline RL addresses interaction risks but struggles with imbalanced data and safety in industrial settings.

Purpose of the Study:

• To develop a novel offline reinforcement learning algorithm for industrial process optimization.
• To address challenges of imbalanced operating conditions and safety sensitivity in offline RL applications.
• To improve decision-making accuracy and operational efficiency in industrial processes.

Main Methods:

• Introduced a novel offline actor-critic algorithm with expert knowledge guidance (EKG-AC).
• Employed a diffusion-transformer-based action generation framework to mitigate out-of-distribution issues.
• Integrated an expert knowledge guidance mechanism for safe and adaptive action generation.

Main Results:

• The EKG-AC algorithm demonstrated superior optimization performance in real-world industrial processes.
• The method generated operational decisions closely aligned with expert decision-making.
• Validated effectiveness in two distinct industrial process scenarios.

Conclusions:

• The proposed EKG-AC algorithm offers a practical and valuable solution for offline reinforcement learning in industrial optimization.
• Expert knowledge integration enhances the safety and adaptability of RL agents in sensitive industrial environments.
• The diffusion-transformer framework effectively addresses challenges in offline RL for complex industrial applications.