Nonzero-Sum Game Reinforcement Learning for Performance Optimization in Large-Scale Industrial Processes.

Area of Science:

• Industrial Process Control
• Artificial Intelligence
• Game Theory

Background:

• Large-scale industrial processes often face challenges in achieving plant-wide performance optimization due to complexity and unknown dynamics.
• Traditional optimization methods may struggle with distributed control and real-time adaptation.

Purpose of the Study:

• To develop a novel technique for plant-wide performance optimization of large-scale industrial processes with unknown dynamics.
• To integrate reinforcement learning (RL) with multiagent game theory for a distributed optimization approach.

Main Methods:

• The plant-wide optimization problem is decomposed into local subproblems within a multiagent framework.
• Nonzero-sum graphical game theory is employed to compute operational indices for reaching a global Nash equilibrium.
• Reinforcement learning agents are developed to solve the game problem using real-time data, without requiring known plant dynamics.

Main Results:

• The proposed technique achieves plant-wide optimal performance through a distributed approach.
• A global Nash equilibrium is reached, ensuring production indices follow target values.
• The stability and global Nash equilibrium of the multiagent graphical game solution are rigorously proven.

Conclusions:

• The integrated RL and multiagent game theory approach provides an effective automated decision algorithm for industrial process optimization.
• The method demonstrates effectiveness using real-time data from a large mineral processing plant.
• This technique offers a robust solution for optimizing complex, unknown industrial systems.