Integrating Deep Model-Based Learning With Modular State-Based Stackelberg Games for Self-Optimizing Distributed...

Area of Science:

• Manufacturing Systems Engineering
• Artificial Intelligence
• Control Theory

Background:

• Model-free modular state-based Stackelberg games (Mod-SbSG) demand extensive real-world interactions for optimization.
• These interactions are often impractical due to cost, time, and safety concerns in industrial settings.
• Previous reliance on digital twins for training Mod-SbSG players is limited by the difficulty in creating accurate system representations.

Purpose of the Study:

• To introduce a novel framework combining deep learning with Mod-SbSG for distributed self-optimization in manufacturing.
• To enhance sample efficiency and reduce reliance on real-world system interactions.
• To develop a method that learns system dynamics using deep learning for policy optimization.

Main Methods:

• Integration of deep model-based learning with modular state-based Stackelberg games (Mod-SbSG).
• Development of deep learning models to predict system dynamics accurately.
• Training of Mod-SbSG players within optimized virtual environments, leveraging learned dynamics.

Main Results:

• The proposed framework successfully replaces the need for accurate digital representations with deep learning.
• Demonstrated effectiveness of single- and multistep predictors for system dynamics.
• Achieved a 77.78% reduction in real system interactions in a laboratory testbed scenario through network reuse and transfer learning.

Conclusions:

• The deep learning-integrated Mod-SbSG framework offers a sample-efficient and practical approach to distributed self-optimization in manufacturing.
• This method significantly minimizes the need for direct interaction with physical systems, overcoming limitations of traditional approaches.
• The framework shows promise for adaptable industrial control systems, enhancing efficiency and reducing operational risks.