A Unified Framework for Dynamics Modeling and Control Design Using Deep Learning With Side Information on...

Area of Science:

• Control Theory
• Machine Learning
• Dynamical Systems

Background:

• Conventional data-driven methods often fail to incorporate essential control properties like stabilizability.
• Ensuring stabilizability is critical for the reliable performance of feedback control systems.

Purpose of the Study:

• To develop a unified deep learning framework for dynamics modeling and control design that explicitly guarantees stabilizability.
• To integrate prior knowledge of stabilizability into neural network-based control strategies.

Main Methods:

• A novel neural network (NN) approach that concurrently learns system dynamics, a stabilizing feedback controller, and a Lyapunov function.
• Embedding stabilizability as a core property within the deep learning framework.

Main Results:

• The proposed framework explicitly guarantees stabilizability in learned models.
• Demonstrated effectiveness across various control problems, including safety control and -gain control.
• Showcased improvements in stability and control performance in diverse scenarios.

Conclusions:

• The developed deep learning framework provides data-driven models with strong control-theoretic guarantees.
• This approach significantly enhances the utility of learned models in practical control applications.
• The methods are applicable to modeling without control design and have been open-sourced.