Feedback control by online learning an inverse model.

Area of Science:

• Control Systems Engineering
• Machine Learning
• Robotics

Background:

• Traditional feedback controllers often rely on accurate models of the plant, which are challenging to develop for nonlinear systems.
• Developing effective control strategies for systems with unknown or complex dynamics remains a significant challenge in engineering.

Purpose of the Study:

• To propose a novel online learning control framework that operates without requiring explicit knowledge of the plant's dynamics.
• To demonstrate the framework's capability for fast and accurate control across diverse applications.

Main Methods:

• The proposed framework utilizes two identical learning modules: one for inverse model creation and another for plant control.
• The inverse model is trained through the exploration of a concurrently developing controller.
• The control module leverages the continuously updated inverse model for real-time control decisions.

Main Results:

• The framework achieved fast online learning and accurate control performance on varied tasks, including nonlinear heating tanks and double inverted pendulums.
• The proposed approach demonstrated effectiveness across systems with slow and fast, linear and nonlinear dynamics.
• Comparative analysis with classical control methods highlighted the framework's advantages in convergence and stability.

Conclusions:

• The novel online learning control framework offers a robust solution for controlling systems with unknown nonlinear dynamics.
• The system's adaptability and learning speed make it suitable for a wide array of control engineering problems.
• This approach advances the field of adaptive control by enabling efficient learning without prior system identification.