Iterative inversion of neural networks and its application to adaptive control

Area of Science:

• Control Systems Engineering
• Machine Learning
• Adaptive Control

Background:

• Traditional neural network controllers pose analysis challenges due to direct integration into feedback paths.
• Existing methods struggle with rapidly changing plant dynamics.

Purpose of the Study:

• To develop a novel control strategy using iterative constrained inversion.
• To enable controllers to adapt online to plant dynamic variations.
• To propose a neural network-based model reference adaptive controller.

Main Methods:

• Learning the forward model of the plant.
• Performing iterative inversion online to generate control commands.
• Demonstrating a neural network-based model reference adaptive controller on a linear system.

Main Results:

• The iterative inversion technique effectively generates control commands online.
• The proposed approach allows controllers to adapt to changing plant dynamics.
• The study highlights the necessity of a dither signal for online dynamic system identification.

Conclusions:

• Iterative constrained inversion offers an alternative to direct neural network controller placement.
• The developed method enhances adaptability and simplifies analysis in control systems.
• Online identification of dynamic systems may require a dither signal.