Model-free nonlinear robust control design via online critic learning.

Area of Science:

• Control Engineering
• Machine Learning
• Robotics

Background:

• Robust control of complex systems often requires detailed system dynamics knowledge.
• Data-driven methods offer an alternative but face challenges in learning without prior information.

Purpose of the Study:

• To develop a data-driven actor-critic scheme for robust control.
• To eliminate the need for system dynamics knowledge and initial stable control.
• To enhance disturbance attenuation performance.

Main Methods:

• A novel performance index is designed, focusing on the decreasing rate of the conventional cost.
• An actor-critic scheme utilizing three neural networks is employed for online optimal control policy approximation.
• The method is validated through numerical simulations and an inverted pendulum experiment.

Main Results:

• The proposed method effectively approximates optimal control policies online.
• The learning process and resulting control policy exhibit significant robustness.
• The approach demonstrates superior disturbance attenuation compared to benchmark methods.
• It relaxes the dependence on initial admissible control.

Conclusions:

• The developed data-driven actor-critic method provides a robust solution for complex control problems.
• This approach is advantageous for systems where dynamics are unknown or difficult to model.
• The method shows practical applicability and improved performance in real-world scenarios.