Reinforcement-Learning-Based Disturbance Rejection Control for Uncertain Nonlinear Systems.

Area of Science:

• Control Theory
• Nonlinear Systems
• Machine Learning

Background:

• Uncertain nonlinear systems pose significant challenges for control design.
• Traditional methods often struggle with complex dynamics and external disturbances.
• Nonsimple nominal models complicate disturbance rejection and optimal policy approximation.

Purpose of the Study:

• To develop a reinforcement-learning-based disturbance rejection control strategy.
• To address uncertainties and nonsimple nominal models in nonlinear systems.
• To achieve practical convergence of system states and optimal policies.

Main Methods:

• An extended state observer (ESO) is employed to estimate system states and total uncertainty.
• Real-time control compensation for total uncertainty is performed.
• An experience-based reinforcement learning (RL) technique is used for online optimal policy approximation.

Main Results:

• The proposed method demonstrates practical convergence of the system state to the origin.
• The developed policy converges to the ideal optimal policy.
• The framework successfully avoids the need for the persistence of excitation (PE) condition.

Conclusions:

• The RL-based control with ESO offers an effective solution for uncertain nonlinear systems.
• The method provides robust disturbance rejection and optimal control policy.
• The absence of the PE condition broadens the applicability of the technique.