Reinforcement learning to adaptive control of nonlinear systems

Area of Science:

• Control Theory
• Artificial Intelligence
• System Identification

Background:

• Nonlinear systems pose significant challenges in control engineering.
• Traditional methods often struggle with complex dynamics and require precise system models.
• Artificial Neural Networks (ANNs) offer a powerful approach for modeling and controlling complex systems.

Purpose of the Study:

• To develop and evaluate a novel reinforcement learning scheme for effective nonlinear system linearization.
• To introduce the Reinforcement Linearization Learning System (RLLS) integrating artificial neural networks.
• To demonstrate the concurrent identification and linearization capabilities of the proposed system.

Main Methods:

• Utilizing feedback linearization theory as the foundational principle.
• Developing a Reinforcement Linearization Learning System (RLLS) with two subsystems: Evaluation Predictor (EP) and a short-term action selector (Linearizing Control - LC, Reinforce Predictor - RP).
• Employing a reference model as the environment to provide reinforcement signals for the linearization process.

Main Results:

• The RLLS successfully achieved concurrent system identification and linearization.
• Simulation results on a pendulum system demonstrated superior control reliability and robustness compared to conventional ANN schemes.
• A Proportional-Integral (PI) controller effectively managed the linearized plant, exhibiting linear system behavior.

Conclusions:

• The proposed reinforcement learning scheme offers an effective method for nonlinear system linearization.
• The RLLS demonstrates significant improvements in control reliability and robustness.
• This approach paves the way for advanced control strategies in complex dynamic systems.