Reinforcement Learning-Based Nearly Optimal Control for Constrained-Input Partially Unknown Systems Using...

Area of Science:

• Control Theory
• Artificial Intelligence
• Robotics

Background:

• Partially unknown systems present significant control challenges.
• Existing methods often require initial stabilization or complete system knowledge.
• Reinforcement learning offers a promising approach for adaptive control.

Purpose of the Study:

• To develop a synchronous reinforcement-learning-based algorithm for input-constrained, partially unknown systems.
• To eliminate the necessity of an initial stabilizing control.
• To ensure the stability and bounded errors of the closed-loop system.

Main Methods:

• Utilized a first-order robust exact differentiator for approximating unknown drift dynamics.
• Employed critic, actor, and disturbance neural networks (NNs) for approximating value, control, and disturbance policies.
• Applied value function approximation to solve the Hamilton-Jacobi-Isaacs equation.

Main Results:

• The proposed control algorithm ensures the stability of the closed-loop system.
• State and weight errors of the three neural networks were proven to be uniformly ultimately bounded.
• Simulation results verified the effectiveness of the developed control strategy.

Conclusions:

• The synchronous reinforcement learning algorithm effectively controls input-constrained, partially unknown systems.
• The method provides a stable and robust control solution without requiring initial stabilization.
• The use of neural networks for approximation proved successful in solving the Hamilton-Jacobi-Isaacs equation.