Model-free reinforcement learning control with zero-min barrier functions for constrained systems.

Area of Science:

• Control Systems Engineering
• Machine Learning
• Robotics

Background:

• Controlling unknown non-affine discrete-time systems presents challenges due to inherent dynamics and physical limitations.
• Ensuring system stability and safety within operational constraints is critical for real-world applications like DC motor control.

Purpose of the Study:

• To develop an adaptive output feedback controller for unknown non-affine discrete-time systems.
• To minimize a cost-to-go function while enforcing input, output, and tracking error constraints.
• To ensure the forward invariance of safe operating regions.

Main Methods:

• Utilized a reinforcement learning algorithm combined with a zero-min barrier function to enforce constraints.
• Employed actor-critic networks implemented via fuzzy rule-emulated networks for adaptive control.
• Derived online learning laws for adaptive controller updates without inner iterations.

Main Results:

• Achieved nearly optimal tracking performance for unknown non-affine systems.
• Successfully enforced symmetric and asymmetric constraints, ensuring safe operation.
• Demonstrated robustness against extreme disturbances and trajectory changes.

Conclusions:

• The proposed adaptive controller effectively balances near-optimal performance with strict constraint satisfaction.
• The method enhances computational efficiency by eliminating inner iterations.
• Validated effectiveness and robustness in a high-gain DC motor torque control system.