Standalone CMAC control system with online learning ability.

Area of Science:

• Control Systems Engineering
• Artificial Intelligence
• Robotics

Background:

• Traditional control systems often require extensive offline training or conventional controllers.
• Adaptive control systems aim to adjust parameters in real-time for unknown or changing dynamics.
• Cerebellar Model Articulation Controller (CMAC) is a type of associative neural network used for function approximation.

Purpose of the Study:

• To propose a novel single-input cerebellar model articulation controller (CMAC) control system.
• To demonstrate online learning and control capabilities without preliminary offline training.
• To ensure system stability and convergence of output error for unknown plants.

Main Methods:

• Implementation of a single-input, differentiable CMAC controller solely responsible for plant control.
• Utilizing gradient descent algorithm for deriving online learning rules for the CMAC.
• Approximating plant sensitivity with a simple formula for applicability to unknown systems.
• Deriving convergence conditions for output error using a discrete-type Lyapunov function.

Main Results:

• The proposed CMAC controller effectively controls plants through online learning at each step.
• The system demonstrates successful control without the need for conventional controllers or offline pre-training.
• Simulations confirm the controller's effectiveness across three distinct plant types.

Conclusions:

• The proposed single-input CMAC offers an effective and adaptive control solution for various plants.
• The online learning approach simplifies implementation and broadens applicability to unknown systems.
• The derived conditions guarantee the stability and convergence of the adaptive control system.