Credit assigned CMAC and its application to online learning robust controllers.

Area of Science:

• Robotics
• Control Systems Engineering
• Machine Learning

Background:

• Conventional Cerebellar Model Articulation Controllers (CMAC) learning schemes distribute errors equally among hypercubes, irrespective of their learning history or credibility.
• This equal distribution can lead to suboptimal learning efficiency and slower convergence rates in complex control tasks.

Purpose of the Study:

• To propose a novel, credit-assigned learning scheme for CMAC that accelerates the learning process.
• To develop an online learning robust controller utilizing the enhanced CMAC for improved performance.
• To demonstrate the controller's ability to accurately trace various trajectories in real-time.

Main Methods:

• Implemented a credit-assigned learning approach for CMAC, using the inverse of hypercube learned times as a credibility measure.
• Developed an online learning robust controller that integrates CMAC with previous control input, current output acceleration, and desired output.
• Introduced an initial trial mechanism to support the early stages of learning.

Main Results:

• The proposed credit-assigned CMAC learning scheme significantly increases learning speed compared to conventional methods.
• The integrated robust learning controller demonstrated accurate online trajectory tracing capabilities.
• The system effectively learned and adapted online, showcasing robust performance.

Conclusions:

• The novel credit-assigned learning scheme offers a substantial improvement in CMAC learning speed and efficiency.
• The developed online learning robust controller, powered by the enhanced CMAC, provides accurate and adaptive control.
• This approach holds promise for real-time control applications requiring fast adaptation and precise trajectory following.