Adaptive categorization of ART networks in robot behavior learning using game-theoretic formulation

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Online robot behavior learning faces challenges with increasing memory demands and difficulties in specifying learning parameters.
• Traditional Adaptive Resonance Theory (ART) networks struggle with operator control over learning accuracy and attention allocation.

Purpose of the Study:

• To introduce an adaptive categorization mechanism for ART networks to address limitations in online robot behavior learning.
• To propose a game-theoretic formulation for adaptive vigilance parameter control in ART networks.

Main Methods:

• An adaptive categorization mechanism was integrated into ART networks for perceptual and action pattern categorization.
• A game-theoretic approach was used to develop a vigilance parameter update rule for controlling category size.
• The proposed mechanism was tested through behavior learning experiments on a physical robot.

Main Results:

• The adaptive categorization mechanism improved category number stability in ART networks.
• The proposed vigilance parameter update rule effectively solved the issue of pre-selecting initial vigilance parameters.
• Experiments demonstrated the enhanced effectiveness of the adaptive categorization mechanism in robot behavior learning.

Conclusions:

• The proposed adaptive categorization mechanism significantly enhances ART network performance for robot behavior learning.
• This approach offers better control over learning processes, addressing key limitations of traditional methods.
• The findings pave the way for more efficient and controllable robot learning systems.