Generalization in the XCSF classifier system: analysis, improvement, and extension

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computational Intelligence

Background:

• The eXternal Classifier System (XCSF) is a learning system that evolves accurate and parsimonious models.
• Generalization in XCSF can be influenced by input range and classifier weight convergence.
• The Widrow-Hoff update rule in XCSF can lead to slow convergence, resulting in suboptimal approximations.

Purpose of the Study:

• To analyze and improve generalization capabilities in the XCSF algorithm.
• To address the issue of slow classifier weight convergence in XCSF.
• To explore methods for enhancing XCSF's ability to evolve more efficient approximations.

Main Methods:

• Theoretical analysis of classifier weight convergence in XCSF using eigenvalue properties.
• Introduction of three novel methods for updating classifier weights: condition-based normalization, linear least squares, and recursive linear least squares.
• Experimental evaluation of the proposed methods on XCSF performance and generalization.

Main Results:

• Input range significantly influences the types of generalizations evolved by XCSF.
• All three proposed methods improve XCSF generalization, with least squares approaches showing superior performance and robustness.
• The system can be extended to incorporate polynomial approximations for enhanced modeling.

Conclusions:

• The proposed weight update methods effectively increase XCSF's generalization capabilities.
• Least squares-based updates offer the most significant improvements in performance and robustness for XCSF.
• XCSF can be extended to handle more complex relationships using polynomial approximations.