A study of structural and parametric learning in XCS.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computational Intelligence

Background:

• Learning classifier systems (LCS) comprise two key components: rule generation via genetic algorithms and parameter adjustment via evaluation.
• These components interact to optimize rule sets, aiming for minimal, fit, and non-overlapping populations.
• The XCS classifier system exemplifies this architecture.

Purpose of the Study:

• To investigate the distinct contributions of the genetic and evaluation components in LCS performance.
• To analyze the role of the genetic algorithm in train/test scenarios versus online learning.

Main Methods:

• Comparison of the XCS classifier system with XCS-NGA (XCS without the genetic algorithm).
• Evaluation of system performance on small Boolean functions using both online learning and train/test approaches.
• Analysis of rule set generalization capabilities and limitations.

Main Results:

• The genetic component plays a crucial role in the train/test approach, an aspect not present in online learning.
• XCS-NGA achieves high accuracy in online learning with sufficient rules, comparable to XCS.
• In train/test scenarios, XCS demonstrates superior generalization compared to XCS-NGA, indicating limitations of the latter.

Conclusions:

• The genetic algorithm is essential for effective generalization in concept learning (train/test) settings.
• Function approximation requirements differ between reinforcement learning (online) and concept learning (train/test).
• The study highlights the distinct roles and importance of different LCS components under varying learning paradigms.