Albert Orriols Homepage

Analysis of Reduction Algorithms for XCS Classifier System

by Albert Orriols-Puig and Ester Bernadó-Mansilla. In Analysis of Reduction Algorithms for XCS Classifier System. In: J. Vitrià, P. Radera, I. Aguiló (Eds.) Recent Advances in Artificial Intelligence Research and Development, IOS Press, 2004, 383-390

XCS is a classifier system that combines reinforcement learning and genetic algorithms to learn a set of rules describing the knowledge
inherent in a dataset. Recent studies have shown that XCS is highly competitive with respect to other classifier schemes. However, these studies have been mainly based on the analysis and improvement of the classification accuracy, paying few attention to the explanatory ability of the system. This paper focuses on the latter aspect. We highlight the high number of rules that XCS evolves in real-world datasets, which makes the human interpretation of the extracted knowledge difficult. Therefore, we investigate several algorithms that reduce the ruleset after training, with the aim of obtaining a simpler explanation of the dataset without loosing much generalization ability. We compare three algorithms in terms of the number of rules that are finally obtained, the loss in classification accuracy and the CPU cost. We also find that training XCS during high number of iterations does not improve the classification accuracy of the system but increases the reliability of the reduction algorithms.

No comments yet

Leave a Reply

You must be logged in to post a comment.