Clustering-based initialization of Learning Classifier Systems

Abstract

The present paper investigates whether an ''informed'' initialization process can help supervised LCS algorithms evolve rulesets with better characteristics, including greater predictive accuracy, shorter training times, and/or more compact knowledge representations. Inspired by previous research suggesting that the initialization phase of evolutionary algorithms may have a considerable impact on their convergence speed and the quality of the achieved solutions, we present an initialization method for the class of supervised Learning Classifier Systems (LCS) that extracts information about the structure of studied problems through a pre-training clustering phase and exploits this information by transforming it into rules suitable for the initialization of the learning process. The effectiveness of our approach is evaluated through an extensive experimental phase, involving a variety of realworld classification tasks. Obtained results suggest that clustering-based initialization can indeed improve the predictive accuracy, as well as the interpretability of the induced knowledge representations, and paves the way for further investigations of the potential of better-than-random initialization methods for LCS algorithms.

Inducing Generalized Multi-Label Rules with Learning Classifier Systems

“…The most prominent example of this class of systems is the accuracy-based UCS algorithm (Bernadó-Mansilla and Garrell-Guiu, 2003;Orriols-Puig and Bernadó-Mansilla, 2008). Additionally, we have recently introduced SS-LCS, a supervised strength-based LCS, that provides an efficient and robust alternative for offline classification tasks (Tzima et al, 2012;Tzima and Mitkas, 2013) by extending previous strength-based frameworks (Wilson, 1994;Kovacs, 2002a,b).…”

“…MlS-LCS also employs a population initialization method that extracts information about the structure of the studied problems through a pre-training clustering phase and exploits this information by transforming it into rules suitable for the initialization of the learning process. The employed method is a generalization for the multi-label case of the clustering-based initialization process presented in Tzima et al (2012) that has been shown to boost LCS performance, both in terms of predictive accuracy and the final evolved ruleset's size, in supervised single-label classification problems.…”

“…Indeed, in recent years, LCS have been modified for data mining (Bull et al, 2008) and single-step classification problems, notably in the UCS (Orriols-Puig and Bernadó-Mansilla, 2008) and the SS-LCS frameworks (Tzima et al, 2012;Tzima and Mitkas, 2013). Their niche-based update and their overall iterative (rather than batch) learning approach has been shown to be very efficient in domains where different problem niches occur (including multi-class and unbalanced classification problems).…”