Meta-Evolutionary Ensembles

Yong Seog Kim

University of Iowa

2001. 5. 10

Ensemble methods have shown the potential to improve on the performance of individual classifiers as long as the members of the ensamble are sufficiently diverse. Individual classifiers have been trained for example on selected subsets of the records or on projections of the feature space to produce diversity. The resulting ensembles reflect a priori decisions about how to allocate records or features across classifiers. In this talk we propose a meta-evolutionary approach in which both individual classifiers and groups adapt. Ensembles compete for member classifiers, and are rewarded based on their predictive performance. Individual classifiers also evolve, competing to correctly classify test points, and are given extra rewards for getting difficult points right. In this way we aim to optimize ensembles rather than form ensembles of individually-optimized classifiers. Our preliminary results on a small number of data sets suggest that this approach can generate ensembles that are more effective than single classifiers and competetive with traditional ensemble methods. In the long run, this approach will provide new insight into how ensembles should be optimally constructed.

This page is maintained by Hee-Ju Chang (
Last update: May 08, 2001