REUCS-CRG
Essam Debie, Kamran Shafi, Chris Lokan · 2013
This paper proposes REUCS-CRG, a Reduct-based Ensemble of sUpervised Learning Classifier Systems with Combinatorial Rule Generation, which is an extension to the classical sUpervised Classifier System (UCS). In REUCS-CRG we build a two-stage ensemble architecture to improve generalization in UCS. In the first-stage, rough set attribute reduction is used to generate a set of reducts with different attribute subspaces, and then a diverse subset of these reducts is selected to train an ensemble of base classifiers. New instances are sent to several UCS-CRGs for classification, which includes a combinatorial rule searching component based on differential evolution algorithm. In the second-stage, a fusion method is used to combine the classification results of individual UCS-CRGs into a final decision. Three combining method are used and their results are compared: simple majority voting, winner-takes-all, and median rule. Experiments on some benchmark data sets from the UCI repository have shown that REUCS-CRG has better performance and better generalization ability than the single UCS and other UCS extensions. It also produces comparable results with other supervised learning methods. The experiments did not show significant differences in the accuracy rates obtained by the three combination methods.