A genetic approach for training diverse classifier ensembles

David Gacquer, François Delmotte, Veronique Delcroix, Sylvain Piechowiak.

Classification is an active topic of Machine Learning. The most recent achievements in this domain suggest using ensembles of learners instead of a single classifier to improve classification accuracy. Comparisons between Bagging and Boosting show that classifier ensembles perform better when their members exhibit diversity, that is commit different errors. This paper proposes a genetic algorithm to design classifier ensembles, using a fitness function based on both accuracy and diversity. The proposed implementation has been run on several UCI Machine Learning datasets and compared to the performances obtained with bagging algorithm and a single classifier of the same nature.

PDF full paper