Learning aggregation for combining classifier ensembles

Creating classifier ensembles and combining their outputs to achieve higher accuracy have been of recent interest. It was noted that when using such multiple classifier approaches the members of the ensemble should be error-independent. The ideal, in terms of ensembles of classifiers, would be a set of classifiers which do not show any coincident errors. That is, each of the classifiers generalized well, and when they did make errors on the test set, these errors were not shared with any other classifier. Various approaches for achieving this have been presented. This paper compares two approaches introduced for training multiple classifiers systems. These approaches are based on the feature based aggregation architecture and the adaptive training algorithm. An empirical evaluation using two data sets shows a reduction in the number of training cycles when applying the algorithm on the overall architecture, while maintaining the same or improved performance. The performance of these approaches is also compared to standard approaches proposed in the literature. The results substantiate the use of adaptive a-dining for both the ensemble and the aggregation architecture.