Weighted combination of neural network ensembles

There exist numerous schemes and methods to determine the output of an ensemble of classifiers. The most common approach being the majority vote. Furthermore, we might expect that an improvement can be achieved if there is a method by which we may weigh the members of the ensemble according to their individual performance. The feature based approach presented an architecture that tries to approach this target. However, if there is a way that the final classification may influence these weights we should expect an increased performance in the overall classification task. In this paper we present a new training algorithm that utilizes a feedback mechanism to iteratively improve the classification capability of the feature based approach. This approach is compared with the standard training method as well as standard aggregation schemes for combining classifier ensembles. Empirical results show that this architecture improved on classification accuracy.