Combination of boosted classifiers using bounded weights

Abstract

A recently developed neural network model that is based on bounded weights is used for the estimation of an optimal set of weights for ensemble members provided by the AdaBoost algorithm. Bounded neural network model is firstly modified for this purpose where ensemble members are used to replace the kernel functions. The optimal set of classifier weights are then obtained by the minimization of a least squares error function. The proposed weight estimation approach is compared to the AdaBoost algorithm with original weights. It is observed that better accuracies can be obtained by using a subset of the ensemble members. © Springer-Verlag Berlin Heidelberg 2005.