Bagging and Boosting performance in Projection Pursuit Regression

Recently, many authors have proposed new algorithms to improve the accuracy of certain classifiers on artificial and real data sets. The goal is to assemble a collection of individual classifiers based on resampling of data set. Bagging (Breiman, 1996) and AdaBoost (Freund & Schapire, 1997) are the most used procedures: the first fits many classifiers to bootstrap samples of data and classifies the units by the majority vote; the second fits many classifiers weighing the data and classifies the units by the weighted majority vote. The success of these methods is in terms of the bias-variance components of the generalization error. In the regression context the application of these techniques has received little investigation. Drucker (1997) and other authors modified AdaBoost to use regression trees as predictors. Our aim is to verify by simulations if boosting or bagging can improve the capability to reduce conjointly training set error and generalization error using Projection Pursuit Regressions as predictors.