In this thesis ensemble learning methods for classification are presented, especially AdaBoost and Arcing algorithms are analyzed. It is theoretically shown why AdaBoost and Arcing are working well and explained what kind of target function is optimized. It is pointed out that AdaBoost's optimization is similar to a annealing process and linear programming. From an asymptotic analysis of AdaBoost's learning process it is shown that those algorithms are not noise robust asymptotically. In order to get a noise robustness, several improvements of AdaBoost are proposed. The concept of slack variables is introduced into AdaBoost and is motivated from the margin's point of view. These improvements are experimentally compared for artificial and real-world datasets. They generally show much better results than the original AdaBoost algorithm and also a better overall performance than the Support Vector algorithm, which is considered to be an excellent classifier.

In German