Performance Enhancement of RBF Networks in Classification by Removing Outliers in the Training Phase

摘要

During data collection and analysis there often exist outliers which affect final results. In this paper we address reducing effects of outliers in classification with Radial Basis Function (RBF) networks. A new approach called iterative RBF (iRBF) is proposed. In which training RBF networks is repeated if there exist outliers in the training set. Detection of outliers is performed by relying upon outputs of the RBF networks which correspond to applying the training set at the input units. Detected outliers have had to be eliminated before the training set is used in the next training time. In this approach we achieve a good performance in outlier rejection and classification with training sets existing outliers.