Robust Training of Radial Basis Function Neural Networks

摘要

Radial basis function (RBF) neural networks represent established machine learning tool with various interesting applications to nonlinear regression modeling. However, their performance may be substantially influenced by outlying measurements (outliers). Promising modifications of RBF network training have been available for the classification of data contaminated by outliers, but there remains a gap of robust training of RBF networks in the regression context. A novel robust approach based on backward subsample selection (i.e. instance selection) is proposed and presented in this paper, which searches sequentially for the most reliable subset of observations and finally performs outlier deletion. The novel approach is investigated in numerical experiments and is also applied to robustify a multilayer perceptron. The results on data containing outliers reveal the improved performance compared to conventional approaches.