Clustering of Regressors for Constructing Radial Basis Function Networks

摘要

This paper presents a learning method for constructing Radial Basis function Neural Networks. In the proposed method, (which will be called Clustering of Regressors (CR)), a clustering algorithm is applied to the regressor vectors and in each cluster the regressor which is the nearest to the codebook vector is selected in contrast to the Orhogonal Least Squares (OLS) algorithm where the regressors are selected after an orthogonalization procedure one by one to decouple the contributions of regressors to the output energy. The CR is independent of ordering among regressors, requires less computations and is more robust in severely noisy conditions in comparison with OLS. By using Vector quantization for clustering process, this paper also reveals the result that, he error vector (whose 2-norm is equal to the const function defiend as total squares error over sampels) is obtained as a linear combination of the quantization error vectors. The performance of CR is compared with those of other learning algorithms in nonlinear system identification problem as well as with that of OLS in function approximation under severely noisy conditions. Results have verified the effectiveness of the proposed method.