A comparative study of radial basis function neural networks in dynamic clustering algorithm

This paper developed two learning procedure, respectively, based on the orthogonal least squares (OLS) method and the "Innovation-Contribution" criterion (ICc) proposed newly. The orthogonal use of the stepwise-regression algorithm of the ICc mages the model structure independent of the selected term sequence and reduces the cluster region further as compared with orthogonal least squares (OLS). as the Bayesian information criteria (BIC) method is incorporate into the clustering process of the ICc, except for the widths of Gaussian functions, it has no other parameter that need tuning ,but the user is required to specify the tolerance ρ, which is relevant to noises and will be difficult to implement in the real system, for the OLS algorithm. The two algorithms are employed to the Radial Basis Function Neural Networks (RBFNN) to compare its performance for different noise nonlinear dynamic systems. Experimental results show that they provide an efficient approximation to the required results for fitting models, but the clustering procedures of the ICc is substantially better solutions than does the OLS.