A novel OLS algorithm for training RBF neural networks with automatic model selection

Orthogonal Least Squares (OLS) algorithm has been extensively used in basis selection problems for RBF networks, but it is unable perform model selection automatically because the user is required to specify the tolerance?, which is relevant to noises and will be difficult to implement in the real system. therefore, a generic criterion that defines the optimum number of its basis function is proposed. In this paper, Not only is the Bayesian information criteria (BIC) method incorporate into the basis function selection process of the OLS algorithm for assigning its appropriate number, but also we develop a new method to optimize the widths of Gaussian functions in order to improve the generalization performance, The augmented algorithms are employed to the Radial Basis Function Neural Networks (RBFNN) to compare its performance for known and unknown noise nonlinear dynamic systems, Experimental results show the efficacy of this criterion and the importance of a proper choice of basis function widths.