Adaptive RBF Neural Network Training Algorithm For Nonlinear And Nonstationary Signal

This paper presents an improved adaptive radial basis function neural network (RBF NN) for nonlinear and nonstationary signal. The proposed method possesses distinctive properties of Lyapunov theory-based adaptive filtering (LAF) in Seng Kah Phooi et. al, (2002). This method is different from many RBF NN training methods using gradient search techniques. A new Lyapunov function of the error between the desired output and the RBF NN output is first defined. The output asymptotically converges to the desired output by proper design of the weight adaptation law in Lyapunov sense. In this paper, we have proved that the design is independent of statistic properties of the input and output signals. The proposed method has better tracking capability compared with the LAF. The performance of the proposed technique is illustrated through the nonlinear adaptive prediction of nonstationary speech signals