An Efficient Algorithm for Feedforward Neural Network Reconstructing and Its Application

Abstract

In this paper, an efficient algorithm based on the pruning method and the Levenberg Marquardt (LM) is presented to design the single hidden layer feedforward neural network (FNN). This new approach can prune the redundant hidden nodes by calculating the Hessian and removing the lines in the matrix for reconstructing the FNN. The proposed pruning hidden nodes (PHN) algorithm can adjust the parameters of the neural networks as well. The proposed PHN is simple and effective and generates a FNN model with a high accuracy and compact structure. In addition, the convergence of both the structures dynamic process and after the modifying is discussed. The PHN is then tested on the non-linear functions approximation to illustrate the effectiveness of our proposed reconstructing scheme. Finally, the PHN is employed to model chemical oxygen demand (COD) concentration in the wastewater treatment process. Experimental results show that the proposed method is efficient for network structure pruning and it achieves better performance than some of the existing algorithms.