Structure automatic change in neural network

Abstract

In this paper, we propose a novel structure automatic change algorithm for neural-network. It can solve the problem that most neural-networks can not change the structure online. This algorithm consists of two main steps: 1) The computation of the neural-network ability to judge whether need to add nodes to the hidden layer or pruning, we use the improved support vector machine (SVM) to decide when and where to change the structure of neural-network hidden layer in this step; 2) Adjusting the parameter of the neural-network, this learning rule for the neural-network is a novel approach based on the modified back-propagation (BP). On the basis of the former methods, we propose a structure automatic changed neural network (SACNN). Finally, the SACNN is applied to track the nonlinear functions, the simulation results show that the results by this neural network perform better than the former growing cell structure (GCS) neural-network. © 2008 Springer-Verlag Berlin Heidelberg.