Network complexity analysis of multilayer feedforward artificial neural networks

Abstract

Artificial neural networks (NN) have been successfully applied to solve different problems in recent years, especially in the fields of pattern classification, system identification, and adaptive control. Unlike the traditional methods, the neural network based approach does not require a priori knowledge on the model of the unknown system and also has some other significant advantages, such as adaptive learning ability as well as nonlinear mapping ability. In general, the complexity of a neural network structure is measured by the number of free parameters in the network; that is, the number of neurons and the number and strength of connections between neurons (weights). Network complexity analysis plays an important role in the design and implementation of artificial neural networks - not only because the size of a neural network needs to be predetermined before it can be employed for any application, but also because this dimensionality may significantly affect the neural network learning and generalization ability. This chapter gives a general introduction on neural network complexity analysis. Different pruning algorithms for multi-layer feedforward neural networks are studied and computer simulation results are presented. © 2010 Springer-Verlag Berlin Heidelberg.