Model structure determination in neural network models

Abstract

Feedforward neural networks (FNN) have been used intensively for the identification and control of chemical engineering processes. However, there is no efficient model structure determination methodology for a particular mapping application. This has resulted in a tendency to use networks that are much larger than required. In this paper a new procedure for model structure determination in feedforward neural networks is proposed. This procedure is based on network pruning using the orthogonal least-squares technique to determine insignificant or redundant synaptic weights, biases, hidden nodes and network inputs. The advantages of this approach are discussed and illustrated using simulated and experimental data. The results show that the orthogonal least-squares technique is quite efficient in determining the significant elements on the neural network models. The results also show the importance of pruning procedures to identify parsimonious FNN models.