Optimal number of neurons for a two layer neural network model of a process

Neural networks are known as powerful tools to represent the essential properties of nonlinear processes because of their global approximation property. However, a key problem in modeling nonlinear processes by neural networks is the determination of neuron numbers. In this paper, a data based strategy for determining number of hidden layer neurons based on the Barrons work, describing function analysis and bicoherence nonlinearity measure is proposed. The proposed algorithm is evaluated for a pH neutralization process. It is shown that this algorithm has acceptable results.