Design issues in applying neural networks to model highly non-linear processes

This paper looks at the selection of some of the design parameters which are crucially important for the training of a valid artificial neural network (ANN) model of processes with strong nonlinearities. Arbitrary selection of data sample time and network structure can result in an ANN model with unacceptable prediction errors. Useful guidelines concerning data sample time and model structure can be obtained by studying local linear models. The Akaike´s final prediction error (AFPE) and Akaike´s information criterion (AIC) penalise overparameterised networks and are therefore useful indicators of model parsimony. They can be used in conjunction with correlation analysis for model selection and validation.