ABSTRACT
This chapter describes the results of an neural networking (NN) experiment based upon one of the databases. The sample itself was split into two parts: a 'training set' of 41 matched pairs from which the NN models were derived, and a hold out sample of 20 matched pairs. A particular problem with iterative models is that initial parameter values have to be chosen before passes are made through the data in order to converge on an optimum solution. In the first experiment there was clear evidence of 'overfitting' the model on the training set. One of the main reasons for developing the neural networking models was to see whether they significantly outperform conventional, statistically determined failure prediction models. In terms of the number of computations required to derive each model, with one neuron 40,000 'epochs' were required, taking some three hours of running time.
