Testing for normality using neural networks

The most commonly used statistical procedures (t, F, chi-squared, ANOVA, regression) assume that samples have been taken at random from normal populations. In some cases the central limit theorem may provide a satisfactory approximation to normality, but, when samples are small, departures from normality can lead users of these procedures to false conclusions. In the paper on work-in-progress the authors describe the results of training an artificial neural network (ANN) to distinguish normal from non-normal samples for random samples of size 30. With little attempt at fine-tuning, the ANN achieves results comparable to those of the best known tests for normality