On performance measures of artificial neural networks trained by structural learning algorithms

Structural learning in multi layer, feedforward neural networks was studied using M. Ishikawa's (1994) modified backpropagation algorithm with forgetting of the connection weights. The proper choice of forgetting constant was investigated previously but no generally accepted method has been established yet. The generalization rate of the trained network is analyzed as a possible means of selecting optimum model parameters. The results are illustrated using R.A. Fisher's (1936) IRIS data and anomaly detection in time series.