Structural learning in multi-layer, feedforward neural networks was studied using Ishikawa's modified back-propagation algorithm with forgetting of the connection weights. The following major features of structural learning were analyzed: (1) knowledge extraction from the skeleton structure of the trained network; (2) independence of the trained neural network of the initial network structure, and (3) improved generalization properties compared to neural networks trained by standard back-propagation. The proper choice of forgetting constant was investigated by several authors but no generally accepted method has been established yet. In this paper, the generalization rate of the trained network is analyzed as a possible means of selecting optimum model parameters. The results are illustrated using Fisher's IRIS data and anomaly detection in time series.