PREDICTION OF QUALITY AND ORIGIN OF BLACK TEA AND PINE RESIN SAMPLES FROM CHROMATOGRAPHIC AND SENSORY INFORMATION - EVALUATION OF NEURAL NETWORKS

The accuracies of neural network and statistical methods were similar for classifying the origin of black teas from their phenolic composition. When the data are non-normal, as was the case for the pine resin samples, the neural network offered a significant improvement. Neural networks were less accurate than stepwise multiple regression as a model for predicting black tea score and price from their chemical composition and sensory attributes. The accuracy improved and the training time was reduced when training variables chosen by stepwise multiple regression were selected. An advantage of the neural network model is that a single model could predict several parameters simultaneously. The selection criterion of neural networks could be estimated by inspection of the most positive weights derived from two-layer trained networks.