The influence of input data preprocessing and of learning error on the performances of ANN systems identifying amphetamines

In the last years, artificial neural networks (ANN) have became the most powerful tools applied for the identification of unknown samples, the recognition of images, the prediction of different phenomena, etc. We are presenting an ANN system, which was optimised in order to enhance its efficiency in discriminating illicit amphetamines according to the substitution pattern associated with the psychotropic activity for which they are abused. The initial input variables are the GC-FTIR absorption intensities measured in each normalized infrared spectrum of the compounds forming the digital database. In order to improve the sensitivity and the selectivity of the ANN system, we have applied two optimization methods: preprocessing of input data using a feature weight spectrum, and changing the order in which the networks use the samples from the training set in the learning process. The order in which the samples were included in the optimized training set was established according to their learning error. The results show that the latter optimization method has a significant impact on the efficiency of the artificial neural network. The performance with which each network identifies the class identity of an unknown sample was evaluated by calculating several figures of merit. The results of the comparative analysis are presented.