Artificial neural networks for determination of enantiomeric composition of α-phenylglycine using UV spectra of cyclodextrin host–guest complexes: Comparison of feed-forward and radial basis function networks

In this work feed-forward neural networks and radial basis function networks were used for the determination of enantiomeric composition of α-phenylglycine using UV spectra of cyclodextrin host–guest complexes and the data provided by two techniques were compared. Wavelet transformation (WT) and principal component analysis (PCA) were used for data compression prior to neural network construction and their efficiencies were compared. The structures of the wavelet transformation–radial basis function networks (WT–RBFNs) and wavelet transformation–feed-forward neural networks (WT–FFNNs), were simplified by using the corresponding wavelet coefficients of three mother wavelets (Mexican hat, daubechies and symlets). Dilation parameters, number of inputs, hidden nodes, learning rate, transfer functions, number of epochs and SPREAD values were optimized. Performances of the proposed methods were tested with regard to root mean square errors of prediction (RMSE%), using synthetic solutions containing a fixed concentration of β-cyclodextrin (β-CD) and fixed concentration of α-phenylglycine (α-Gly) with different enantiomeric compositions. Although satisfactory results with regard to some statistical parameters were obtained for all the investigated methods but the best results were achieved by WT–RBFNs.