A new method based on counterpropagation network algorithm for chemical pattern recognition

This paper develops an approach for chemical pattern recognition realized by using a modified counterpropagation network (CPN). Combining Kohonen network and least mean-square (LMS) learning enables the modified CPN algorithm to solve the supervised type classification problem. First, patterns are separated into subgroups in which there are no samples of different categories. The Kohonen layer of neurons is used to establish the clusters. During the training process the updating of weights deals with merely winning neurons in stead of a domain of neurons as done in the standard counterpropagation network (SCPN) algorithm. Second, based on the obtained subgroups patterns of the same category are grouped together by LMS in the output layer. Thus the proposed method yields a piecewisely linear discriminant boundary for classification. Compared with the existing methods the proposed procedure is competent for more complex problems of classification and efficiently lightens the computational burden of network training. Experimental investigations of simulated and real data sets showed that the proposed method outperformed standard linear discriminant analysis (SLDA) method in the cases of partially linear inseparability. Compared with the conventional neural work methods, i.e. the standard backpropagation network (BPN) and SCPN, it also showed better learning efficiency and classification performance.