On optimal adaptive classifier design criterion

The authors develop a Bayes consistent classifier design criterion, the GMEE (generalized minimum empirical criterion), from an analysis of classification error. The criterion has been applied to the design of neural network classifiers. The results of two examples indicate that the GMEE can yield optimal neural network classifiers. It is also demonstrated that a neural network classifier is Bayes optimal if it is selected by the GMEE. Hence, the results provide a theoretical foundation for the connectionist approach to classification problems. These results can also be extended to the optimal design of other types of neural network, e.g., radial basis function networks.<>