Minimum misclassification error performance measure for layered networks of artificial fuzzy neurons

Summary form only given. Neural network techniques that directly minimize misclassification error (MME) are more appropriate for automatic classification than least mean square (LMS) techniques. While LMS is designed to give a best fit to the ´humps´ of the data distribution, MME is sensitive to the ´tails´ that have overlapping minima. It is shown that the MME energy landscape can be nonconvex. Cauchy simulated annealing was used to find the global minimum, and the resulting solution is shown to give a lower misclassification rate than the linear-LMS solution. MME layered networks of fuzzy neurons circumvent the formidable difficulty in the estimation of Bayesian probability that puts emphasis on approximating the ´humps´ instead of the ´tails´