A multisurface method for pattern classification

摘要

A method is presented to build a multisurface classifier for a two category set of samples. The constructed classifier has some properties that make it a classifier of high quality and apparently a quasioptimal classifier. The properties imposed on the classifier are based on a generalization of Fisher's criterion function that is suitable for a non-linear classifier.The desired classifier is built in two stages. In the first stage an initial classifier is built. In the second stage the classifier is modified to obtain the desired properties. Modification is performed by use of the gradient descent procedure.