Algorithms for Generating Computationally Simple Piecewise Linear Classifiers.

摘要

Piecewise linear classifiers constitute a group of classifiers often used in real time pattern recognition. In most cases they are reliable and sufficiently fast. In this work, new algorithms for generating piecewise linear classifiers are developed, which can easily handle multi class problems. Using only a small number of discriminant functions, they attempt to create a classifier with low error rate. In brief, the concept consists of first splitting the sample space of a given class into two subsample spaces. The samples belonging to a given class which are lying in the same (sub)sample space are said to belong to the same subclass. Next, the (sub)classes are separated using linear classifiers. Then, a new (sub)class is split and a classifier based on this splitting, is created. This process continues until the overall performance is not improved by further splitting. Our classifiers have been compared with other relevant classifiers such as Bayes classifier (for known distributions), Bayes classifier with estimated densities, the nearest neighbor rule as well as previously developed piecewise linear classifiers. So far tests have shown that our algorithms are working very well. They produce fast and reliable classifiers which in some cases have been found superior to the classifiers used for the comparison. Keywords: Cluster analysis, Monte Carlo simulation. (KR)