A new piecewise linear classifier

An algorithm for generating piecewise-linear classifiers (PLCs) is reported. This approach can easily handle multiclass problems. The concept consists of first applying a linear classifier to the design set. If it fails to separate the classes, then the sample space of one of the classes is divided into two subsample spaces. A PLC based on this splitting is designed and evaluated. This sequence of splitting, redesigning, and evaluating continues until the overall performance is no longer improved. The classifier is compared with other relevant classifiers. Its ability to produce fast and reliable classifications is demonstrated.<<ETX>>