Growing construction of conlitron and multiconlitron

Abstract

Based on the concepts of conlitron and multiconlitron, we propose a growing construction technique for improving the performance of piecewise linear classifiers on two-class problems. This growing technique consists of two basic operations: SQUEEZE and INFLATE, which can produce relatively reliable linear boundaries. In the convexly separable case, the growing process, forming a growing support conlitron algorithm (GSCA), starts with an initial conlitron and uses SQUEEZE to train a new conlitron, moving its classification boundary closer to the interior convex region and fitting the data distribution better statistically. In the commonly separable case, the growing process, forming a growing support multiconlitron algorithm (GSMA), starts with an initial multiconlitron and uses INFLATE and SQUEEZE to train a new multiconlitron, making its classification boundary adjusted to improve the generalization ability. Experimental evaluation shows that the growing technique can simplify the structure of a conlitron/multiconlitron effectively by reducing the number of linear functions, largely keeping and even greatly improving the level of classification performances. Therefore, it would come to play an important role in the subsequent development of piecewise linear learning, with the main goal to improve piecewise linear classifiers in a general framework.