A Magnitude-Based ART2 Classifier: Structure and Algorithms
Jiaoyan Ai, Shange Wei, Lihua Zhang, Shuliang Sun · 2006
In an ART2 (A) network, processes of pattern-matching and vigilance-testing must be based on the value of similarity of two patterns. The measurement of similarity in classical ART2 is something about the phases of pattern vectors. It doesn't work well when we must take the magnitude of a pattern into consider, or when process a great of samples which distribute as clustering groups. Based on this, the authors propose a new modified ART2 version with Euler-distance as similarity measurement. In the new one, the magnitude of an input pattern is extracted before the pattern being sent to STM-F1 (short term memory field) and it will keep working during calculating mediate input pattern and vigilance testing. We have introduced three special functions together to help measure the similarity between an input pattern and the pattern stored in weight vector. The magnitude of the input pattern isn't lost through the whole matching and testing processes. Experiment in this paper showed that the new ART2 classifier performed better than the classical one when grouping clustering data