A thesis submitted in partial fulfillment of the requirements for the degree of Master of Engineering, School of Engineering of Technology

Thesis (M.Eng.) - Asian Institute of Technology, 1994

This thesis investigates Adaptive Resonance Theory 1 (ARTl) as a pattern clustering algorithm. Inherent characteristics of the model are analyzed. In particular the vigilance parameter, p, and its role in classification of patterns is examined. The experiments show that the vigilance parameter as defined by Carpenter & Grossberg does not necessarily increase the number of categories with its value, but decrease also, against the claim made by them. Hence, the lemma, "Increasing p increases the total number of clusters learned and decreases the size of each cluster" stated by Barbara Moore (1989), an MIT AI researcher, is not always valid. A modified vigilance test criteria has been proposed, which takes into account, the problem of subset & superset patterns and stably categorize, arbitrarily many input patterns in one list presentation when the vigilance parameter is closer to one. The proposed method also performs much better with regard to classification of patterns and reduces the number of list presentations required for stable category learning. Better perfo1mance of new similarity criteria with regard to noisy patterns also has been shown with experimental results.