V. Palade, R.J. Howlett, and L.C. Jain (Eds.): KES 2003, LNAI 2774, pp. 1-6, 2003.  Springer-Verlag Berlin Heidelberg 2003 Putting the Utility of Match Tracking in Fuzzy ARTMAP Training to the Test Georgios C. Anagnostopoulos 1 and Michael Georgiopoulos 2 1 ECE Department, Florida Institute of Technology 150 West University Boulevard, Melbourne, Florida 32901, USA anagnostop@email.com http://www.fit.edu/~georgio/ 2 School of EE & CS, University of Central Florida 4000 Central Florida Boulevard, Florida 32816, USA michaelg@mail.ucf.edu Abstract. An integral component of Fuzzy ARTMAP's training phase is the use of Match Tracking (MT), whose functionality is to search for an appropriate category that will correctly classify a presented training pattern in case this particular pattern was originally misclassified. In this paper we explain the MT's role in detail, why it actually works and finally we put its usefulness to the test by comparing it to the simpler, faster alternative of not using MT at all during training. Finally, we pre- sent a series of experimental results that eventually raise questions about the MT's utility. More specifically, we show that in the absence of MT the resulting, trained FAM networks are of reasonable size and exhibit better generalization performance. 1 Introduction Fuzzy ARTMAP (FAM) [1] is a neural network architecture based on the principle of adaptive resonance theory developed in [2]. The network is capable of learning in- put-output domain associations in an on-line or an off-line fashion. As a special case, when the output domain consists of a collection of class labels, FAM can be used as a classifier. In the sequel, when we refer to FAM, we will actually be referring to the FAM classifier. FAM enjoys several desirable properties of learning including the dual support for off-line (batch) and on-line (incremental) learning as well as the property of learning stability: using fast learning its training phase completes in a finite number of list presentations (epochs). FAM follows an exemplar-based learn- ing paradigm and crystallizes its acquired knowledge in the form of categories, whose geometric representations are hyper-boxes embedded into the input domain. Learning in the presence of new data evidence occurs when either existing categories are updated or new categories are created. An integral part of FAM's training phase is the Match Tracking (MT) mechanism. When an already-existing, chosen category initially misclassifies a training pattern,