Adaptive resonance theory : characterization of ART1 and development of similarity measures
- Bangkok : Asian Institute of Technology, 1993
- 50 leaves + 1 online resource
- Thesis ; no. CS-93-5 .
- Asian Institute of Technology. Thesis ; no. CS-93-5 .
A thesis submitted in partial fulfillment of the requirements for the degree of Master of Engineering
Thesis (M.Eng.) - Asian Institute of Technology, 1993
In the present work, Adaptive Resonance Theory 1 (ART1) as a clustering algorithm is analyzed. Under some network modeling assumptions and enhanced template definitions, some of its emergent but inherent characteristics related to similarity and learning are investigated and demonstrated to understand more about the output it generates. The memory capacity in bits (each bit bears an essential feature of a pattern), its upper and lower limits, and the capacity variation upon learning of input patterns in real time, which Carpenter & Grossberg avoided in their original work, are derived. A new idea for a similarity metric in novelty detector (orienting-subsystem) is proposed to extend its operation over arbitrary pattern sequences in the real-time for one pattern list presentation. The authenticity of the proposed method for fresh and noisy pattern environments compared to Carpenter & Grossberg method is also illustrated with some examples. Moreover, the present work on this paradigm points out the ways of extending this theory to look into stability-plasticity problems in the context of some real world applications.