Fusion ARTMAP: Clarification, Implementation and Developments

This report claims to clarify how the fusion ARTMAP neural network operates. Here we explore the case of multiple-input sensors with single-teacher channel architecture and RT sensor models using independent baseline vigilance parameter. Two similar algorithms have been interpreted and implemented from the fusion ARTMAP architecture proposed by Asfour et al [1], [2]. One uses unsupervised Fuzzy ART [3] modules to form a compressed code within each channel before activation of a global recognition code. The other generates simultaneously the compressed and the global, recognition codes. The latter algorithm seems more closely to resemble the fusion ARTMAP architecture as set out in [1],[2]. Simulations are shown and compared with Fuzzy ARTMAP [5].