Symmetric Adaptive Resonance Theoretic Model (SMART)

Adaptive decoding of binary information from symmetric memoryless channels is necessary when the probability distribution of the transmitted codewords and the statistics of the noise which corrupts them are short-term (but not long-term) stationary. The learning process which is called for should redraw the borders of the decision regions at the receiver on the basis of estimates of the unknown probability distributions. We employ the structure which Carpenter and Grossberg propose in their ART model with several important changes, called for by the problem constraints. We show the relations between our architecture and parallel asymptotically-Maximum-Aposteriori-Probability classifiers, and compare classification performance between the SMART network and a standard classifier, which categorizes and erases on the basis of Hamming distance.