Off-line structural risk minimization and BARTMAP-S

Abstract

BARTMAP-S (Simplified Boosted ARTMAP) is a neural network architecture with which structural risk minimization can be performed, although indirectly. BARTMAP-S is trained in an online fashion, consistent with the original way intended for the fuzzy ARTMAP neural network architecture. We propose an extension to BARTMAP-S for conducting off-line learning. Consequently, this alternate mode of learning will allow us to conduct structural risk minimization more directly. We describe the new architecture and present some empirical results to demonstrate the usefulness of structural risk minimization in learning with an ARTMAP-based neural network.