An empirical comparison of ID3 and back-propagation

Abstract

AI and connectionist approaches to learning from examples differ in knowledge-base representation and inductive mechanisms. To explore these differences we experiment with a system from each paradigm: 1D3 and back-propagation. We compare the systems on the basis of both prediction accuracy and length of training. The systems show distinct performance differences across a variety of domains. We identify aspects of each system that may account for these performance differences. Finally, we suggest paths for cross-paradigm interaction. 1