Pruning fuzzy ARTMAP using the minimum description length principle in learning from clinical databases

Fuzzy ARTMAP is a family of neural network architectures based on adaptive resonance theory (ART) in which supervised learning can be carried out. However, it usually creates more categories than needed causing an overfitting problem namely in the performance of the network test. In order to avoid this problem, a confidence-based pruning method that eliminates those categories that are either less useful or less accurate and an alternative method based on the minimal description length (MDL) principle are proposed. The Cameron-Jone's error encoding scheme and Quinlan's modifier for theory encoding are used to estimate the fuzzy ARTMAP theory description length.