Abstract
This thesis studies fuzzy ARTMAP overfitting avoidance based on the Minimum Description Length (MDL) principle. Fuzzy ARTMAP is a good neural network architecture in which rules can be extracted and reviewed by human experts, but it tends to create too many categories than needed. This paper argued that creating fuzzy ARTMAP categories is a tradeoff between complexity and accuracy, and vigilance test and match tracking are not enough to solve the tradeoff. MDL gives a good measurement for both complexity and accuracy. A MDL-based fuzzy ARTMAP pruning algorithm is developed. Experiments showed that it provides better performance and far fewer categories than the original fuzzy ARTMAP. We also showed that it gives bettes performance over other machine learning systems on a number of benchmark medical databases. In particular, MDL-based pruning has a stronger theoretical basis than the previous confidence-based pruning algorithm. This thesis also studies the problem of missing value and nominal attribute treatment, which is common in medical databases.