Enhancing Pattern Classification with Relational Fuzzy Neural Networks and Square BK-Products

This research presents several important developments in pattern classification using fuzzy neural networks and BK-Square products and presents extensions to max-min fuzzy neural network research. In this research, the max and min operations used in the fuzzy operations are replaced by more general t-norms and co-norms, respectively. In addition, instead of the £ukasiewicz equivalence connective used in network of Reyes-Garcia and Bandler, this research introduces a variety of equivalence connectives. A new software tool was developed specifically for this research, allowing for greater experimental flexibility, as well as some interesting options that allow greater exploitation of the merits of the relational BK-square network. The effectiveness of this classifier is explored in the domain of phoneme recognition, taxonomic classification, and diabetes diagnosis. This research finds that the variance of fuzzy operations in equivalence and implication formulae, in complete divergence from classical composition, produces drastically different performance within this classifier. Techniques are presented that select effective fuzzy operation combinations. In addition, this classifier is shown to be effective at feature selection by using a technique which usually would be impractical with standard neural networks, but is made practical through the unique nature of this classifier.