This is a method that is similar to standard proximity style instance based learning methods, differing in that it uses fuzzy logic to determine the ``proximity'' of a given instance to the training set's instances. This would be done (at least initially) by the time and space expensive technique of storing each of the instances of the training set as a disjoint fuzzy rule. Validation instances are then fuzzily compared to each of the rules and weighted according to their ``proximity'' in feature space (how true rule is given the instance). While storing every training instance would be too expensive, the results to this would prove particularly interesting in suggesting possible ways to trim down the number of instances required.
This method was suggested because during the construction of the fuzzy rules, the human found himself feeling as though he was overfitting the data by adding rules for many different cases in the dataset, some covering relatively few instances. However, the fuzzy nature of the rules meant that these rules weren't weighted as much as more important rules, since they are less true. This then suggested the approach of having a fuzzy rule for each instance.
The main reason this wasn't attempted in this project was the fact that the fuzzy logic toolkit used ([Hillegas, Fuzzy Java]) required the fuzzy rules to be hard coded into the program. Allowing instance based rules would have required modifying the toolkit to suit this, which could not have been feasibly achieved within the scale of this project.