Unbounded Recurrent Fuzzy Min-Max Neural Network for Pattern Classification

A prominent stumbling block in use of the original fuzzy min-max neural network (FMN) and other FMN-based algorithms is their sensitivity to the expansion coefficient or hyperbox expansion parameter. Specifically, these algorithms need to be trained with its different values in the range [0,1], and it is adjusted appropriately to obtain 100 percent accuracy for training data set with a minimum number of hyperboxes. Hence, these algorithms execute multiple passes over the training data for different values of the expansion coefficient. Moreover, usually, its value has to be retuned for the addition of new patterns or classes to the existing classifier. Consequently, as expected these algorithms do not learn in a single pass. This paper proposes an Unbounded Recurrent FMN (URFMN). In the URFMN, several modifications are proposed such as removal of the expansion coefficient parameter, membership functions for different types of hyperboxes, and metamorphosis from feed-forward to the recurrent topology with a novel learning algorithm which works in two phases- offline and online phase. Moreover, during the online phase, it allows the addition of new patterns or classes without the need for retraining. Hence it is an online adaptive algorithm. Its performance is evaluated using the benchmark data sets.