A multiclass neural network classifier with fuzzy teaching inputs

Abstract

A multiclass neural network classifier with fuzzy teaching inputs is proposed. The classifier creates each class by aggregating a fuzzy prototype and several fuzzy exemplars in the hidden layer. Fuzzy inputs and all the nodes in the hidden layer are represented by trapezoidal fuzzy numbers. The classifier is trained by a two-pass learning algorithm. In pass one, a very fast one-epoch algorithm PECFUH (Prototype Expansion and Contraction of FUzzy Hyperbox) or FUNLVQ (FUzzy Number's Learning Vector Quantization) is used to train the prototypes. These prototypes will classify as many fuzzy input intances as possible. Afterward, exemplars that mean the exceptions, like the “holes”, in pattern space will be generated and expanded in pass two to classify those fuzzy input instances that cannot be correctly classified by the prototypes. A few-epoch FENCE (Fuzzy Exemplar Nested Creation and Expansion) training algorithm is proposed to create the exemplar nodes. Due to the training in pass one, the number of exemplar nodes is reduced and the learning speed is very fast during pass two. In addition, on-line adaptation is supplied in this model and the computational load is lightened. Also, nonlinearly separable instances and overlapping classes can be handled well. Furthermore, this classifier has good generalization ability for the training instances with don't-care information. The experimental results manifest that the training and recalling are fast. At the same time, they illustrate that required nodes are few.