A novel granular neural network architecture

We introduce a novel granular neural network (GNN) architecture based on the multi-layer perceptron architecture. The GNN uses linguistic terms as connection weights, and uses the operations of linguistic arithmetic to update those connection weights. The GNN has been implemented in a Java-based simulation environment, with support for both regression and classification learning tasks. We present the results of a preliminary experimental comparison between the GNN and the c4.5 decision tree algorithm on two benchmark datasets. Our results show that the GNN was slightly more accurate than c4.5 on both datasets.