A granular neural network: Performance analysis and application to re-granulation

The multi-granularity problem is one of the key open problems in Granular Computing. Multiple descriptions of the same phenomena may use very different information granulations, complicating any comparison or synthesis of those descriptions. One method for solving this problem is to transform all observations to a common granulation; however, this granulation must be adequate to capture all important facets of the phenomena. Determining this “natural” granulation could be done by inductively learning and comparing multiple granular representations of the phenomenon, but this requires a dedicated learning architecture. We present the Granular Neural Network, a novel adaptive neural network architecture that employs granular values and operations at the level of individual neurons. The Granular Neural Network is based on the multiplayer perceptron architecture and the backpropagation learning algorithm with momentum. It uses the operations of linguistic arithmetic to manipulate granular connection weights, which are represented by linguistic terms. We test the performance of the Granular Neural Network on three well-known benchmark datasets, and then explore its use in determining the “natural” granularity of a dataset.

► A granular neural network (GNN) architecture using linguistic connection weights is proposed.
► The GNN is compared against multilayer perceptrons and the C4.5 decision-tree learner.
► On three benchmark datasets, the GNN is at least as accurate as the contrasts.
► An exploration of learning at multiple granularities shows the GNN can help find the “natural” granularity of a dataset.