Evolving granular classification neural networks

The objective of this study is to introduce the concept of evolving granular neural networks (eGNN) and to develop a framework of information granulation and its role in the online design of neural networks. The suggested eGNN are neural models supported by granule-based learning algorithms whose aim is to tackle classification problems in continuously changing environments. eGNN are constructed from streams of data using fast incremental learning algorithms. eGNN models require a relatively small amount of memory to perform classification tasks. Basically, they try to find information occurring in the incoming data using the concept of granules and T-S neurons as basic processing elements. The main characteristics of eGNN models are continuous learning, self-organization, and adaptation to unknown environments. Association rules and parameters can be easily extracted from its structure at any step during the evolving process. The rule base gives a granular description of the behavior of the system in the input space together with the associated classes. To illustrate the effectiveness of the approach, the paper considers the Iris and Wine benchmark problems.