GMDH and RBFGRNN networks for multi-class data classification

This work explored the application of Group Method of Data Handling (GMDH) and Radial Basis Functions Generalized Regression Neural Network (RBFGRNN) for a multi-class data classification with real valued attributes. The Dataset employed has 336 instances, with each instance consisting of a name and seven predictive inputs and one of five different classes. The Karhunen Loeve Transformation (KLT) technique is used to avoid redundancy in the data keeping more than 95% of the information in the original data. A four layer GMDH network with six best error outputs passed to the succeeding layer gave the best result. Before applying a multidimensional RBFGRNN network, a DIANA like hierarchical clustering algorithm has been employed to determine the number of cluster centers and organize the data around those centers. Finally, the multi-dimensional RBFGRNN is applied on the transformed data and the performance of the network is evaluated The regressed output of the network highly resembles the actual desired class outputs.