GRNN with Double Clustering

The hybrid combination of three techniques has yielded a pattern recognition and estimation technique with greatly improved training speed, orders-of-magnitude speed improvement for testing and readout, and sometimes improved accuracy as well. It is useful for problems with high dimensionality and noisy data. The techniques used are clustering, kernel regression with adaptive parameters, a second level of clustering, and the formation of a binary decision tree. When previous versions of general regression neural networks have been used as the system identification component for control systems, the most important problem has been that the estimation speed limits the iteration time in the feedback loop. The new hybrid technique was designed specifically to overcome this limitation.