Weighted Parzen windows for radial basis function network design

This paper shows how the weighted Parzen window (WPW) technique can be used for radial basis function network (RBFN) design. The WPW training algorithm uses an agglomerative hierarchical clustering procedure to find the RBFN centers and weights. This approach reduces storage requirements as it selects the centers and weights. It is shown that RBFNs can be designed using the WPW technique so that they are functionally equivalent to some statistical techniques. Experimental results are reported for two practical applications, laser-weld classification and handwritten character recognition. The results show that WPW designed RBFNs outperform some neural techniques in these applications