Adaptive RBF neural networks for pattern classifications

The viewpoints are presented that the centers and widths of kernels in RBF networks should be determined by a self-learning procedure, that a new kernel naturally comes into being according to which class some labeled patterns are misclassified to, and going a step further, that a current kernel be deleted if its effect on the test set is too trivial to be worthy of mention. As a result, a kind of cascade RBF-LBF networks consisting of a single-layer RBF and LBF ones are proposed. A classification application shows that the proposed adaptive algorithm is able to optimally determine the structures and parameters of the RBF-LBF networks in accordance with the characteristics of sample distribution, has higher convergence rate and classification precision as well as many other advantages, compared with the feedforward two-layered LBF and RBF networks. The cascade RBF-LBF networks have a clear advantage for dealing with such questions as multiple distribution regions and irregular shapes for one class in multi-dimension spaces