Configuring the radial basis function neural network

The most important factor in configuring an optimum radial basis function (RBF) network is the training of neural units in the hidden layer. Many algorithms have been proposed, e.g., competitive learning (CL), to train the hidden units. CL suffers from producing dead-units. The other major factor Which was ignored in the past is the appropriate selection of the number of neural units in the hidden layer. The frequency sensitive competitive learning (FSCL) algorithm was proposed to alleviate the problem of dead-units, but it does not alleviate the latter problem. The rival penalized competitive learning (RPCL) algorithm is an improved version of the FSCL algorithm, which does solve the latter problem provided that a larger number of initial neural units are assigned. It is, however, very sensitive to the learning rate. This thesis proposes a new algorithm called the scattering-based clustering (SBC) algorithm, in which the FSCL algorithm is first applied to let the neural units converge. Then scatter matrices of the clustered data are used to compute the sphericity for each k, where k is the number of clusters. The optimum number of neural units to be used in the hidden layer is then obtained. The properties of the scatter matrices and sphericity are analytically discussed. A comparative study is done among different learning algorithms on training the RBF network. The result shows that the SBC algorithm outperforms the others