A quickly trained ANN with single hidden layer Gaussian units

Radial basis functions are used in approximation and interpolation problems and in artificial neural networks. Nonlinear radial basis functions at the single layer hidden units are effective in generating complex nonlinear mapping and, at the same time, facilitate fast linear learning. A model and an algorithm are proposed to arrive at a near optimum initial configuration very quickly. The position of the hidden units in the input space and the connection weights from the hidden units to the output units are optimally set. Simulations on this initial configuration are performed. Different parameters are further trained and their effects are studied experimentally