Multilayer Radial-basis Function Network and its Learning

The radial basis function (RBF) network may serve as a good alternative for multilayer perceptron (MLP), since RBFN structure is much easier, and learning speed is faster. This article aims to present the architecture of the multilayer radial basis function network and the usage of the backpropagation algorithm as one of the approaches for finetuning the weights of a neural network based on the error rate. Using Epanechnikov kernels as activations function allows to avoid the undesirable phenomena during the tuning like vanishing and exploiding gradient. In addition, because input signal of the last layer is linearly depends of the tuning synaptic weights, this allows optimize the speed of learning process. In result error backpropagation algorithm for multilayer network was proposed, whose layers in fact are radial-basis function neural network.