From multilayer perceptrons to radial basis function networks: a comparative study

A special additional input, which is the sum of the squares of the other inputs, is added to the standard multilayer perceptron, so that the multilayer perceptron works similarly as the radial basis function network with localized response. Specially, we will show a three-layered multilayer perceptron with exponential activation function and this kind of additional input is naturally a generalized radial basis function network which can by trained with the well developed training strategies of multilayer perceptrons. A comparative study is also conducted between multilayer perceptron, with additional inputs and radial basis function networks trained by various methods.