Gradient descent training of the middle layer parameters in radial basis function neural networks
Radial basis function neural networks have been gaining popularity in recent years for use in approximating unknown functions. A drawback of using radial basis functions in a neural network is that a certain amount of trial and error goes into choosing the parameters of the basis functions such as the placement and the widths of the functions. Along with the time and effort that is involved, it is unlikely that using trial and error or some heuristic method to generate the parameters will find better than a fair approximation to the best possible values. This thesis discusses the use of gradient descent techniques to modify the basis function parameters during the training of the radial basis network. Using this technique, normalized error rates have been achieved on the Mackey-Glass time series data that compare favorably with other neural network training methods.
Thesis94.S331.pdf
1.4 MB
Adobe PDF
575731d5cf6ace7ab898fabf93d02d2c