What are the main factors involved in the design of a Radial Basis Function Network ?

摘要

Many papers have been published concerning the obtention and analysis of new algorithms for neural network learning, and in particular the paramenter optimization of a Radial Basis Function network system (RBF). However, there has been no analysis in the influence in RBF design for the selection of the following factors: which of the many avalible functions to select for the non-linear functions within the hidden neurons; selection of distances; number of neurons in the hidden layer; number of bits assigned to weight resolution, etc. The present statistical study was motivated by the great variety of alternatives that a designer has to take into account when developing an RBF. Thus, complementing the existing intuitive knowledge, it is necessary to have a more precise understanding of the significance of the different alternatives. In the present contribution, the relevance and relative importance of the parameters involved in such the design are investigated by using a statistical tool, the ANalysis Of the VAriance (AN0VA)[2]. In order to analyze the behaviour of the RBF, three different groups of examples were used: time series forecasting, approximation of functions from samples and classification problems.