Determining Radius of RBF for Improving Convolutional Radial Basis Function Network (CRBFN)

抄録

Aim of this study is implementing a more precise method for function approximation when dealing with large-scale problems. In this study, we conducted experiments by using Golinski’s Speed Reducer (GSR) benchmark problem (7 design variables). In Convolutional Radial Basis Function Network (CRBFN), with the increasing of the convolution times, the radius of basis centers appear an accelerating upward trend, while the result of evaluation criterion for each iteration shows a slowing downward trend. For they have opposite characteristics, we attempted to use convolutional ways to combine them together. In this experiment, we used 520 sets of data. Each data set includes the design variables, the corresponding function values, and the constraints on the design variables. We used these data sets to train CRBFN and our network to approximate the target function, and the results proved our method was always better.