A novel RBF neural network with fast training and accurate generalization

Authors:
Lipo Wang;Bing Liu;Chunru Wan
Affiliations:
College of Information Engineering, Xiangtan University, Xiangtan, Hunan, China;School of Electrical and Electronic Engineering, Nanyang Technology University, Singapore;School of Electrical and Electronic Engineering, Nanyang Technology University, Singapore
Venue:
CIS'04 Proceedings of the First international conference on Computational and Information Science
Year:
2004

Citing 7
Cited 1

Matrix computations (3rd ed.)

Matrix computations (3rd ed.)
Neural Networks: A Comprehensive Foundation

Neural Networks: A Comprehensive Foundation
Self-Organizing Maps

Self-Organizing Maps
Comparing support vector machines with Gaussian kernels to radialbasis function classifiers

IEEE Transactions on Signal Processing
Performance evaluation of a sequential minimal radial basis function (RBF) neural network learning algorithm

IEEE Transactions on Neural Networks
Selecting radial basis function network centers with recursive orthogonal least squares training

IEEE Transactions on Neural Networks
Analysis of input-output clustering for determining centers of RBFN

IEEE Transactions on Neural Networks

Global structural optimization considering expected consequences of failure and using ANN surrogates

Computers and Structures

Quantified Score

Hi-index	0.00

Visualization

Abstract

For the reason of all the centers and radii needed to be adjusted iteratively, the learning speed of radial basis function (RBF) neural networks is always far slower than required, which obviously forms a bottleneck in many applications. To overcome such problem, we propose a fast and accurate RBF neural network in this paper. First we prove the universal approximation theorem for RBF neural networks with arbitrary centers and radii. Based on this theory, we propose a new learning algorithm called fast and accurate RBF neural network with random kernels (RBF-RK). With the arbitrary centers and radii, our RBF-RK algorithm only needs to adjust the output weights. The experimental results, on function approximation and classification problems, show that the new algorithm not only runs much faster than traditional learning algorithms, but also produces better or at least comparable generalization performance.