An Improved Algorithm Using B-Spline Weight Functions for Training Feedforward Neural Networks

Abstract:

An improved algorithm using B-splines as weight functions for training neural networks is proposed. There is no need for training neural networks or solving linear equations. The most important advantage is that we can get the forms of weight functions by the given patterns directly. Each of weight function is a one-variable function and takes one associated input point (input neuron) as its argument. The form of each weight function is a linear combination of some B-splines defined on the sets of given input variables (input knots or input patterns), whose coefficients are associated with the given output patterns. Some examples are presented to illustrate good performance of the new algorithm.

• Periodical: Applied Mechanics and Materials (Volumes 278-280)
• Pages: 1301-1304
• DOI: https://doi.org/10.4028/www.scientific.net/AMM.278-280.1301
• Online since: January 2013
• Authors: Dai Yuan Zhang
• Keywords: Algorithm (GA), Artificial Intelligence (AI), B-Spline, B-Spline Weight Function, Neural Network (NN), Noninterpolatory Approximation, Weight Function

References

[1] D. Y. Zhang. New theories and methods on neural networks [M]. Beijing: Tsinghua University Press, 2006. (in Chinese).

[2] D. Y. Zhang. New algorithm for training feedforward neural networks with cubic spline weight functions [J]. Systems Engineering and Electronics, 2006, 28(9): 1434-1437. (in Chinese).

[3] D. Y. Zhang. New algorithm for training neural networks based on generalized Chebyshev polynomials [J]. Systems Engineering and Electronics, 2008, 30(11): 2274-2279. (in Chinese).

[4] D. Y. Zhang. A new Algorithm of Neural Networks with B-Spline Weight Functions [C]. / Proceedings/2010 International Conference on Artifical Intelligence and Education. Hangzhou: IEEE Press, 2010. 782-785.