Training Neural Networks by Continued Fractional Weight Functions

Abstract

To solve the training problems for complex problems, a new learning algorithm based on continued fractional weight function for training neural networks is proposed, which is different from that of polynomial weight functions. The continued fractional weight functions using interpolation methods are more suitable for training the patterns obtained by some rational problems. The analysis of generalization is also presented in this paper. At last, to illustrate the power of the new learning algorithm, a simulation example is presented to show that the new algorithm proposed in this paper has good performance both on generalization and calculating precision.