Improved Three-term Conjugate Gradient Algorithm For Training Neural Network

Abstract

A new three-term conjugate gradient algorithm for training feed-forward neural networks is developed. It is a vector based training algorithm derived from DFP quasi-Newton and has only O(n) memory. The global convergence to the proposed algorithm has been established for convex function under Wolfe condition. The results of numerical experiments are included and compared with other well known training algorithms in this field.