A synthetically backpropagation algorithm

Kejun Wang,Hongzhang Hongzhang,Guobin Li
1999-01-01
Abstract:This paper presents a synthetically backpropagation algorithm for multilayered forward neural networks. A new general index function that consider the effect of absolute error and relative error on NN learing and performance and the back-propagation technique based on searching output space are proposed and used in the algorithm. The algorithm has both a dynami-cal adaptive regulation learning rate and a variable momentum coefficient, and has ability of self regulation active characteristic, eliminating flat phenomenon and convergence no equilibrium phenomenon during training. The contrast experiments indicate that the algorithm has more fast convergence speed than BP algorithm and can achieve a global optimal solution.
What problem does this paper attempt to address?