The study of learning algorithm the BP Neural Network based on extended BFGS method

In this paper, a new BP learning algorithm based on extended BFGS method is presented in order to get the solutions for some learning problems of traditional BP Neural Network, such as the slow rate of convergence and poor stabilization. Introduce the modified Newton descent method to the BFGS method which is used to obtain good variables. The single variable search, which the ordinary BFGS arithmetic often has to do, is not required in the extended BFGS method; but the optimized results can be reached step by step. Numerical test is carried out by the extended BFGS method and the others learning method of BP Neural Network, and comparisons of the results demonstrate that the new algorithm is better than the others in numerical stability and convergence.