The high dimension of the learning error function for BP networks and the difficult computation complexity are incestigated. A simple modified conjugation-gradient decent algorithm (MPARTAN) is proposed based on improving the gradient BP algorithm. That the computation complexity of this algorithm is not higher than that of the BP momentum algorithm. Compared with FR conjugation algorithm, this algorithm has better stability and fast speed quality of convergence. It is also investigated that the convergence theorems for this algorithm and comparison of the computing results by two computing examples for the promoted three algorithms: BP momentum algorithm, FR conjugation-gradient algorithm and the novel MPARTAN algorithm.
umeihart D E, Hinton G E, Williams R J. Learning internal representations by error propagation[A].In:Rumelhart D E, McClelland J L, eds. Parallel Distributed Proceeding[C]. Cambridge MA:MIT Press, 1986. 318��362.��
umeihart D E, Hinton G E, Williams R J. Learning representations by back��propagating errors[J].Nature.1986,323(6088):533-536
echt��Nielsen R. Theory of the backpropagation neural network[A].In:Proceedings of the 1989 International Joint Conference on Neural Networks [C].NewYork:IEEE Press,1989.593��599.��
hah B V , Buehler R J, Kempthorne O. Some algorithms for minimizing a function of several variables[J].J Soc Indust Appl Math.1964, 12 (1):74-92
uang Deshuang. An analysis of structure properties for feedforward neural networks[A].In: 1998 Int Conf on Neural Networks and Brain Proceedings[C]. Beijing: Publishing House of Electronics Industry, 1998. 463��466.