An Improved BP Neural Network Model Based on Quasic-Newton Algorithm

Aiming at the low learning rate, bad stability and local minimum problems in standard and some improved BP neural network, in this paper we proposes a novel BP neural network model which concentrates on two aspects: the choice of learning rate and the learning algorithm. In the new model we use Quasic-Newton algorithm to replace gradient descent algorithm or other learning algorithms, thus the new model not only avoids the local minimum problem but also mends the learning rate. On the other hand, the choice of the learning factor includes two keys, the expertise and the final output of neural network. By means of the two keys, we propose a kind of self-adaptive learning factor which can improve the learning ability and real-time learning ability of neural network. At last, several classical examples are utilized to validate the proposed new BP neural network. The simulations show the feasibility and validity of the proposed BP neural network compared with BP neural network based on gradient descent algorithm or Levenberg-Marquardt algorithm.