The Study Of Algorithm Of BP Neural Network

The learning algorithm of neural network has always been an important problem in both research and application fields of artificial neural networks, especially to the study of the learning(design) of feedforward neural networks. The backpropagation neural network(BPNN) is a kind of multi-level feedforward neural networks, and it is one of the most typical models of artificial neural network which is widely used. But, there isn't any fixed pattern or formula can be used for setting the structure of the BP neural network. In addition, although the learning algorithms of the BP neural network have been greatly improved in many aspects they still have many disadvantages such as local minimum easy,slow convergence, and so on. These problems severely limit the performance of BP neural network and the promotion of its application. Therefore, the paper proposed a new learning algorithm of the BP neural network based on the adjustment of weight and threshold value,which has some improvement in terms of convergence speed.There are two parts in the contents of this thesis. The first part mainly introduces how to select the training sample of BP neural network, especially for the high-dimensional data. The selection of BP neural network training sample has a strong impact on the generalization ability of the network, and the selection of training sample from the high-dimensional data is especially difficult. The paper uses the method of factor analysis to pretreat the data from a large sample, and then clustering the result. In this way, both of the dimensions of data noise and variables of sample can be reduced, besides the training sample selected from the large sample will has almost the same characteristics as the large sample, and finally, both of the dimensions of variables and the number of samples can be reduced.In the second part, the paper proposed a BP algorithm based on a partial adjustment of the weight and threshold value. According to the characteristics of biological neuron in learning and memory formation, only some neurons were stimulated to produce the output for the specific training samples, while the other part of the neurons weren't stimulated. There are large difference between this part of the neuron's output and target, and then we need this part neurons weight and threshold value to adjust. Therefore the algorithm proposed in this paper only adjust the weight and the threshold value of the local neurons, and this can accelerate the learning speed of the network.