The Research Of BP Neural Network Method Base On Gray Theory

BP(Back Propagation) neural network has a very wide range of applications in function approximation, pattern recognition, classification, data compression and so on. However, BP neural network model for most of the problem of uncertainty. So, It is need to study on the information uncertainty and ambiguity for practical application.BP in the traditional network model, the training samples are often in the form of single-valued, but in practical applications, the attribute values of training samples is sometimes uncertainty. At this point, if it is still the traditional method of BP neural network training and prediction, it will significantly affect the network's generalization ability. At present, the training samples for the characteristics of the gray, many experts and scholars rough fuzzy neurons and neurons in an in-depth research, and has made some progress. However, the theoretical knowledge through a combination of gray, to deal with training samples of uncertainty research is still relatively small. Based on the above analysis, this paper proposes an approach based on gray neurons BP neural network method. The method is first analyzed the feasibility of gray neurons; theory and then combining the characteristics of gray, a gray neurons of the design method; Finally, according to the design of neurons to build gray BP neural network prediction model and an example of the model of rationality and effectiveness.Besides, In order to improve the study quality of BP NN effectively, it is necessary to think about the study training of NN from the new angle. Wu Yan, Wang Shoujue etc put forward the FBBP arithmetic. Put the input regulation of NN and the common reverse prevalence arithmetic together, and regulate the input vector data and authority result to minimize the NN's error function. However, in the process of regulation input vector, input vector result may increase ceaselessly and influence the study ability and extensive ability of network. So, the people put forward the new arithmetic based on FBBP-GBP arithmetic. The arithmetic not only put the authority result regulation method and input vector regulation method together, but also do the gray regulation to the input vector. After the decreasing of input, optimizing of NN input, The error of the FBBP arithmetic has been solved and the NN training is reasonable. So the effective study is come into true. After the experiment, it is found that the method is effective. It not only keeps the invariability of prime classification, but also has the advantage of quick training speed and high extensive ability.