The Research And Simulation Of The Algorithm In Ascertaining The Structure Of The BP Network

The artificial neural networks developed quickly. Having been researched for many years, many kinds of artificial neural networks with different abilities to deal with information had been found and applied to many areas of information. BP neural network is used widely, it is reported that 80-90 percent of the neural networks are BP neural networks or its transmutation.It has been proved that the BP neural networks with three layers can implements any non-liner projection. But in actual training, the result is not satisfied because its Structure is not reasonable or the initial weights are inappropriate. To solve the problem and improve the accuracy and training efficiency, a method, dynamically spreading the hidden layers is designed in this article. In the process of training, a hidden layer can be added to the BP network according to the actual situation, the hidden layer adopted liner function, so it won't change the expression ability of the BP neural network. Considering the advantage of the agent arithmetic—easy to work with other arithmetic, and can find the result in global process, agent arithmetic is used to train the parameters of the neural networks.At the end of this article a correlative experiment is designed to test the correctness. A BP neural network is established according to the need of the enterprise and the relationship between each parameter in actual product process. Then the relative code of the program is written in Matlab environment to predict the value of the catalyst, and the process is simulated. According to the result, it's concluded that combining the agent arithmetic and the method of dynamic spreading the hidden layers is feasible, which will improve the precision of the BP neural networks and has an important meaning.