本文介绍了神经网络中应用广泛的BP算法,讨论了BP神经网络的优点、存在的问题及其训练过程, 并在此基础上用MATLAB实现了数字识别的仿真。最后,本文还比较了几种改进的BP训练算法。

BP Neural Network principle and MATLAB Simulation

Xiong Xin Nie Mingxin

School of Information Engineering School of Information Engineering Wuhan University of Technology Wuhan University of Technology Wuhan, P. R. China 430070 Wuhan, P. R. China 430070 niemx@sohu.com

Abstract

This paper introduces the prevalent BP algorithm in neural network, and discusses the goodness、problem and training process of BP neural network, as well as using MATLAB software to simulate the numbers on the basis of it. At last, several improved BP training algorithms have been compared in the paper.

Keywords: BP neural network; number recognition; MATLAB

1 Introduction

The development of neural network is rapid since the first neural network model——

[1]

MP model came up in 1943. Hopfield neural network proposed in 1982 and opposite phase broadcast algorithm proposed by Rumelhart in 1985 make the neural network of Hopfield model and multilayer feedforward model to be the prevalent neural network model. They are effective in many applications of fields such as speech recognition, mode recognition, image processing and industry controlling.

Neural network is an theory which is imitative of the biological processing model to get the function of information intelligent processing. It treats with the pattern information which is hard to be expressed in certain language by the method from bottom to top and parallel distribution way formed by self-study, self-organization and non-linear dynamics. Neural network is a parallel and distributed information processing network architecture. It is generally composed by massive neurons, each of which has only one output that can connect many other neurons. The reciprocity between neurons is embodied by their connected weighs. The output of neurons is its input function. The types of functions in common use have linear function, Sigmoid function and value function.

There are two phases of positive transmitting processing and error reverse

transmitting processing in the study

[2]

processing of BP neural network. The signal inputted from outside spreads to the output layer and gives the result through processing layer for layer of neurons in input layer and hidden layer. If the expected output can’t be obtained in output layer, it shifts to the conversed spreading processing and the true value and the error outputted by network will return along the coupled access formerly. The error is reduced by modifying contacted weight value of neurons in every layer and then it shifts to the positive spreading processing and revolves iteration until the error is smaller the given value. Take a three layer network for example, the network is composed of N input neurons, K hidden neurons and M output neurons(as showed in fig.1). O2pm and O1pk are the output value of output layer and hidden layer respectively. w2km and w1nk are the connected weight value from the hidden layer to the output layer and from the input layer to the hidden layer respectively. Suppose the input studying sample is Xpn, so its corresponding expected output value is tpm.

Fig.1 BP neural network configuration