Rule Extraction From Neutral Network And The Applications In The Early-warning Of Enterprise

For a long time, the knowledge which is studied by neutral network is called hidden knowledge or hidden rules for its difficulty to express.So neutral networks is also named 'black boxes'.The express of neutral networks is a hot topic in this field. People not only want the neutral network to implement the mapping from inputs to outputs, but also want to know what is the knowledge hiding in the shade of network, that is, to get rules automatically from data in order to use it again in the practice.We ever use linear function to approximate sigmoid function and extract the relation between inputs and outputs within the given precision. But when the question is nonlinear, its precision is worse.So we learn from others to improve it. When the inputs and outputs are continuous, we use the subsection linear function to approximate the sigmoid function in the multi-layer perceptions (MLPs). We implement the algorithm which includes the subsection linear theory; pruning techniques of the multi-layer perceptrons and extraction rules methods. We also make use of it in early warning of enterperise. We draw some useful conclusion.The paper presents a new method to extract rules from trained neutral networks whose input vectors and output vectors are continuous. On the base of studying on the MLPs, analyze more using the subsection linear theory. The paper expresses the relation ship between variables by means of knowledge, and gives clear explanation.