The Research On The Design Of Fuzzy Classifier Based On Neural Networks

Pattern recognition problems of large scale(PRPLS) refer to those with a feature space of high dimensionality and a data set with large amount of samples that belong to many different classes. Many practical pattern recognition problems belong to this type, which means the decision boundaries in these problems are usually very complex and it is difficult and time consuming to build the corresponding classifiers.Recently a covering method for feed-forward neural network design has been proposed by professor Zhang Ling. Based on the sphere neighborhood model, this method transforms the design of neural classifiers to a geometrical covering problem. With this method, it is direct and convenient to construct a neural classifier. Experimental results show that the covering method is promising for solving PRPLS.Fuzzy theory is considered to be most fit tool to solve many artificial intelligence problem. From the beginning, pattern recognition plays an active role in the application of fuzzy theory. This paper try to combine fuzzy theory with the sphere neighborhood model, and then put forward a fuzzy algorithm based on CSN(covered with sphere neighborhood). The performance of sphere neighborhood model is discussed and get deeply into in this paper.The decision boundaries in PRPLS are usually complex and sometimes it is difficult to get ideal result with only a single classifier. So it is necessary to build multi-stage classifiers, which gradually determines the class of a given sample with multi-steps. This paper designs a two-stage classifier, which is based on the combination of the covering approach, fuzzy set theory and the possible set based method. Experimental results show that this classifier has a good performance.