Research Of Classification Ability And Training Algorithm Of Feedforward Neural Network

This thesis is supported by National Foundation of Science (No. 60071026) and NUDT fund (No. OOJI .4.4.DZ0106). As an emerging subject, artificial neural network (ANN) has been seen vast development ever since start, and been widely used in the fields of pattern recognition,optimization, image process, signal process, prediction etc, and provided novel and effective method for resolving relevant problem in these fields. As an emerging subject, the fundamental theory of artificial neural network is still in the stage of development, such as classification ability, selection of structure, selection of learning algorithm, network implementation etc. All these put restriction not only on the application of neural network in practice, but also on the neural network itself for further development. As the basis of other kind of neural network, feedforward neural network(FNN) have always been the model that people study most and apply widest. And the main aspect of the FNN Application is pattern recognition and classification, so we chose the classification ability and training algorithm of FNN as the main subject of this paper.The paper firstly summarizes the history of ANN, discussed several problem of the application of ANN in the field of pattern recognition. Then preliminary study was carried on the classification ability and network structure of FNN by constructing fundamental decision boundaries in experimental simulation. These conclusions will help further studies on the classification mechanism and classification ability of FNN. After that, the initial method try to avoid most of local minima by properly apply reverse operation to the output neurons of the FNN is provided. Then the question of robust classification is discussed, we studied the computational complexity of base-set robust perceptron under the condition of large training set, and the incremental learning algorithm is provided. At last, based on the base-set robust perceptron and SVM, a geometric training algorithm is provided for support-vector based robust perceptron (SVRP) to resolve linear separable problem. Experimental simulation indicts that its robustness and speediness.The key techniques used in this thesis are: ①computer simulation; ②incremental learning; ③statistical learning theory; ④linear proramming; ⑤BP training algorithm; ⑥linear space theory.In this paper, studies were carried on the classification ability and network structure of FNN by constructing fundamental decision boundaries in experimental simulation. It lacks theoretical deduction, and the number of constructed decision boundaries is not enough. Constructing more and common decision boundaries and theoretical inducation are needed to obtain theoretic conclusion on classification ability and structure selection. The support vector robust perceptron can only be applied to linear separable problem; the extending of it to nonlinear separable problem will be a new research direction in future.