A Study Of Algorithms Of Neural Networks As Classifiers And Their Application In Text Classification

Based on the theory of statistics, this dissertation investigates neural network classifiers realized with software simulation in the computer. After a brief summary of the existing studies of the neural networks as classifiers and the problems that are still waiting to be solved, the dissertation begins to focus on the study of the cover algorithms in the constructive learning methods with the aim to shorten the learning time, simplify the network structure and improve the classification precision. The dissertation mainly deals with the following items.Firstly, the dissertation puts forward a probability model of cover algorithm, and, with the help of maximum likelihood estimation of finite mixtures of models, optimizes the cover algorithm with the expectation maximization algorithm. The dissertation takes the coverage result from the original cover algorithm as a pretreatment, and sees the coverage number of a certain class of samples as component numbers in the finite mixtures of models. In this way the dissertation solves the problem of calculating the number of components in the finite mixtures of models, while in the past researchers usually depended on their own subjective estimation. The dissertation holds that the number of coverage in a certain class of samples determines the number of components in the finite mixtures of models, and that every coverage of a class of samples could be seen as a Gauss distribution. Then, with the help of maximum likelihood estimation of finite mixtures of models, one could optimize the cover algorithm with the expectation maximization algorithm. Such a model extends the cover algorithm's range of application, and simulation result shows that the new algorithm has improved the examination precision.Secondly, the dissertation studies the relationship between parameters of the cover algorithms and the examination precision and reaches a new conclusion. That is, when the k-dimensional samples of the original space project onto the k+1 dimensions of the feature space, the radius of the hyper sphere, R, has nothing to do with the examination precision and the number of refused samples. Moreover, the paper studies the relationship between the value of the parameter n in the coverage...