Abstract
In this thesis, we propose a pruning procedure base on the concept of Support Vector Machine（SVM）to discover a subset that possess the characteristics of some learning function form a large data set. In this procedure, three major stages of pruning are implemented for a given data set, we identify the quadrant of learning function firstly, then we calculate the central point, and take the coordinate axes of the central point to be our first separate hyperplane. Based on this separate hyperplane, we classify the data into two clusters and prune away unnecessary data. Second, based on these asymmetrical characteristics of the learning function, we apply regression analysis to obtain the regression line of the remaining data to be our second separate hyperplane and prune away unnecessary data. Under the condition of minimum error, the Nelder-Mead simplex method is used to obtain the learning function of the remaining data. At the third stage, we use the residual analysis to evaluate the resultant function. Three levels of standard deviation are used to prune unnecessary data. Then, we use the Nelder-Mead simplex method again to obtain the best-fit learning function. And the trade-off between accuracy and pruning data is carried out. Finally, we can obtain the best-fit learning function of the remaining data. Three models were used to demonstrate the proposed pruning procedure and the results are promising in application.