Research On Selective Ensemble Algorithm Based On Support Vector Machine

Support Vector Machine (SVM), which is based on statistical learning theory, has proved its performance in processing small samples, non-linear and high-dimensional data. With features of global optimization, simple structure and wide applicability, it has been applied in various fields. Yet it’s strongly affected by the selection of kernel function and parameters and it has some difficulty in high dimensional data processing. These defects may negatively afFect the stability and generalization of SVM.By far, a single SVM is usually used in the training procedure of SVM while studies on multiple SVMs are not plenty. By applying diverse classifiers, ensemble learning can greatly improve the generalization of the learning system. Selective ensemble learning has enhanced the performance not only on problems which require larger calculation amount and storage, but on increasing accuracy of classification of the ensemble system. Thereof, selective ensemble has become a major approach of the studies on SVM.This thesis focuses on the theoretical basis, implementation and adoption of SVM and ensemble learning. Given the defects of SVM on ensemble learning and selective ensemble learning, a selective ensemble learning of SVM based on harmony search algorithm (HSA) is proposed. Under this approach, an initial sample set is obtained through sampling the sample set based on bagging principle. The initial set is thereafter trained in SVM and the ensemble classifiers are subsequently selectively integrated through harmony search. Drought brings great loss to agricultural production. Since agricultural meteorological data is a kind of non-linear high-dimensional data, selective ensemble method can be applied to prediction of agricultural data. Based on data of agricultural meteorological data in Anhui province, a prediction model is developed to provide a novel method of drought forecast.