Support vector machine (SVM) is a novel pattern recognition technique. Due to its complete theoretical basis as well as excellent performance, SVM has become a hotspot in the area of pattern recognition. In practical applications, previous SVM models exhibit some deficits such as too loose bound on the structural risk, lack of self-adaptability to the data set during the learning course, and strong sensitivity to the noise.This thesis is focused on the study of SVM, and has acquired the following achievements.1. A full survey is provided on the history and state-of-the-art of SVM, with emphasis on its theoretical system as well as its various models.2. In order for an SVM to be more robust to noise, a new SVM model (i.e., the support vector machine based on adjustive boundary (SVMAB) is proposed. Then, the dual objective function of SVMAB is obtained by calling the Lagrange Theory. According to the errors gained during the training course, SVMAB treats different samples in different ways. In this way, the effect of noise on the training of SVMAB is well controlled. Through universal simulations on man-made data with or without noise points as well as those of Titanic and Breast Cancer, we conclude that SVMAB outperforms the classical L1-SVM in that it needs less supporting vectors but enjoys better generalization performance. |