Research On Acme-set Of Convex Hull In Support Vector Classification

SVM is a potential learning machine that is newly developed. It becomes hotspot of research due to its excellent capability in the pattern recognition and regression estimation field. SVC is the application of SVM on pattern recognition.In order to achieve SVM, it is needed to figure out the quadric optimization problem with constraints. For a SVC with n samples, n variables are available, and a n × n scale matrix is needed to be stored in EMS memory once at a time. But if n is very large, it's hard to do so. In order to reduce the amount of training samples and improve learning speed, we suggest using the acmes set instead of the whole sample set to train a SVC, which is based on the traits of the sample acmes that is representative for classification and of small amount.To implement the idea above, an algorithm for extracting acmes is proposed. Furthermore, it is extended to the kernel feature space. Based on the acme set, a criterion of two-kind separability is put forward. Experiments indicate that the algorithm and the criterion are feasible. Meanwhile, performances of the SVC trained by the acme set almost equal to the one trained by the whole sample set while the former one is faster than the latter one.