【作者】 林关成； 李亚安； 李国辉；

【Author】 LIN Guan-cheng~(1,2),LI Ya-an~1,LI Guo-hui~1 (1.School of Marine Engineering,Northwestern Polytechnical University,Xi’an,710072,China; 2.Department of Media and Communications Engineering,Weinan Teachers College,Weinan 714000,Shaanxi)

【机构】 西北工业大学航海学院； 渭南师范学院传媒工程系；

【摘要】 针对支持向量回归(Support Vector Regression,SVR)的传统训练算法在大样本下内存开销大、训练速度慢的缺点,在分析SVR数学形式的基础上,充分利用二次规划自身的结构特点,将阈值量引入支持向量回归的目标函数中,采用扩展变形的方法,提出了一种用于SVR的简化连续过松弛(Successive Over Relaxation,SOR)训练算法,并将其与传统的训练算法进行对比。仿真结果表明,简化的SOR算法是可行、有效的。它在保证拟合精度的情况下,克服了传统训练算法对系统内存要求较高的缺陷,减少了训练时间,提高了运算效率。更多还原

【Abstract】 For the shortcomings of the traditional training algorithm for support vector regression under the large data sets,it proposes a simplified successive over relaxation training algorithm for SVR using expansion and deformation method based on the analysis of SVR mathematical form.The algorithm makes full use of the structural characteristics of quadratic programming and the threshold quantity was introduced in the objective function of support vector regression.Thus it compared with the traditional training algorithm.The simulation results show that the simplified SOR algorithm is feasible and effective.In the case of ensuring fitting accuracy it overcomes the shortcomings of higher system memory requirements for the traditional training methods,reduces the training time and improves the computational efficiency.更多还原