摘要
在序列最小优化(Sequential Minimal Optimization,SMO)算法训练过程中,采用标准的KKT(Karush-KuhnTucker)条件作为停机准则会导致训练后期速度下降。由最优化理论可知,当对偶间隙为零时,凸二次优化问题同样可以取得全局最优解。因此本文将对偶间隙与标准KKT条件同时作为SMO算法的停机准则,从而提出了改进停机准则的SMO算法。在保证训练精度的情况下,提高了SMO算法的训练速度。通过对一维和二维函数的两个仿真实验,验证了改进SMO算法的有效性。
In SMO training process, standard KKT stopping criteria will lead to training speed declining along with training progress. According to the optimum theory, if the dual gap is zero, convex quadratic optimization problem will also obtain global optimal solution. Therefore, an improved stopping criteria of SMO is proposed in this paper, it combines the duality gap and standard KKT condition as the stopping criteria. This algorithm can improve the training speed without training accuracy decrease. Two cases experimental simulating results corroborate the efficiency of this algorithm.
出处
《计算机工程与应用》
CSCD
2014年第16期31-34,61,共5页
Computer Engineering and Applications
基金
国家高技术研究发展计划(863)(No.2009AA05Z203)
江苏高校优势学科建设工程资助项目
关键词
支持向量机回归
序列最小优化算法
对偶间隙
KKT条件
停机准则
Support Vector Regression
Sequential Minimal Optimization （SMO）
duality gap
Karush-Kuhn-Tucker（KKT）
Stopping Criteria
