论文部分内容阅读
二阶工作集选择的SMO(顺序最小优化)算法是目前SVM(支持向量机)求解的高效率方法,然而实践中发现SMO算法在训练过程中依然存在训练时间过长的问题.针对这一问题,提出·一种目标函数值辅助的SMO改进算法,算法首先设计了目标函数值随训练迭代次数变化的经验性实验.经验性实验结果表明,该变化呈铰链函数形态,在一定的迭代次数后目标函数值在很长的一段时间里变化甚微,甚至出现微小的升降波动现象.基于上述实验结果,改进算法跟踪目标函数值的变化,待训练进入目标函数值变化曲线对应的水平区域后就终止算法.改进算法测试及k-CV实验表明,其在保证改进前预测能力的前提下,可以使训练效率提高至少20%.测试及k-CV(k分组的交叉验证)实验表明,改进算法能够保持改进前的预测能力.
The SMO (Sequential Minimal Optimization) algorithm selected by the second-order working set is an efficient method for solving SVM (Support Vector Machine). However, in practice, it is found that SMO algorithm still has long training time in training process. , An improved SMO algorithm based on the objective function value is proposed. The algorithm first designs an empirical experiment of changing the objective function value with the number of training iterations. The experimental results show that the change is in the form of a hinge function. After a certain number of iterations The objective function value changes little in a long period of time, and even slight fluctuations occur.According to the above experimental results, the improved algorithm tracks the change of the objective function value, and after it is trained into the horizontal area corresponding to the target function value curve Termination algorithm.The improved algorithm test and k-CV experiment show that the training algorithm can improve the training efficiency by at least 20% on the premise of improving the predictive ability.Experimental and k-CV (k-group cross-validation) experiments show that the improved algorithm Ability to maintain predictive power before improvement.