Keerthi的SMO算法的偏置计算改进

指出Keerthi的SMO算法存在的问题。该算法由于采用“取中法”求偏置，在优化条件不满足的情况下，偏置值有可能出现偏差，从而劣化SVM的建模性能。该文从SVM回归的原问题出发，导出求偏置的新方法并将其归结为一维凸函数最优化问题，将新算法应用于高斯函数的回归和记忆非线性功率放大器的预失真器的建模中，结果显示了新算法的正确性和有效性，建模精度提高10%左右。

Improvement of Bias Computing on Keerthi's SMO Algorithm

A source of error in Keerthi’s Sequential Minimal Optimization(SMO) regression algorithm is pointed out. The bias value is acquired by ‘median-finding’ technique in this algorithm, when the optimality condition does not hold, the bias value may be inaccurate, and make the Support Vector Machine(SVM) model performance worse. By analyzing the primal problem of SVM regression, a new kind of computing method is derived to get bias, and proved to be an optimization problem of one dimension convex function. The modified algorithm is confirmed to be correct and effective when it is used to estimate Gaussian function and model predistorter of nonlinear memory power amplifier, the SVM model trained by the new algorithm performs more accurately by 10% than Keerthi’s one does.