摘要： 使用SVM进行分类，超参数的选择非常重要，它直接影响分类的性能。在实际应用中，最优SVM算法参数选择还只能是凭借经验、实验对比、大范围的搜寻或者利用软件包提供的交叉确认功能进行寻优。而拟牛顿算法，可在一个校验集上最小化一个经验误差估计来优化SVM的超参数，使超参数在分类任务中达到最优值，从而取得理想的分类结果。该文对拟牛顿算法进行了探讨，并将其应用在基于SVM的羽绒识别系统中，实验结果表明，该算法是有效的，与未经过超参数优化的SVM分类器相比，羽绒的识别率有了较大提高。

关键词: 支持向量机, 超参数, 拟牛顿算法, 经验误差估计

Abstract: Support vector machines (SVM) are efficient in a large number of real-world applications. However, the classification results highly depend on the parameters of the model. For any SVM classification task, the best values for these parameters are usually picked by experience, experiment compare and large-scale search, or using cross validation providing by software package to optimize. For optimizing SVM hyper-parameters, this optimization scheme minimizes an empirical error estimate using a quasi-Newton optimization method on the validation set. The method shows satisfactory results in feather and down category recognition.

Key words: Support vector machines (SVM), Hyper-parameters, Quasi-Newton optimization method, Empirical error estimate

中图分类号: 

• TP391