参数自适应决策的多分辨率核支持向量分类器

Multi-Resolution Kernel Support Vector Classifier Based on Parameters Automatic Selection

为克服经典支持向量分类器(SVC)训练算法中参数的选择需要多次人工调整的缺陷,本文提出了基于多分辨率核的支持向量机参数自适应调节策略。首先通过分析非线性核映射的特征空间超平面的最小VC维数,提出了多分辨率核函数参数的自适应优化准则。然后通过迭代求解获得最优泛化能力的多分辨率核参数数值。多分辨率核函数方法保持了经典SVC训练算法结构风险最小化的原则,克服了经典SVC选择单一参数的缺陷。仿真实验结果表明本文提出的算法能够自适应的选择合适的核参数达到最优泛化能力。

In this paper in order to solve the problem of manual selection parameters in classical support vector classifier, we proposed a multi-resolution parameterized kernel method in support vector classifier to slelect optimal kernel parameters automatically. First- ly a multi-resolntion kernel parameters optimal scheme is presented by deeply analyzing the high dimensional feature space produced by nonlinear mapping. Then the optimal kernel parameters could be obtained by iterative computing. This novel method not only preserves the excellence of structural risk minimization that existing in general support vector classifier, but also overcomes the limitation of repeating manual definition kernel parameters in classical support vector classifier. The experimental results prove that the new multi-resolution kernel support vector classifier is valid and efficient.