To solve the problem of the SVR( Support Vector Regression) automatic model selection, this paper presents an optimal approach based on a gradient descent algorithm for model parameters of a support vector machine for regression. This paper gets the local optimal model parameter by minimizing the model selection criteria R2w2 using a gradient descent algorithm over the set of parameters. Based on the Riemannian geometry, a conformal transformation suitable for SVR is proposed and kernel function is modified in a data-dependent way. The general ability is further enhanced as a result. Simulated results show that the approach is very effective.%为解决SVR(支持向量回归)自动模型选择的问题,提出一种基于梯度下降算法的支持向量回归机模型参数优化方法.通过最小化模型选择准则R2w2,对核参数集采用梯度下降算法得到局部最优的模型参数.依据黎曼几何为理论,提出一种适合于SVR的保角变换,对核函数进行数据依赖的改进,进一步提高SVR的泛化能力.仿真试验的结果验证了该方法的有效性.
展开▼