基于信息冗余检验的支持向量机时间序列预测自由参数选取方法

Redundancy-test-based hyper-parameters selection approach for support vector machines to predict time series

支持向量机建模中的一个关键和难点问题是自由参数的设置.不同于以往应用残差的简单统计量选取最佳模型的方法,本文提出通过检验模型在训练集上的拟合残差是否不含冗余信息作为选择自由参数的依据.进一步提出应用全向相关函数(omni-directional correlaton function,ODCF)检验残差信息冗余并给出应用方法,并从理论分析和数值仿真两方面给出该方法正确性的证明.在两个典型的非线性时间序列(年均太阳黑子数和Mackey-Glass数据)上进行了实验,实验结果优于相关文献记载及基于校验集方法的预测性能.

The selection of hyper-parameters is a crucial point in support vector machine modeling. Different from previous method of choosing an optimal model by using basic statistics of residuals in, the new approach selects hyper-parameters by checking whether there is redundant information in residual sequence. Furthermore, omni-directional correlation function （ODCF） is used to test redun- dancy in residual, and the accuracy of the method is proved by theoretical analysis and numerical simulation. Experiments conducted on benchmark time series, annual sunspot number and Mackey-Glass time series, indicating that the proposed method has better per-formance than the recorded in the literature.