基于核函数回归估计理论的软件可靠性预测建模引起诸多研究者的兴趣.此类研究中,核函数选择问题尤为重要.然而目前还很少有针对所给软件失效数据进行核函数选择或者构建核函数的工作.在14个常用软件失效数据集上应用配对t-检验对基于核函数理论的软件可靠性预测模型中核函数选择问题进行研究.使用的核函数回归估计方法包括核主成分回归算法、核偏最小二乘回归算法、支持向量回归算法、相关向量回归算法;核函数包括高斯核函数、线性核函数、多项式核函数、柯西核函数、拉普拉斯核函数、对称三角核函数、双曲正割核函数、平方正弦基核函数.实验结果表明:不同类型的核函数在不同数据集上表现差异较大,高斯核函数在所有数据集上表现较为稳定,预测结果最好.
The research of software reliability models based on kernel theory has arousing inter- ests of numerous researchers, and a quite remarkable study is how to choosing or constructing suitable kernel functions. By employing paired t-test, this paper compares the predictive perform- ance of different kernel functions in kernel based software reliability modeling on 14 distinct fail- ure data sets. The kernel functions used is as follows: Gaussian Function, Linear Function, Pol- ynomial Function, Symmetric Triangle Function, Cauchy Function, Laplace Function, Hyper- bolic Secant Function and Squared Sin Cardinal Function, kernel regression methods used inclu- ding: Kernel Principal Components Regression, Kernel Partial Least Squares, Support Vector Regression and Relevance Vector Regression. The experimental results show that there is great prediction discrepancy among different kernel functions and Gaussian Function suit best in all 14 failure data sets.