Journal of Systems Engineering and Electronics ›› 2010, Vol. 32 ›› Issue (12): 2725-2729.doi: 10.3969/j.issn.1001-506X.2010.12.47

• 软件、算法与仿真 • 上一篇    

基于多层稀疏先验的回归样条变量选择算法

邓海松1,马义中1,邵文泽2   

  1. 1. 南京理工大学经济管理学院, 江苏 南京 210094;
    2. 南京理工大学计算机科学与技术学院, 江苏 南京 210094
  • 出版日期:2010-12-18 发布日期:2010-01-03

Variable selection for regression splines using hierarchical sparseness prior

DENG Hai-song1,MA Yi-zhong1,SHAO Wen-ze2   

  1. 1. School of Economics and Management, Nanjing Univ. of Science and Technology, Nanjing 210094, China; 
    2. School of Computer Science and Technology, Nanjing Univ. of Science and Technology, Nanjing 210094, China
  • Online:2010-12-18 Published:2010-01-03

摘要:

在计算机试验中,复杂现象的仿真拥有数目庞大的输入变量。因此,筛选出对输出有重大影响的输入变量显得至关重要。针对计算机试验的变量选择问题,提出一种基于贝叶斯多层稀疏先验的回归样条变量选择算法。新算法能够同时进行非重要输入变量的自动删除和重要输入变量系数的自适应估计。不同于计算机试验中已有的变量选择算法,新算法不需要调节控制稀疏性的超参数。通过快速算法进行数值求解,试验结果表明:新算法不仅能够更精确地实现变量选取,而且能够大大地降低计算复杂度。

Abstract:

In computer experiments, simulation of complex phenomena requires a large number of inputs, and identifying the inputs which make a notable impact on the outputs is of crucial importance. A new Bayesian variable selection algorithm is proposed for computer experiments via a hierarchical sparseness prior. The new algorithm is not only capable of deleting insignificant variables and estimating coefficients of significant variables simultaneously, but also has no necessity to adjust the sparseness-controlling hyperparameters. Numerical implementation is carried out by a kind of fast algorithm and experimental results show that the new approach not only yields more accurate variable selection but also is of low computational complexity.