基于Jeffreys非信息超先验的贝叶斯变量选择算法  

Algorithm of Bayesian Variable Selection Using Jeffereys' Noninformative Super Prior

作　　者：邓海松[1] 马义中[1] 邵文泽[2] 

机构地区：[1]南京理工大学经济管理学院,南京210094 [2]南京理工大学计算机科学与技术学院,南京210094

出　　处：《系统仿真学报》2011年第9期1801-1805,1818,共6页Journal of System Simulation

基　　金：国家自然科学基金(70672088)

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,筛选出对输出影响重要的输入变量是至关重要的。我们通过Jeffreys非信息超先验,提出了一种新的贝叶斯变量选择算法。不同于计算机试验中存在的变量选择算法,我们的方法不需要调节控制稀疏性的超参数。新的变量选择方法通过EM(expectation-maximization)算法求解,试验结果表明,我们的方法不仅取得了理想的效果,而且大大地减少了计算的负担。In computer experiments,simulation of complex phenomenon requires a large number of inputs and identifying the inputs which most impact the outputs is of crucial importance.A novel algorithm of Bayesian variable selection was proposed for computer experiments via a Jeffreys＇ noninformative super prior.Different from existed algorithms of variable selection in computer experiments,the proposed algorithm has no necessity to choose the sparseness-controlling hyperparameters.Implementation was carried out by an EM（expectation-maximization） algorithm and experimental results demonstrate that the new approach not only yields state-of-art performance but also has low computational cost.

关 键 词：计算机试验 变量的选择 元模型 稀疏性 插值 

分 类 号：O212.6[理学—概率论与数理统计]