一种改进的Bayesian网络结构学习算法

被引:13
作者
羌磊
肖田元
乔桂秀
机构
[1] 清华大学自动化系,清华大学自动化系,清华大学自动化系北京,北京,北京
关键词
Bayesian网络; 数据挖掘; 知识发现; 机器学习;
D O I
暂无
中图分类号
TP181 [自动推理、机器学习];
学科分类号
081104 ; 0812 ; 0835 ; 1405 ;
摘要
基于模型选择的 Bayesian网络 ( BN)结构学习是 NP难的可行解搜索过程 .针对现有算法在复杂系统求解中时间效率低的问题 ,提出了一种新的基于最小描述长度 ( minimal description length)理论的结构学习算法 I-B &B-MDL .这种算法将独立性测度与预测估计相结合 ,在学习过程中引入小计算量的独立性测试为 MDL搜索提供启发性知识 ,限制可行解搜索空间 ,从而加速问题求解过程 .针对新算法讨论了改进策略对求解精度的影响 ,并结合算例分析了独立性测试的阶数选择问题 .通过对一实际问题进行验证表明 ,在保证结果精度的前提下 ,新算法在时间性能上比仅基于预测估计的 B & B-MDL 有较大改进
引用
收藏
页码:1221 / 1226
页数:6
相关论文
共 12 条
[1]  
Modeling by shortest data description. J Rissanen. Automatica . 1978
[2]  
A th eory of inferred cau sation. JPearl,T Verm a. Proc ofthe2 ndInt’’lConf . 1991
[3]  
Th e recovery of causal poly- trees fromstatistical data. G Rebane,J Pearl. Uncertainty inA rtificialIn telligence . 1989
[4]  
Causation, prediction and search. P Spirtes,C Glymour,R Scheines. In: Lecture Notes in Statistics . 1993
[5]  
A new approach for learning belief networks using independence criteria. M Luis,de Campos,J Huete. International Journal of Approximate Reasoning . 2000
[6]  
Selection of the stochastic model based on the minimum description length principle and state decomposition. J Suzuki. Journal for Information Processing Society . 1992
[7]  
Learning Bayesian belief networks based on the MDL principle: An efficient algorithm using the branch and bound technique. J Suzuki. IEICE Transactions on Information and Systems . 1999
[8]  
Using hidden nodes in Bayesian networks. C K Kwoh,D F Gillies. Artificial Intelligence . 1997
[9]  
Objective probabilities in expert systems. L E Sucar,D F Gillies,D A Gillies. Artificial Intelligence . 1993
[10]  
Approximating discrete probability distributions with dependence trees. C K Chow,C N Liu. IEEE Transactions on Information Theory . 1968