摘要
着重研究了小数据集条件下结合凸约束的离散贝叶斯网络(Bayesian network,BN)参数学习问题,主要任务是用先验知识弥补数据的不足以提高参数学习精度.已有成果认为数据和先验知识是独立的,在参数学习算法中仅将二者机械结合.经过理论研究后,本文认为数据和先验知识并不独立,原有算法浪费了这部分有用信息.本文立足于数据信息分类,深入挖掘数据和先验知识之间的约束信息来提高参数学习精度,提出了新的BN参数学习算法—凸约束条件下基于数据再利用的贝叶斯估计.通过仿真实验展示了所提算法在精度和其他性能上的优势,进一步证明数据和先验知识不独立思想的合理性.
In this paper, parameters learning of discrete Bayesian networks（BNs） with small data sets with convex constraints is investigated, and the main task is improving the accuracy of parameter learning through offsetting the lack of data with prior knowledge. Data and prior knowledge are often mechanically integrated in most existing algorithms because they are treated independent. However, after a theoretical study, they are found dependent on each other, and the existing algorithms have dissipated this relevance. A novel parameter learning algorithm — Bayesian estimation based on data reutilization under convex constraints, is proposed with deeply mining the information between data and prior knowledge based on classification of data information. Finally, simulations demonstrate the advantages of novel algorithm in precision and other indexes, which in turn tells the dependance between data and prior information.
出处
《自动化学报》
EI
CSCD
北大核心
2015年第12期2058-2071,共14页
Acta Automatica Sinica
基金
国家自然科学基金(60774064
61573285)
教育部博士点基金(20116102110026)资助~~
关键词
贝叶斯网络
参数学习
小数据集
数据信息分类
数据再利用
Bayesian network（BN）
parameter learning
small data sets
classification about data information
reutilization of data
