摘要
在贝叶斯网络结构学习的过程中,如何采集先验信息并合理利用它对于构建准确的网络结构非常重要。鉴于此,依据有先验信息的贝叶斯网络结构学习的三个环节:先验信息的采集、先验信息的融合和网络结构的优化,首先讨论了现有先验信息获取方法的不足,并提出了基于信念图的先验信息获取方法;其次针对所获取的先验信息通常具有一定的不确性,对最小描述长度测度进行了改进以融合非确定性先验信息;最后依据问题特性对模拟退火算法进行了适当的修改以更好地优化网络结构。实验表明,提出的结构学习方法能够有效地提高网络结构的学习精度。
In the process of Bayesian network structure learning, how to gather prior information and use it effectively are very important for building an exact network structure. Therefore, according to the three stages of Bayesian network structure learning with prior information： information gathering, information fusion and the optimization of learned network structure. Firstly, the deficiency of currently available methods that is used for obtaining prior information is discussed, and then a new method of gathering prior structure information based on so-called belief map is proposed. Secondly, the minimum description length score is modified so that it can fuse uncertain prior information. Finaly, a simulated annealing method is revised appropriately according to the characteristics of the problem for searching the optimal structure. Experimental results show that the proposed method can improve the precision of structure learning efficiently.
出处
《系统工程与电子技术》
EI
CSCD
北大核心
2012年第12期2585-2591,共7页
Systems Engineering and Electronics
基金
国家自然科学基金(61074107
91024015)资助课题
关键词
贝叶斯网络
结构学习
信念图
非确定性先验信息
模拟退火
最小描述长度
Bayesian network
structure learning
belief map
uncertain prior information
simulated annealing
minimum description length
