Journal of Systems Engineering and Electronics ›› 2009, Vol. 31 ›› Issue (4): 938-946.

• 软件、算法与仿真 • 上一篇    下一篇

DBN结构学习度量分解性能分析

肖秦琨1,2, 高晓光2, 高嵩1, 王海芸2   

  1. 1. 西安工业大学电子信息工程学院, 陕西, 西安, 710032;
    2. 西北工业大学电子信息工程学院, 陕西, 西安, 710072
  • 收稿日期:2008-01-01 修回日期:2008-07-06 出版日期:2009-04-20 发布日期:2010-01-03
  • 作者简介:肖秦琨(1974- ),男,副教授,博士,主要研究方向为动态贝叶斯网络.E-mail:xiaoqinkun@tsinghua.edu.cn
  • 基金资助:
    国家自然科学基金(90205019);中国博士后科学基金(2008043410);陕西省教育厅专项科研基金(07JK277);“十一五”兵器预研支撑基金(62301110115,62301110408)资助课题

Study on metric decomposition for DBN structure learning

XIAO Qin-kun1,2, GAO Xiao-guang2, GAO Song1, WANG Hai-yun2   

  1. 1. School of Electronic Information Engineering, Xi’an Technological Univ., Xi’an 710032, China;
    2. School of Electronic Information Engineering, Northwestern Polytechnical Univ., Xi’an 710072, China
  • Received:2008-01-01 Revised:2008-07-06 Online:2009-04-20 Published:2010-01-03

摘要: 针对动态贝叶斯网络(DBN)结构学习中涉及的度量分解问题,提出了DBN度量分解后的相关性能。首先,细化了DBN的贝叶斯信息度量(BIC)及贝叶斯-狄里克莱(BD)度量公式,通过表达式的分析,讨论了分解后的相关性质,进而提出了由分解公式提供给DBN结构学习的相关性能。其次,通过设计的性能分析仿真实验,验证了提出的若干设想,即将BN结构学习算法移植到DBN结构学习的可行性及分解降低算法复杂度等问题,并提出了寻找DBN快速结构学习算法的有效思路。

Abstract: Some correlative properties on dynamic Bayesian networks(DBN) structure metric decomposition for DBN structure learning are proposed.Firstly,DBN’s Bayesian information matric(BIC) and Bayesian-Dirichlet metric(BD) decomposition formula are further divided into two parts.Some characters are discussed based on the decomposition formula,and more useful properties are developed.Secondly,a simulation model is designed to verified properties.The properties include two problems,one is the transplantation problem that many static state Bayesian networks(BN) structure learning algorithm can be used to DBN structure learning,the other is computation complexity problem that DBN structure learning time can be lower through DBN structure decomposition.In the end,a good idea is presented for finding a faster and efficient DBN structure learning algorithm.

中图分类号: