摘要： 目前结构学习的算法普遍存在收敛性差、精确度低、易陷入局部最优等问题。为此,提出一种新的网络结构学习算法。通过保留飞蛾-烛火优化算法的整体框架,借鉴遗传算法的杂交、变异等操作,替换原算法的位置更新方法。变异操作时考虑节点间的互信息,对不同节点采取不同的变异动作,保障结构返回的稳定性。实验结果表明,该算法能够较快地学习到评分最优的网络结构,且获得的结构和标准的网络结构最相似。

关键词: 贝叶斯网络, 全局搜索, 飞蛾-烛火优化算法, 遗传算法, 互信息

Abstract: At present,Bayesian structure learning algorithms commonly have the disadvantages of poor convergence,low accuracy and easily trapping in the local optimum.A novel structure learning algorithm is proposed in this paper,which retained the sort framework of Moth-flame Optimization(MFO),defines the crossover operator and variation operator by borrowing the ideas from genetic algorithm to replace the location update strategy of MFO.The mutual information between nodes is considered during mutation period to increase the possibility of returning a similar solution.Simulation results show that the novel algorithm can learn the optimal Bayesian network structure quickly and the returned outcome is more approximate to the standard networks.

Key words: Bayesian network, global search, Moth-flame Optimization(MFO) algorithm, Genetic Algorithm(GA), mutual information

中图分类号: 

• TP301.6