基于改进蝙蝠算法的贝叶斯网络结构学习

The Bayesian Network Structure Learning Based on the Improved Bat Algorithm

针对蝙蝠算法在搜索评分阶段易陷入局部最优且收敛精度低,以及基于蝙蝠算法的贝叶斯网络结构学习不完善等缺点,将模拟退火算法的思想引入到蝙蝠算法中,并对某些蝙蝠个体进行高斯扰动,提出了一种改进蝙蝠算法的贝叶斯网络结构混合学习算法.混合算法首先应用最大最小父子节点集合算法(Max-min parents and children,MMPC)来构建初始无向网络的框架,然后利用改进的蝙蝠算法进行评分搜索并确定边的方向.最后把应用本算法学习的ALARM网,和蚁群算法(MMACO)、蜂群算法(MMABC)进行比较,结果表明本混合算法具有较强的学习能力和更好的收敛速度,并且能够得到与真实网络更匹配的贝叶斯网络.

In this paper an improved bat algorithm for Bayesian network structure hybrid learning algorithm is presented by introducing the simulated annealing algorithm into the bat algorithm and associating Gaussian perturbations to some individual bats to improve the defects of the bat algorithm that is easy to fall into a local optimum in the search- score stage and has low convergence and the imperfection of Bayesian network structure learning based on the bat algorithm. Firstly,the max- min parents and children( MMPC) algorithm is applied to construct the initial framework of the undirected network in the hybrid algorithm,and then the improved bat algorithm is used to the score-searching and the edges' orienting. Finally,the proposed algorithm is applied to learn the ALARM network and it is found that our algorithm has more learning capabilities and better convergent rates and can obtain more real Bayesian network structure than the max- min ant colony optimization( MMACO) algorithm and the max- min artificial bee colony( MMABC) algorithm.