基于进化计算的贝叶斯网络结构复合学习算法

Hybrid Learning Algorithm for Bayesian Network Structure Based on Evolutionary Computation

提出一种贝叶斯网络结构复合学习算法.该算法将EM算法、蒙特卡罗抽样算法、进化算法结合起来,用EM算法、蒙特卡罗抽样算法将不完整的数据集转换成完备的数据集,再利用进化算法进化网络结构.这种算法能够克服EM算法容易陷入局部最大值的缺陷,对于缺省数据处理是基于后验网络的,网络结构随进化计算不断优化,得到的补充数据可信度比较高,网络学习效率高、运算性能好.

The authors have presented a Hybrid Learning algorithm （HL algorithm） for Bayesian Network structure, which combines EM algorithm, Monte Carlo sampling algorithm and evolutionary algorithm. The HL algorithm uses EM algorithm and Monte Carlo sampling algorithm to convert uncompleted data to completed data and uses evolutionary algorithm to evolve the structure of Bayesian Network, which overcomes the defect of gaining local maximum of the EM algorithm. Since data processing in HL algorithm is based on posterior networks structures and the structures of Bayesian Network are optimized with evolution computing, the HL algorithms are of higher reliability of complementary data, higher learning rate, and better operational performance.