基于MMPC-FPSO贝叶斯网络混合结构学习方法

A Hybrid Structure Learning Method Based on MMPC-FPSO for Bayesian Networks

针对贝叶斯网络结构学习的过程中网络结构规模随节点数增加呈指数增长,导致网络结构搜索空间增大,进而导致网络结构学习算法效率低下的问题,提出一种基于最大最小父子集合约束与萤火虫粒子群搜索算法的贝叶斯网络混合结构学习方法。首先,针对粒子群算法在解决贝叶斯网络结构学习过程中,随机初始化网络结构种群导致算法搜索效率低下,网络结构准确性低的问题提出一种基于改进的最大最小父子集合算法的种群约束方法。其次,针对传统的基于粒子群评分搜索方法速度慢,精度低,易陷入局部最优的问题,提出一种基于萤火虫算子的粒子寻优策略。最后,为了验证所提方法的正确性和优越性,将上述方法用于3种标准网络的结构学习。仿真结果表明:所提算法与传统的基于粒子群的结构学习方法相比,所得的贝叶斯信息准则评分与标准网络评分的差距分别缩小了68.7%、65.5%、34.1%。

Aimed at the problems that in the process of bayesian network structure learning,the network structure size increases in exponential with the number of nodes,in leading to the expansion of the network structure search space,and,in turn,hampering the efficiency of network structure learning algorithms,a bayesian network hybrid structure learning method,MMPC-FPSO,is introduced in combination with maximum-minimum parent-child set constraints(MMPC)and firefly particle swarm optimization(FPSO).Firstly,in view of addressing the issues of low algorithm efficiency and inaccurate network structure due to random initialization of the network structure population in the process of bayesian network structure learning using particle swarm algorithms,a population constraint method is proposed based on the improved MMPC algorithm.Secondly,in view of tackling the problems of slow speed,low accuracy,and susceptibility to local optima in traditional particle swarm-based scoring search methods,a particle optimization strategy based on the firefly algorithm is presented.Finally,in order to validate the correctness and superiority of the proposed method,the three standard networks are applied to the structure learning.The simulation results demonstrate that the gap between the obtained BIC scores and the scores of standard networks is reduced by 68.7%,65.5%,34.1%,respectively by the proposed algorithm,compared to the traditional particle swarm-based structure learning methods.