摘要
为改善当前贝叶斯网络结构学习算法易陷入局部最优、过早收敛和寻优效率低的问题,进行了混合简化粒子群算法优化贝叶斯网络结构学习的研究。该算法利用最大支撑树约束搜索空间,并提出V-结构与条件相对平均熵相结合的初始定向策略,然后利用爬山策略建立初始粒子群,再利用改进的粒子群优化算法和遗传算法对初始种群迭代优化,在迭代过程中提出条件交叉和变异策略避免粒子的随机发散更新,并结合副粒子增缓策略更新未优化粒子,避免算法陷入局部最优。该算法与其他算法在4种标准网络下进行了仿真实验。实验结果表明,所提算法在ASIA、CAR、CHILD、ALARM网络中相比于MMHC、GS、BNC-PSO、PC-PSO算法BIC评分分别平均高5.775%、5.8%、0.475%、2.75%;汉明距离HD更低,正确率ACC更高。
In order to improve the problems that the current Bayesian network structure learning algorithm tends to fall into local optimization,premature convergence and low optimization efficiency,a hybrid simplified particle swarm algorithm is proposed to optimize Bayesian network structure learning(BNs-HsPSO).The algorithm uses the maximum support tree to constrain the search space,and proposes an initial orientation strategy combining V-structure and conditional relative average entropy(CRAE),and then uses the mountain-climbing strategy to establish the initial particle swarm,uses the improved particle swarm optimization algorithm and genetic algorithm to iteratively optimize the initial population,proposes a conditional crossing and mutation strategy in the iterative process to avoid random divergence update of particles,and updates the unoptimized particles in combination with the sub-particle slowing strategy to avoid the algorithm falling into local optimum.The algorithm is compared with other algorithms in simulation experiments under four standard networks.The experimental results show that the proposed algorithm has an average higher BIC score of 5.775%,5.8%,0.475%,and 2.75%compared to MMHC,GS,BNC-PSO,and PC-PSO algorithms in ASIA,CAR,CHILD,and ARM networks,respectively;the Hamming distance HD is lower and the accuracy ACC is higher.
作者
刘浩然
李晟
崔少鹏
王念太
蔡炎滨
时倩蕊
张力悦
LIU Haoran;LI Sheng;CUI Shaopeng;WANG Niantai;CAI Yanbin;SHI Qianrui;ZHANG Liyue(School of Information Science and Engineering,Yanshan University,Qinhuangdao,Hebei 066004,China;The Key Laboratory for Special Fiber and Fiber Sensor of Hebei Province,Qinhuangdao,Hebei 066004,China)
出处
《计量学报》
CSCD
北大核心
2024年第2期269-278,共10页
Acta Metrologica Sinica
基金
国家重点研发计划(2019YFB1707301)。
关键词
智能算法
贝叶斯网络
粒子群优化
遗传算法
自定义交叉和变异概率
副粒子增缓策略
intelligent algorithm
Bayesian network
particle swarm optimization
genetic algorithm
custom crossover and variance probabilitie
paraparticle accretion and mitigation strategie
