【摘要】 从数据库中学习贝叶斯网络结构是一个NP难问题.针对此问题,本文提出一种基于遗传算子的粒子群优化算法.首先,利用最大权生成树算法得到初始种群,然后采用遗传算法中的变异和交叉规则优化初始种群,结合贝叶斯网络的结构特点,并设计粒子位置更新策略将学习贝叶斯网络结构的过程转化为粒子寻找最优位置的过程.在学习过程中利用贝叶斯信息标准值作为粒子的适应度函数值,在保证求解质量的同时,加速了搜索过程;为了避免过早收敛,对局部较优的部分粒子和全局极值采用混沌优化策略.最后,利用标准的Alarm和Asia网络模型,验证了本文算法的有效性及可行性.与其他算法相比,新算法在保持较快收敛速度的前提下,具有更好的求解质量.更多还原

【Abstract】 The learning structure of Bayesian networks from a data base is a NP-hard problem. To overcome this problem,This paper proposes a method which integrating particle swarm optimization( PSO) and genetic operators. Firstly,this method used the Maximum Weight Spanning Tree( MWST) to generate the candidate networks. Then,we use the mutation and crossover operators of genetic algorithm( GA) to optimize the initial populations. Considering the characteristics of the structure,we designed an update strategy for particle location which transformed the process of learning the structure of Bayesian Networks into finding the optimal location of particles replaces. In the learning process,this paper used the standard values of Bayesian information criterion instead of the fitness function values of particles,so that it improved the search process of particles and keeped better solution quality. At the same time,we used chaos optimization to avoid getting the local optical solution. Finally,Simulation results based on Alarm and Aisa networks showgreat efficacy and feasibility of our proposed algorithm,and compared with other algorithms,the solution quality precedes the other algorithms while the convergence speed is faster.更多还原