基于MMHC算法的贝叶斯网络结构学习算法研究

贝叶斯网络是人工智能领域研究不确定环境下知识表示和因果推理的有效工具之一,迄今为止已经提出了许多贝叶斯网络结构学习算法.MMHC算法是一种较新的贝叶斯网络结构学习算法,该算法的评分搜索阶段应用了贪婪搜索算法,但该算法容易陷入局部最优而无法得到全局最优网络,针对该缺点,在MMHC算法的评分搜索阶段应用模拟退火、随机重启爬山搜索、禁忌搜索3种搜索策略取代贪婪搜索,详尽的实验结果表明在MMHC算法中这3种搜索算法的效果普遍优于贪婪搜索,其中模拟退火搜索学习效果最好,MMHC算法的评分搜索阶段可以用模拟退火搜索替代贪婪搜索达到提升算法的学习效果.

STUDY ON LEARNING BAYESIAN NETWORK STRUCTURE BASED ON MMHC ALGORITHM

Bayesian network is an important knowledge representation and reasoning tool under uncertain conditions,there are state-of-the-art Bayesian network structure learning algorithm.Tsamardinos presented a new algorithm for Bayesian network structure learning,called max-min hill-climbing(MMHC).Greedy search algorithm used in the search-score stage of this algorithm,but it is easy to get into the local optimum.In order to overcome this drawback,an improved algorithm was proposed.The algorithm applied simulated annealing,random repeated hill-climbing search,tabu search instead of greedy search in the search-score stage.Detailed results of a complete experiment show that these three search algorithm is generally superior to greed search,simulated annealing is the best.MMHC algorithm applied simulated annealing in order to improve the performance of algorithm.