贝叶斯网络结构学习与应用研究

发布时间：2018-05-11 13:15

本文选题：贝叶斯网络 + 结构学习　；参考：《中国科学技术大学》2017年博士论文

【摘要】：进入到21世纪以来,人们对于人工智能领域的研究取得了极大的进步。作为表示不确定性知识的一种重要方法,贝叶斯网络已经成为了人工智能研究中的热点和重要成果之一,建立有效的贝叶斯网络结构也是学习和应用贝叶斯网络的基础和核心。在贝叶斯网络结构方面,以传统的专家知识构建结构为代表的方法已经逐渐被通过数据学习结构的方法取代。然而,由于可能的网络结构数多,会导致搜索空间巨大。通过训练数据学习贝叶斯网络的方法通常存在着精度低或复杂度高等问题,使得所学习的结构与实际相差较大,从而对贝叶斯网络的推理及实际应用带来很大的影响,因此研究合适的贝叶斯网络结构学习方法具有重要的理论和实际应用的意义。本文针对贝叶斯网络的结构学习及应用问题进行研究。主要围绕根据结构的复杂性,提出相应的结构学习方法,进一步地将贝叶斯网络知识应用于多标签分类中。本文的主要工作和创新性可以概括如下。第一,针对搜索评分的结构学习方法存在着搜索效率低以及容易陷入到局部最优结构问题,提出一种基于离散人工蜂群的结构学习方法。首先,将贝叶斯网络的有向无环图结构抽象成人工蜂群中的食物源,这样将结构学习问题转化成人工蜂群算法中寻找最优食物源问题;然后采用人工蜂群的搜索框架,结合差分进化算法中的变异、交叉算子,从可能的搜索空间中寻找评分最高的食物源;同时利用深度优先搜索的方法,提出一种有环图修正算法;最后通过建立对应的马尔科夫链对算法的搜索过程进行描述,理论上分析算法能够依概率1收敛到最优结构。实验仿真表明所提的基于离散人工蜂群的结构学习算法学习到的结构评分更优,且收敛性更好。第二,针对结点数庞大会导致搜索评分的结构学习方法复杂度高的问题,提出一种混合约束与搜索评分的结构学习方法。首先,利用条件独立性测试从数据中学习出结点所对应的无向独立图结构;然后利用递归的思想对无向独立图进行递归分解,从而构建出多个子图结构;对于每个子无向图结构,利用搜索评分的方法来确定出结构中的边及其对应的方向;最后,提出子结构合成规则将这些子结构重新组合成对应的有向无环图结构。实验仿真表明混合约束和搜索评分的方法能够更有效地处理结点数众多的结构学习问题。第三,考虑到多标签分类中标签之间存在着相关性的问题,提出基于贝叶斯网络的多标签分类方法。首先利用训练数据学习标签贝叶斯网络模型,其中每个结点代表着一类标签,结点间的边和条件概率分布描述标签之间的相关程度;然后利用最大似然估计的参数学习方法学习结构中的参数;对于每个测试集,通过计算其与训练样本的欧式距离来选择与其相邻的样本集,并将相邻样本集中的标签类别信息作为证据输入到标签贝叶斯网络中;最后利用联合树推理来预测测试集最可能属于哪些类标签中。
[Abstract]:Since twenty-first Century, people have made great progress in the field of artificial intelligence. As an important way to express uncertain knowledge, Bayesian networks have become one of the hotspots and important achievements in the research of artificial intelligence. The establishment of an effective Bayesian network structure is also the learning and application of Bayesian networks. The foundation and core. In the Bias network structure, the method represented by traditional expert knowledge construction has gradually been replaced by the method of data learning structure. However, the number of possible network structures can lead to a huge search space. The method of learning the Bias network through training data usually has a low precision. The problem of high degree of complexity makes the difference between the structure and the reality, which has a great influence on the reasoning and practical application of Bayesian network. Therefore, it is of great significance to study the proper Bayesian network structure learning method. This paper aims at the structure learning and application of Bayesian network. The main work and innovation of this paper can be summarized as follows. The main work and innovation of this paper can be summarized as follows. First, the structure learning method for the search score is low in search efficiency and easy to fall into the part of Bias. A structure learning method based on the discrete artificial bee colony is proposed. First, the Bias network's directed acyclic graph structure is used to abstract the food source of the adult worker swarm. In this way, the structure learning problem is transformed into the adult worker swarm algorithm to find the optimal food source problem. The mutation and crossover operator in the evolutionary algorithm search for the highest grade food source from the possible search space; at the same time, we use the method of depth first search to propose a loop graph correction algorithm. Finally, the search process of the algorithm is described by establishing the corresponding Markov chain, and the theoretical analysis algorithm can converge to the probability 1. The experimental simulation shows that the structure learning algorithm based on the discrete artificial bee colony learning algorithm has better structure and better convergence. Second, a structural learning method with mixed and search scores is proposed for the problem of high complexity of the structure learning method which can lead to the search score. Using the conditional independence test, the undirected independent graph structure corresponding to the node is learned from the data. Then the recursive idea is used to decompose the undirected independent graph, and then the multiple subgraph structure is constructed. For each subgraph structure, the edge and its corresponding direction in the structure are determined by the method of search score. Finally, The substructure synthesis rules are proposed to recombine these substructures into corresponding directed acyclic graphs. Experimental simulation shows that the method of mixed constraint and search score can deal with many structural learning problems more effectively. Third, considering the problem of correlation among tags in multi label classification, it is proposed based on bainite leaf. The multi label classification method of the network. First, we use training data to learn the label Bayesian network model, in which each node represents a class of tags, the edge of the node and the conditional probability distribution describe the correlation between the tags; then the parameter learning method of the maximum likelihood estimation is used to learn the parameters of the structure; for each test set, each test set is used. By calculating the Euclidean distance from the training sample, we choose the sample set adjacent to it, and input the label category information of the adjacent samples into the label Bayesian network. Finally, the joint tree reasoning is used to predict which class labels are most likely to belong to the test set.

【学位授予单位】：中国科学技术大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP18

【相似文献】