【作者】 范敏；

【导师】 黄席樾；

【作者基本信息】 重庆大学 ， 控制理论与控制工程， 2008， 博士

【摘要】 从网络环境的海量数据中获取新的科学知识、新的规律和决策支持信息是信息时代研究的热点。有效地分析和挖掘广域分布海量数据的知识发现方法是研究工作的核心。贝叶斯网络(BN)结合图论和统计学方面的知识,提供了一种自然的表示因果信息的方法,用于表达随机变量之间复杂的概率不确定性,发现数据间的潜在关系,它在海量数据的知识发现中具有重要的研究意义和广阔的应用前景。本文以贝叶斯网络为基础理论,以重庆市科技攻关重点项目(CSTC,2006AA7024)“三峡库区水环境安全预警平台与科学决策关键技术研究”为应用背景,针对网络环境下的数据智能学习问题,围绕科学决策理论进行基于贝叶斯网络的学习与决策方法研究,从贝叶斯网络学习、贝叶斯网络分类器学习、概率关系模型学习和基于BN的多agent群决策支持系统四方面开展以下研究工作:①分析了贝叶斯网络的建模流程,着重讨论从数据中学习贝叶斯网络结构。通过分析两种经典的结构学习方法(K2和MCMC算法)的基本思想,将两种算法的优点和模型平均的思路结合起来,提出一种改进的贝叶斯网络结构学习算法。仿真实验表明改进算法可以在无先验知识的情况下以较快的收敛速度获得较正确的稳定结构,且不易受初始结构影响,能正确反映节点间的依存关系。该学习算法具有较好的鲁棒性,能为科学决策提供有价值的决策依据,并可以推广于各应用领域的海量数据中进行变量间的相关性分析与建模研究。②在对常见贝叶斯网络分类器分析的基础上,研究将属性变量间聚集的层次关系引入分类器中,从数据中学习属性变量间的层次关系,提出构造层次朴素贝叶斯分类器(SAHNB)的学习算法。算法思路是首先借助节点间的条件互信息值来确定可能聚集节点的范围,然后再通过模拟退火算法来搜索评分较高的模型。通过实验将SAHNB分别同NB、TAN和GBN-SA进行比较,结果表明在多数数据集上SAHNB具有较好的分类效果,并且模型引入的潜在节点可以实现对某些属性节点状态的聚集,能够进行语义解释,从而为决策者提供相应的分类规则。SAHNB分类器模型用于水质预警系统,获得较好的实际应用效果。③针对实际应用问题,结合概率关系模型(PRM)构建的特点,提出建设基于PRM的水体富营养化风险预测模型思路,在这一领域研究中是一种新的尝试。学习基于PRM的水体富营养化风险预测模型,首先根据先验知识构造每个参考变量的候选父节点集,并且利用关系数据库的标准化SQL语言对候选父节点集进行数据预处理,主要包括多集操作、聚集操作、离散化处理等,以减轻启发式搜索评分算法的计算复杂度,然后采用类K2算法的搜索评分方法进行模型结构学习,并且通过构建分类指示器来细分不同蓄水期的水体富营养化风险预测模型。应用分析显示该方法在断面水质富营养化风险分析上可以取得初步成效。④研究多agent技术与贝叶斯网络结合建模的方法。设计了基于贝叶斯网络和效用理论的个体agent结构,研究了反馈控制的双向学习机制;设计了多agent群决策支持系统的五层组织结构和面向决策任务的agents组织过程建摸流程;定义了多agent的协商模型,提出了基于贝叶斯信念模型学习和推理的冲突协商、效用偏好调整和启用外界推荐的协商交互流程。本文所构建的贝叶斯网络学习与决策方法能有效的进行数据挖掘与知识发现,实现科学决策。并通过实际应用验证了方法的可行性。更多还原

【Abstract】 The retrieval of new knowledge, rule & decision support information from large scale data in network environment is a hot research point during information time. It focuses on studying the knowledge discovering methods of efficiently analyzing and mining large scale data. Bayesian Network (BN), combined with the knowledge of graph theory and statistics, prompts a nature method to express casual relation. It can be used to express complex uncertainty of random variables and discovery latent relation among data. Therefore, Bayesian Network has an important research value and wide application foreground in KDD domain.This dissertation has study learning & decision methods based on Bayesian Network basic theory, which takes science and technology plan program of ChongQing (CSTC 2006AA7024)”Research on Water Environment Security Early Warning Platform and Key Technology of Science Decision in Three Gorge Region”as application background. Aiming at the problem of data intelligent learning, focusing on Bayesian network learning, Bayesian Network classifiers learning, Probability Relation Model learning and multi-agent group decision support system based BN, the main contributions of this dissertation are as follow:①It analyzes the modeling flow of Bayesian Network and focuses on learning Bayesian Network structure from data. According to analyzing classical Structure Learning methods (K2 and MCMC algorithms), an improved Bayesian Networks Structure Learning algorithm is proposed which combined with the merits of above two algorithms and the idea of model averaging. Experiment results show that the proposed algorithm can cover shortages of K2 and MCMC algorithms and can quickly achieve a comparative correct and steady model structure without priori knowledge. The algorithm has better robust which can be application in variables correlation analysis and modeling of large scale data for science decision.②A learning algorithm for modeling hiberarchy Na?ve Bayesian classifier SAHNB is proposed which learns hiberarchy relations between variables by aggregating variables. It firstly uses Conditional Mutual Information between variables to derermine the scale of latent aggregate nodes, then usee Simulation Anneal algorithm to search the classifier with higher score. Experiment results show that SAHNB has better classification effct comparing with NB, TAN &GBN-SA. Moreover, the nodes adding in model can aggregate the states of attribute nodes and explain semantic in order to supply correlation classification rules. SAHNB classifier has been applied in water quality security early warning system and received better results.③According to the fact problem, it proposes a construct method of risk forecasting model for water eutrophication combined with the merits of Probability Relation Model. This is a new research taste in the domain. The construct flow is as follow: firstly it sets candidate parent nodes for each correlation variable by expert knowledge; then uses SQL language to process candidate parent node sets, including in multi-set operation, aggregate operation and discretization, in order to relax computation complex of heuristic search-score algorithms; finally, it learns the model structure by the similar K2 algorithm and divide the model by classified indicators. This model has received the first step result on application research.④It has research on modeling methods combining MAS with Byaesian Network. Based on Bayesian Network and Utility Theory, it designs the single agent structure and studies a bidirectional learning mechanical with feedback control. Then it designs a hiberarchy structure of multi-agent group decision support system and a dynamics organization method faced on decision tasks. Finally, it defines a multi-agent negotiation model and proposes a negotiation algorithm based on Bayesian belief model for conflict, emphasis and persuading.In conclusion, above works can provide efficacious data mining and knowledge discovering methods for science decision. Moreover, it’s tested that these methods are feasible in applitions.更多还原