【作者】 王洪春；

【导师】 张勤；

【作者基本信息】 重庆大学 ， 控制理论与控制工程， 2005， 博士

【摘要】 人工智能研究的目的无非是用机器模拟人脑的思维,人类的思维是多样性的,虽然很多思维现象体现为对确定性信息的处理,然而更多的现象却体现了各种各样的不确定性,而且,客观世界中的绝大部分现象都是不确定的。因此,真正的人工智能系统要能很好反映人脑思维的不确定性并能对各种无所不在的不确定性信息进行处理。于是,如何表示和处理知识的不确定性也就成为人工智能研究的重要课题之一,也是人工智能面临的一大难题。动态因果图由张勤教授1994 年提出,它与信度网类似,是概率论与图论结合的一种数学工具,其特点是提供不确定知识的表达和灵活的推理方法:用节点表示事件或变量,有向边表示因果关系,并用连接强度来表示因果关系的强度,支持由原因到结果的正向推理方式和由结果到原因的反向推理方式以及正反向混合推理方式。但因果图与信度网相比又具有一些自己独特的优点,在不确定性知识间的因果关系表达更加方便,尤其在故障诊断领域更有独特优势。因此对因果图的进一步研究不仅具有重要的学术意义,而且具有很好的实用价值和经济价值。论文围绕着因果图的知识表达、学习、推理进行了讨论和研究,主要内容包括: 在扼要介绍了一些比较常见的不确定性知识的表示和推理方法:证据理论、确定性因子、模糊逻辑与模糊推理、粗糙集理论、主观Bayes 方法、信度网的基本知识和面临的困难之后,比较详细地阐述了因果图的基本知识,主要的推理算法以及对一些问题的处理方式方法。针对目前因果图不包括自学习机制、推理的先验知识完全由领域专家提供的问题,提出了采用统计的方法学习因果图参数的方法。包括:在数据完备时用后验分布的数学期望——条件期望估计,数据不完备时,用类似期望最大化(EM)算法,学习离散因果图参数的算法,以及用信息熵学习相关度的方法,而且用实例验证了它们的有效性和可行性;采用含参数的EM 算法(EM(η)),进行在线因果图参数(连接强度)的学习,使学习出的参数能适应环境的变化而适时调整,并阐述了它的优越性和离线因果图参数学习的区别,同时在理论上论证这种方法的正确性;用经典的统计方法:参数估计、非参数估计、半参数化估计方法学习连续因果图参数(基本事件和连接事件的概率密度函数)方法;给出了一个学习因果图结构的途径。从而较好地解决了因果图知识获取的关键问题,对丰富因果图理论和因果图的应用都有着十分重要的意义。针对信度网研究已比较成熟,已有许多现成的算法和实用的推理软件,提出更多还原

【Abstract】 The aim of Artificial Intelligence research is no more than to simulate the thought of human brain with machine, the thought of human is various, although many thought phenomena are behaved the disposal of certainty information, and more phenomena are behaved various uncertainty, many phenomena in reality world are uncertain. Therefore, the really Artificial Intelligence system has to reflect the uncertainty of human brain and deal with uncertain information immanence. And then, how to represent and deal with the uncertainty of knowledge is the basis of Artificial Intelligence research, it is a puzzle of Artificial Intelligence must be faced with. Dynamic Causality Diagram was first proposed by professor Zhang Qin in 1994, it is a mathematics tool combined with probability and graph theory, just like the Belief Network, its characteristic is to provide the method of uncertain knowledge representation and agility reasoning, it adopts nodes to represent random variables in the domain and directional edges between nodes to represent causal relationship between variables, linkage intensity to represent the strength of the link between these variables, it supports the forms of reasoning from cause to effect and from effect to cause and together. Dynamic Causality Diagram has some advantages compared with Belief Network, it is more convenience to represent causal relationship, and the superiority is taken on especially in the Fault Diagnosing field. Therefore the more researches on Causality Diagram have not only the academic significance but also practical and economical value. This dissertation discusses and studies to surround the knowledge representation, learning, reasoning, and the main contents include: To state the basic knowledge, primary algorithm and the way of dealing with some problem of Causality Diagram relative particular, after introducing some familiar uncertain knowledge representation and reasoning and the difficulties of them: Evidential theory, Certainty factor, Fuzzy logic and fuzzy reasoning, Rough set theory, Subjective Bayesian method, Belief network. Owing to the problem of Causality Diagram don’t include self-study mechanism at present and the prior knowledge of reasoning is supplied completely by field experts, a method of learning Causality Diagram’s parameters with statistical approach is presented. It includes the discrete Causality Diagram’s parameters learning algorithms: the mathematical expectation of posterior distribution---conditional expectation estimation when the data is complete, and analogous EM algorithm when the data is incomplete, as well as learning correlative degree method with information entropy, and the experiment results show that the algorithms are effective and feasible; Learning Causality Diagram’s parameters online algorithm adopt with EM algorithm that contains parameter---EM(η), which makes the leaned parameters can adapt to circumstance, its superiority and difference with offline parameters learning are explained, and its correctness is proved in theory; Learning the continuous Causality Diagram’s parameters is presented by classical statistical approaches: parameter estimation, non-parametric estimation, semiparametric estimation; an approach of learning Causality Diagram’s structure is present. It solves the knowledge acquisition key problem of Causality Diagram preferably, it is important for abundance Causality Diagram theory and application of Causality Diagram. An algorithm that how converts Causality Diagram into Belief Network is presented, because the Belief Network has many ready reasoning algorithms and applied software. It includes the linkage intensity of Causality Diagram transfer to CPT of Belief Network, and the structure of Causality Diagram transfer to the structure of Belief Network. The channel of solving problem is developed because the model represented by Causality Diagram can be solved by Belief Network model. Production rule knowledge representation is popular and common, but there are many shortages to represent knowledge and reasoning, according to the relation of production rule and Causality Diagram, discuss the method and course are proposed in this paper to convert production rule set into Causality Diagram, correspond to give a method of knowledge acquisition based on Causality Diagram, and give an example to this conversion. An approximate reasoning algorithm that adopt max and min operator based on Causality Diagram’s characteristic and matrix transform has been presented to improve the deficiency of logic operation complexity and computation complexity, it can raise the computation velocity in Causality Diagram reasoning. Owing to the occurrence probability of event shows fuzzy and random, the fuzzy number is inducted into the causality diagram in this paper, with the fuzzy number replace the probabilities of basic events and linkage events, and it can solve the difficulty of obtaining the precision probability value as well as solve the problem of the fuzzy and random of the occurrence probability of event. This method was applied to the fault diagnosis of stow machine and pressure vessel.更多还原