人工智能中若干模糊问题的研究-智能与分布计算实验室

人工智能中若干模糊问题的研究

姓名	符海东
论文答辩日期	2002.05.10
论文提交日期	2005.03.02
论文级别	博士
中文题名	人工智能中若干模糊问题的研究
英文题名	Research on a Few Approximate Problems in the Artificial Intelligence
导师1	卢正鼎
导师2
中文关键词	人工智能;模糊问题;模糊推理;模糊决策;数据挖掘
英文关键词	Artificial intelligence;Approximate problems;Fuzzy reasoning;Fuzzy decision making;Data-mining
中文文摘	模糊理论在人工智能中有着重要的运用。人工智能要研究的一个重要的课题就是对不精确、不完整、不确定的信息加以有效处理。由于人们对人类思维中处理模糊性问题的规律还未能有一个十分明确的认识，因此，对于这类问题，人们至今并没有一个统一的模式。虽然现在已有许多不同的方法处理模糊问题，它们各有特点，但都存在着局限性和不足,不能适应广泛的要求。研究目标之一就是提出更好更有效的算法或方法，以使得算法或方法能够适应更广泛的条件，得到更为满意的结果。针对人工智能中模糊问题的三个重要的领域进行了深入的研究和探索，做了许多工作和取得了下面的一些成果： (1)在模糊推理方面，提出了多重蕴涵稀疏规则的模糊近似插值推理方法。首先分析了经典的合成规则推理方法，并给出了这些方法的使用条件。分析了Marsala和Bouchon-Meunier多重蕴涵稀疏规则条件下的插值推理，指出了它的局限性，并使用实验数据证明了在许多场合这种算法并不适用，特别是不能保凸，而保凸是模糊控制的基本要求。之后，提出了一个新的面积相似的线性插值推理方法，克服了Marsala和Bouchon-Meunier线性插值推理方法的不足。这种新算法根据相似性原理，从输入事实和推理规则的几何图形和重心位置入手，使得几何图形的面积成比例以及它们的重心的水平距离成比例，适应任何形式的稀疏规则下的插值推理，结果永远是保凸且正规的，具有直观的物理意义解释，为智能控制系统提供了有力的工具。 (2) 在模糊决策方面，对基于Vague集理论的模糊决策方法进行了深入研究，提出了三种决策方法：1.提出了理想方案的概念和侯选方案对满足多评价指标的不确定性的概念，利用侯选方案与理想方案的Vague集距离来求出最佳方案；2.对约束条件，从它出现的可能性和不出现的可能性以及未知是否出现的可能性三个方面去综合处理，使得决策更加准确和有效；3. 创造性地将基于Vague集合的推理运用于模糊决策，强调从整体上去考虑对各个评价指标的满足程度。将评价指标和侯选方案之间的关系用一组基于Vague集的推理规则来表示，将决策者的要求用一组Vague集来表示，经过模糊推理等过程最后得到决策结果。 (3)在知识学习和数据挖掘方面，提出了基于Rough集差异矩阵的数据挖掘算法,这种方法在比较信息系统中记录的属性值的差异的基础上，获得各种确定性知识和可能性知识，并给出置信度。这种算法分为3个学习阶段，有初始学习阶段、增量学习阶段、规则约简阶段；对于不完整的信息系统数据挖掘困难的现状，也提出了新的方法解决这个问题，提出了利用条件属性不完整等价划分关于决策属性等价划分的上、下不完整近似集计算的算法，能够从数据不完整的信息系统中获得知识，并且同时能预计数据不完整属性的值。
英文文摘	Fuzzy theory is being importantly applied in the field of artificial intelligence. One of the important projects of artificial intelligence is to research the capability how to process inexact unintegrity uncertain information efficiently. Because the rules of fuzzy questions in mind can not be recognized very clearly, there are not unique model for these questions in human being’s mind. Even though now there are many different methods for fuzzy questions, which have own properties, but there exists limit and fault in it so can not meet widely requirement. The goal is to propose better arithmetic to get more satisfying result under more widely conditions. This paper researched deeply three important fields about fuzzy questions: (1) For approximate reasoning, this paper proposes an interpolative reasoning approach of multiple implications approximate reasoning. At first, introduce the CRI reasoning methods, and constraint conditions of them. Approximate reasoning is proved an interpolator. This paper analyses Marsala and Bouchon-Meunier interpolative reasoning and its restrict under multiple implications rules condition, and then propose a new linear interpolative reasoning method, which overcomes the fault of MB linear interpolative reasoning and keep result convex and shaping. (2) For Fuzzy decision making, this paper researched Fuzzy decision making method based on vague set, and proposed three decision making method: 1. The concept of ideal alternative and the concept of uncertainty for alternative to satisfy multicriteria are proposed and the best choice can be gotten by calculating the distance between alternative and ideal alternative. 2. For constraint condition, three aspects that the probabilities of appearance, disappearance of situations and the probability that we do not know whether it will appear are processed together so that decision making is more exact and effective by using it. 3. The relation between multicriteria and alternative is represented to a group of reasoning rules, requirements of decision maker is represented to a vague set. Through approximate reasoning and some other procedures we can get best choice from the alternatives. (3) For knowledge learning and data mining, this paper proposed a data mining method based on discernibility matrix which can get exact knowledge and possible knowledge and give a confidence factor based on comparing the difference of properties recorded in information system. The method included three learning phases ? initial learning, increment learning and rule simplifying. And for the actuality of that it is difficult to mine data from incomplete information system, this paper also proposed a new method which can get knowledge from information system of incomplete data and can forecast property value of incomplete data using upper and lower incomplete approximate set calculation about condition property incomplete equivalence partitioning.