【作者】 施明辉；

【导师】 周昌乐；

【作者基本信息】 厦门大学 ， 基础数学， 2008， 博士

【摘要】 随着科学技术的高速发展,智能信息处理已成为众多学科领域研究的热点。当前中医现代化的进展迫切需要先进的智能信息处理技术的支撑。中医诊断现代化无疑是中医现代化的重要方面。其中,中医智能诊断是中医诊断技术与智能信息处理技术相结合的较好切入点,其必须解决的核心问题和关键技术在于中医智能辨证。早期的研究实践表明,中医智能辨证的关键环节在于知识的处理,包括知识的表示、获取、发现与利用等方面。其中所面临的许多问题与困难也是当前人工智能领域研究的热点与难点。基于软计算思想的一系列新型智能信息处理技术的兴起,为更好地解决这些问题与难点带来了机遇,而其自身也可从解决问题的过程中获得新的启迪,丰富其研究内容与成果。论文研究并分析了软计算方法在中医辨证智能诊断领域中的研究现状、基本方法及面临的困难,对不确定性知识的表示及处理,归纳与模拟人类专家的经验并建立相应的信息模型,从大规模数据中获取或发现知识,以及计算大规模信息系统的属性约简等关键技术的研究作了有益的探索,旨在为中医辨证计算化的研究与实现提出新的思路、方法和技术,也为人工智能领域中相关的难点问题提出新的解决办法。论文主要研究内容如下:第一章首先阐述了论文研究的时代背景及学科交叉特色;然后指出了目前中医智能诊断研究面临的挑战与意义,以及软计算在中医辨证计算化研究中具有的独特作用;最后概述了粗糙集理论及应用的研究进展,着重分析了粗糙集在知识获取与属性约简两方面的研究内容与意义,以及粗糙集在中医智能诊断方面的初步应用与存在的问题。第二章首先介绍了中医辨证的基本概念、辨证原理和辨证方法;然后分析了软计算在中医辨证智能诊断研究中的优势与难点;最后详细阐述了中医辨证智能诊断的软计算方法研究的进展。本章归纳总结了基于模糊集理论的中医辨证诊断方法和基于模糊集理论的中医证型的模糊聚类方法;分析了神经网络在中医辨证智能诊断中的应用研究现状、基本观点、一般方法、存在的问题、解决问题的思路,并介绍了基于神经网络的中医辨证智能诊断研究整体思路的初步设想和所做的相关研究工作;回顾并总结了粗糙集理论在中医辨证智能诊断中的一般步骤;概要介绍了当前多技术融合方面的相关研究工作与研究趋势。针对知识获取这一智能系统开发的瓶颈问题,第三章和第四章分别针对人类专家的两种思维方式—“聚焦”和“层级聚类”—进行了深入探讨,发现已有的模拟这些思维方式的分类规则提取方法的局限性:它们在聚焦机制的排除过程和鉴别过程中都采用覆盖准则,导致其鉴别过程只能适用于在二者之间进行。为此,提出了改进办法。改进思路的基本出发点是:若在鉴别过程中采用精度准则,则可以使鉴别过程在多者之间进行,进一步地,还可以与属性约简方法相结合,消除冗余属性。第三章针对“聚焦”思维方式,提出了分类规则提取算法REFM;第四章针对“层级聚类”思维方式,提出了诊断规则提取算法REHC。针对计算大型信息系统的所有属性约简(包括计算其所有最小属性约简)这一NP-hard问题,第五章首先考察了分辨函数的一系列等价形式;然后提出了约简分辨图的概念,以及深度优先搜索的三项原则:成员独占原则(MEP)、友人劝阻原则(FPP)、陌生人吸纳原则(SEP);进而阐述了基于约简分辨图计算属性约简的完整理论及算法CARRDG,并从理论上严密论证了算法CARRDG的高效性与完全性;最后用六种典型的UCI数据进行实验验证。UCI数据实验表明:对于多数信息系统,算法CARRDG计算所有属性约简的时间小于0.5秒;对于对象数达到20000的信息系统,算法CARRDG的剪枝率可达90%以上,且可在几分钟内计算出所有属性约简。算法CARRDG虽然是针对属性约简的计算问题提出的,但其实质上解决的是将合取范式快速转化为析取范式并进行简化的问题,因而具有广阔的应用空间。故第五章的理论价值不仅在于为计算所有属性约简(包括计算所有最小属性约简),提出了新的观点、思路、理论和方法,而且在于给出了采用基于约简分辨图的启发式搜索,解决逻辑表达式转化与简化中的组合爆炸问题的新思路。第六章首先提出了学习型中医辨证诊疗系统的构想,分析了论文研究成果在此构想中的应用方式及意义;然后总结了论文的主要工作与创新点;最后阐明了目前研究工作中有待完善之处、存在的困难及未来的研究方向与前景。更多还原

【Abstract】 With the rapid development of modern science and technology, intelligent information processing has become hot point in many research fields, thus the corresponding technologies have become the urgent supporting power for the modernization of traditional Chinese medicine (TCM). While the modernization of diagnosis in TCM is one of the important facets of the modernization of TCM, intelligent diagnosis in TCM appears to be the perfect research entry for combing intelligent technologies with diagnostic technologies in TCM, and its core problem and key technology lie in the intelligent syndrome differentiation in TCM (SDTCM)(syndrome differentiation is an unique concept of TCM).Based on the earlier research activities, it has been emerged that the key point of intelligent diagnosis in TCM may be related to knowledge processing including knowledge representation, knowledge acquisition, knowledge discovery and knowledge utilization, etc., and the difficult problems occurred are always the important research topics in the area of artificial intelligence. Fortunately, a series of advanced intelligent technologies based on soft computing have brought great opportunity for solving these problems, and they, in return, also can attain new enlightenment and enrich their research content and harvest.In this dissertation, the state-of-the-art of soft computing for the intelligent SDTCM as well as its fundamental methods and difficulties are studied and analyzed. The other research topics involve in representing uncertain knowledge, designing information model by concluding and simulating experiences of human experts, discovering knowledge from large data, and reducing large data, etc. The purpose of this dissertation is to provide some new ideals, methods and technologies for the realization of the computing in SDTCM, and to propose several new resolutions for the related hard problems in the area of artificial intelligence.The main content is as follows.In the first chapter, the background and the significance of the research are interpreted and the research activities and applications about rough set theory are reviewed. It is also pointed out that the critical point of the intelligent diagnosis in TCM is the realization of the computing of SDTCM, and the main corresponding problems lie in knowledge processing including knowledge acquisition, knowledge discovery and knowledge exploiting. As for the rough set theory, the state-of-the-art is outlined, and the research works based on it are analyzed involving knowledge acquisition, attribution reduct as well as the application in the intelligent SDTCM.In the second chapter, first, the fundamental concepts, principles and methods of SDTCM are introduced. Second, the advantages and difficulties of soft computing in the research on the intelligent SDTCM are analyzed; Finally, the state-of-the-art of the intelligent SDTCM is interpreted in detail: 1) the diagnostic methods and clustering methods of syndromes for SDTCM based on fuzzy set theory are concluded ; 2) the stat-of-the-art, fundamental viewpoints, general methods, existing problems with corresponding resolution ideas of the applications based on neural network for SDTCM are analyzed, and our initial idea and research works for this area are introduced; 3) the general steps of the practical methods based on rough set theory for SDTCM are reviewed and concluded; 4) the research works for the fusion of multi-technology are also introduced in brief.Aiming at the problem about knowledge acquisition, which is a neck problem when developing an intelligent system, in the third and fourth chapters, two kinds of human experts’ thinking ways, the focusing mechanism and the hierarchical clustering mechanism, are respectively examined, and the limitations of the existed approaches to extracting classification rules by modeling these two thinking ways are carefully examined. The focusing mechanism is clearly represented by three ordered processes: exclusion process, discrimination process and combination process. One of the main limitations of the approaches existed is that they can only do discrimination between two classes, since they exploit coverage criteria in the discrimination process. The point departure for the improvement is: if exploiting accuracy criteria in the discrimination process, then the discrimination process may be done among many classes. Furthermore, they can be combined with the methods for computing attribute reducts. For the focusing mechanism, the algorithms REFM is proposed in the third chapter. For the hierarchical clustering mechanism, the algorithm REHC is proposed in the fourth chapter.To challenge the NP-hard problem computing the total attribute reducts (including the total minimal attribute reducts) in large information systems, in the fifnth chapter, a series of equivalent form of discernibility function are examined; then an important and novel concept, i.e. reduct disernibility graph (RDG), is proposed; furthermore, the complete theory for computing attribute reducts based on RDG as well as the corresponding algorithm CARRDG are proposed and interpreted; the effectiveness and completeness of the algorithm CARRDG are also proved; at last the results of the experiment on six typical UCI data sets show: the algorithm CARRDG can compute total attribute reducts within 0.5 seconds for the general information systems, and within several minutes and more than 90% trim rate for the large information systems even with 20000 objects. It should be noted that although the algorithm CARRDG is proposed for the special problem, in fact it solves, however, the problem of the transformation and simplification of logic expressions from conjunctive normal form to disjunctive normal form, so it has wide variety of application fields, and it maight also open a new window for solving the combination exposition problems in the real-world problems.In the sixth chapter, the tentative idea for the future structure of the SDTCM system with learning mechanism is proposed, and the application means and the significance of our research fruits in the structure are interpreted. At last, the main research work and the existing problems of this dissertation are concluded and the future of the research is also prospected.更多还原