节点文献
智能融合的定性映射模型及其属性计算网络实现技术的研究
【作者】 许广林；
【导师】 冯嘉礼；
【作者基本信息】 上海海事大学 ， 电力电子与电力传动， 2008， 博士
【摘要】 虽然不同的智能方法具有不同的特点,从而能够解决各自特定问题。但是,无论是专家系统、人工神经网路、模糊集方法、支持向量机等,都存在着各自的局限。因此怎样把不同的智能方法与技术整合起来取长补短就成为人工智能一个重要的研究领域。当人工智能领域的学者与工程技术人员尝试把两种以及两种以上的智能方法和技术有机地融合为一体,就形成所谓智能融合(IntelligenceFusion)。本论文试图通过回答以下五个问题来讨论定性映射与智能融合的关系:第一,各种分属不同学派的人工智能方法能否融合;第二,它们能够在什么样的框架下可以融合;第三,它们的融合机制是什么;第四,怎么实现智能融合模型。第五,智能融合模型有那些方面的应用。对第一个问题,我们的回答是肯定的,其理由如下:从认识论角度讲,所谓各种人工智能方法不能融合的看法,与大脑本身是一个统一而和谐的事实是相矛盾的,也就是说,尽管人脑神经系统是一个开放的复杂巨系统,但各神经子系统的运行却是既相互独立,又相互协同的,这种情况表明,人脑中枢指挥系统存在着一个统一而和谐的调度机制,否则,各子系统之间既相对独立又相互协同的运行是不可能实现的。然而,由于人脑神经中枢的这种统一而协调的调度和运行机制没有被揭示出来,使人工智能专家不仅不能利用这种机制去指导自身的研究,而且,即使已提出的各种人工智能方法分别模拟的就是人脑各不同子系统的功能,但由于缺乏一个统一而协调的机制,使人工智能没能成为一门成熟的科学,就象一盘零散的钻石和珍珠,因缺乏一条金链而无法串成一条项链。由此可见,能否将人脑统一调度各神经子系统的机制揭示出来,已不仅仅是脑科学研究的一个基础理论问题,而且,也是能否将各种人工智能方法融合起来的一个核心技术问题。反过来,假如能为各种人工智能方法的有机融合找到一个合理而有效的机制,及其哲学原理、数学理论、方法、模型和计算机实现技术,那么,我们就应该回过头来问一问,各种人工智能方法融合机制的原理、理论、方法、模型和技术,对人脑思维和智能运行机制的揭示和发现,会有哪些启发意义和促进作用?假若人脑思维与智能的统一而和谐的运行机制,与人工智能方法的融合机制具有某些相似或相同之处的话,那么,人工智能融合机制的探索和揭示,就可为人脑思维与智能之秘的探索和揭示,提供一条新的研究途径、方法和手段。为回答第二个问题,我们以医疗和故障诊断操作的“If…,then…”型符号表示为例,不仅找到了导致符号方法与连接方法产生分裂的根源,而且,提出了一种能将它们融合在一起的数学模型——定性映射模型。为回答第三个问题,我们讨论了定性映射与专家系统、人工神经元网络、模糊集合、Bayesian计算、支持向量机和遗传算法等,多种分属不同学派的人工智能方法之间的关系,并指出,定性映射的定性基准是联系专家系统和人工神经网络的一座桥梁,通过它及其平移、伸缩和叠加变换,不仅使定性映射与人工神经元及其网络可以相互转化和定义,而且,还可得到一种基于叠加变换的定性基准边界模糊化机制,和一种基于Bayesian计算的模糊隶属度确定方法。因定性映射可看作是事物属性量——质特征之间转化过程的一个数学表达,为反映同质不同量特征在转化程度上的差异,我们提出了(量——质特征)转化程度函数的概念,并证明了转化程度函数的基准变换,恰好诱导出一个径向基函数人工神经元。此外,定性基准的细分(或剖分)变换,还可诱导出一个子定性映射簇,及其张成的Hilbert空间和一个支持向量机,而定性基准剖分的极限则可诱导出一个遗传算法。为回答第四个问题,我们将定性映射的输入——输出关系,等价地用一个逻辑计算(或电路)单元加以模拟实现,并用串联、并联、加权(偏置)和反馈等方式,将若干个定性映射(逻辑计算或电路)单元整合为一个可重组、可学习、可调整的属性计算网络,由第二和第三个问题的答案可知,分属不同学派的多种人工智能方法,都可用属性计算网络分别加以模拟和实现,也就是说,属性计算网络是一个能融合多种智能方法的智能计算平台。为了回答第五个问题,在属性计算网络智能平台上,我们以手写汉字、手写英文字符、手绘图形、交通标志和船用主机故障诊断等为例,通过软件编程对属性智能计算网络平台进行训练和调试,不仅实现了特征抽取、模式识别和故障诊断等,通常利用人工神经网络实现的功能,而且,还以生产竞争力和高考招生为例,通过人机对话学习方法,利用属性智能计算网络平台,对决策者心理评判标准随环境、条件、经验和心理偏好等变化而变的情况进行了动态的学习和跟踪,并以此为依据,对基于决策者心理评判基准非线性变化的综合评估和全局决策进行了动态模拟。
【Abstract】 Although different intelligent methods are in possession of distinctive characteristics to solve their individual problems, since all the expert system, neural network, fuzzy set methodology and support vector machine have their respective limits, how to integrate different intelligence methods with technology for mutual supplementation has become an important research field in artificial intelligence. When the scholars engaged in this field and engineering technicians attempt to combine two or more intelligence methods with technology, the intelligence fusion takes shape.The authors of the thesis discuss the relation between qualitative mapping and intelligence fusion by answering the following five questions: 1. Can the artificial intelligence methods of different academic schools be fused? 2. Under what kind of framework can they be fused? 3. What is the fusion system? 4. How can the intelligence fusion model be realized? 5. What are the application fields of the intelligence fusion model?For Question 1, we have affirmative answer for the following reasons:With respect to the epistemology, the opinions that various artificial intelligence methods cannot be fused are in conflict with the fact that human brain is united and harmonious. In other words, although the neural system of human brain is open and extremely complicated, the center nerve system has united and harmonious control mechanism due to mutual independent but coordinative operation of neural subsystems or otherwise such operation would not be realized.However, as such united and harmonious control and operation mechanism in nerve center of human brain have not been exposed; artificial intelligence experts fail to utilize the mechanism to direct research. Moreover, even if the artificial intelligence methods do simulate the functions of human brains subsystems, people cannot use this united and harmonious mechanism to integrate all the intelligence methods into one framework. Under such circumstance, whether we can figure out the mechanism of exclusive control of each neural subsystem by human brain is no longer the problem of basic theory in brain science but a core technology problem relating to fusion of artificial intelligence methods.On the contrary, if we do find the effective fusion mechanism for all the artificial intelligence methods, we should ponder whether it is the mechanism of united and harmonious operation of human brain.For Question 2, by taking the "If... then..." symbols used in medical and failure diagnosis, we not only find the source that tears apart the symbol method and connection method, but also suggest the mathematics model for their integration, namely the qualitative mapping model.For Question 3, we discuss qualitative mapping and expert system, artificial neural network, fuzzy set, Bayesian calculation, support vector machine and genetic algorithm as well as the relations of artificial intelligence methods of different academic schools. We point out that the qualitative benchmark of qualitative mapping bridges the expert system and artificial neural network, which, together with its movement, expansion and superposition, not only makes qualitative mapping and artificial neural cell and network mutually interchangeable and definable, but also generates qualitative benchmark boundary fuzzy mechanism based upon superposition and a fuzzy membership definition based upon Bayesian calculation. Furthermore, subdivision (or dissection) of qualitative benchmark may induce a qualitative mapping sub variety and its Hilbert space and a support vector machine while the dissection limit of qualitative benchmark may induce genetic algorithm.For Question 4, we equalize the input/output relation of qualitative mapping to a logic calculation unit and integrated several qualitative mapping (logic calculation) units into a Attribute Computing Network that can be restructured, studied and adjusted by methods of series, parallel, weight (offset) and feedback. Judging from the answers to the second and third questions, the artificial intelligence methods affiliated to different academic schools can be simulated and realized by Attribute Computing Network. In other words, the Attribute Computing Network is an intelligence calculation platform combining various intelligence methods.For Question 5, on this Attribute Computing Network intelligence platform, we take examples of handwriting Chinese characters, handwriting English characters, handmade pictures, traffic signals and failure diagnosis of ship’s main engine and realize training and adjustment of property intelligence calculation network platform by software programming so as to realize the functions of artificial neural network such as feature extraction, mode identification and failure diagnosis. What’s more, by taking the examples of production competitiveness and college entrance examinations, adopting human/computer interplay method and utilizing intelligence calculation network platform, we track and analyze the mental judgment standards and their changes of the decision-makers that change with the environment, conditions, experience and psychological inclination, on the basis of which we conduct dynamic simulation of comprehensive assessment of the decision-makers based upon the dynamic changes of mental judgment standards.