【作者】 何斌；

【导师】 朱学峰；

【作者基本信息】 华南理工大学 ， 控制理论与控制工程， 2005， 博士

【摘要】 机器学习已经被广泛应用到各种智能系统中，它是计算机系统有目的地自动增进知识并改善系统功能的基本手段，是使机器具有智能的根本途径之一。 针对智能学习系统目前仍不具备或仅具备有限的智能学习能力的问题，为了解决智能应用系统中存在的动态性、复杂性和非结构化的学习问题，特别是控制系统和决策系统中不相容问题的学习问题，本论文把可拓逻辑引入机器学习方法之中，提出了“基于可拓逻辑的机器学习”的研究课题，并从形式化和模型化的角度，较系统地研究了“基于可拓逻辑的机器学习”，提出了三种基本的创新学习形式，即“相似学习”、“逆向学习”和“基于问题的学习”，研究了相关的学习策略，建立了适合于解决不相容问题的机器学习方法，对于提高现有机器学习方法解决不相容问题的能力，改善和提高智能系统的性能，具有理论和实践价值。 本论文的主要研究内容和研究结果如下： (1)研究相似学习，建立了相似性的形式化模型，引入δ-相似的概念和δ-相似可拓元的概念，然后给出了测度相似性的定量函数，提出了解决不相容问题的相似替代原理，它是解决不相容问题的有效方法。还建立了三条相似性推理基本规则。研究表明，相似推理是类比推理的推广。相似学习既丰富了相似理论的内容，也扩展了机器学习的相关内容。 (2)研究逆向学习，建立了四种逆向元，讨论了基本元的逆开拓，分析了变换的逆开拓，研究了逆向解的逆向度计算方法，探讨了逆向推理问题，并给出了相关的逆向推理规则。然后，给出了逆向学习的基本步骤。研究表明，逆向学习可运用逆向开拓和逆向推理展开。 (3)研究基于问题的学习，以可拓逻辑为基础，系统性探讨了问题的表示、问题的分析、问题的求解和问题的管理等几个问题；给出了问题的可拓表示方法，探讨了问题的发散性分析方法、共轭性分析方法、关键分析方法和传导分析方法；研究了问题求解的变换方法，特别综合分析了目标变换的上下位蕴含变通方法、转换桥方法和临界方法；指出，除条件变换和目标变换外，利用可拓变换进行环境变换，也是解决不相容问题的重要方法。 以上三种学习方式都是以可拓逻辑为工具，并在集成各种相关学习理论(包括心理学学习理论、机器学习理论和组织学习理论)的基础上发展而成的，都是面向不相容问题求解的，并利用各种恰当的变换和推理进行学习的学习方式。 本论文是把可拓逻辑应用于机器学习的一项探索性研究工作，是以作者十年来进行可拓逻辑和人工智能的研究工作为基础的。所做的探索性研究尝试提供一种新的学习方法，这种学习方法不仅使计算机智能系统解决不相容问题和矛盾问题成为可能，而且使机器学习理论的研究内容和研究方法得到扩展，也为更有效地解决以往的学习问题提供更多还原

【Abstract】 Machine learning has now been applied to all kinds of intelligent systems. It is the basic means to increase knowledge automatically and improve system’s performance purposefully. It is also a significant ability for achieving machine intelligence.Considering the inadequate and limited learning ability of intelligent systems for the time being, and in order to deal with dynamic, complicated and non-structured control and decision systems, especially the incompatible problems in these systems, this paper combines extension logic with machine learning, and proposes the research on machine learning based on extension logic. In this paper, machine learning based on extension logic is studied systemically from the viewpoint of formalization and modeling. Three kinds of basic innovation learning, including similarity learning, inverse learning and learning based on problems, are discussed, and their inference and relevant learning tactics are explored as well.The purpose of the research hopes to raise intelligence standard to some extent for intelligent systems to solve incompatible problems, and thus reach the target to improve the intelligent performance for intelligent systems in the future.The main result of the paper is listed as follows:(1) The paper studies similarity learning. Formalized models for describing similarity is set up, the concepts of δ-similarity and δ-similarity extension elements is introduced, the quantitative calculation for measuring similarity is given, similarity substitution principles for solving incompatible problems are put forward, three basic similarity reasoning rules are developed. The result shows that similarity reasoning is the generalization of analogy. Similarity learning enriches both the content of similarity theory and that of machine learning.(2) The paper studies inverse learning. Four inverse elements are given, the inverse extension is analyzed, the quantitative calculation for measuring inverse solutions is discussed, inverse reasoning is studied and relevant inverse reasoning rules are developed. The result shows that inverse learning can be expanded by inverse extension and inverse reasoning.(3) The paper studies the learning based on problems. Based on extension logic, the representation, analysis, solving and management of problems are discussed. The extension representation of problems, the divergent analysis methods, conjugate analysis methods, key analysis methods and conduction analysis methods are studied. The transforming methods of goal transformation for solving problems, including the upper and lower implicative transformation method, transforming bridge method and critical method, are introduced. Besides the goal transformations and condition transformations, the environment transformations by extension transformations are of significance in solving incompatible更多还原