【作者】 赵辉；

【导师】 王普；

【作者基本信息】 北京工业大学 ， 检测技术与自动化装置， 2015， 博士

【摘要】 案例推理(Case-based reasoning,CBR)是人工智能领域中一种新的推理技术和机器学习方法,目前在模式识别、智能控制、决策支持等领域均有典型应用。其中,在模式分类领域中,案例推理分类器的综合性能包括分类准确率、求解效率以及可靠性等指标是影响CBR成功应用于现实领域的重要因素。如何提高案例推理系统的综合性能与学习能力,获得高准确率、高效率、高可靠性的CBR分类器具有重要的研究意义。然而,在其发展过程中,有待解决的首要理论问题是CBR的懒惰学习机制决定了其创造性思维能力的不足,另外,缺乏一套系统化的理论和方法来建立CBR的动态学习模型等。这些科学问题都是当前普遍关注而尚未完全解决的问题,从而限制了CBR综合性能的提高。因此,对CBR推理模型开展进一步的方法研究具有重要的理论价值。本文针对CBR的四个阶段,即案例检索、案例重用、案例修正和案例存储进行了系统研究,借鉴无线通信领域中的注水原理,引入案例评价环节,利用选择性的案例维护策略设计了一种新的具有动态学习能力的案例推理分类模型。取得的主要成果如下:(1)针对案例检索环节常见的属性权重分配方法存在学习能力不足的问题。借鉴无线通信领域中的注水原理并结合分式规划思想,提出一种权重动态迭代学习方法对属性权重进行优化分配。该方法首先将案例属性类比为子信道,将每个属性值的波动性值作为权重分配的指标得到属性容量的表达式,并通过构建拉格朗日函数来求得静态初始权重;其次依据案例库中检索到的近邻案例信息构建了属性权重迭代调整的指标,并根据系统性能变化情况对属性权重进行迭代调整,从而得到合理的权重结果,为后续进行的案例间相似度的准确评估奠定了基础;(2)针对传统CBR系统对求解效果进行事后评价导致重用效果不佳的问题,在原有的推理求解框架基础之上(案例重用之后)增加了案例评价环节。通过引入案例建议结论的可信度计算指标,对案例重用得到的建议解答进行可信度评价:首先利用案例重用得到目标案例的建议结论,并根据此建议结论将检索到的近邻案例划分为同类近邻与异类近邻两类;其次引入可信度评价指标来计算出建议结论的具体可信度值,然后设定可信度阈值,来据此评价建议结论是否可信以及可信的程度;最后评价为可信的案例可直接重用建议结论,而对评价为不可信的案例,则提出基于“二次检索”的结论修正策略来得到最终的结论,从而得到一种具有动态适应能力的评价-修正策略。可信度评价环节的引入为用户与决策者提供了采纳建议结论的依据,降低了不可信结论带来的不确定性风险,为提高CBR系统的可靠性提供了保障;(3)针对案例存储环节的增量学习机制会使得案例库规模不断增长从而导致推理效率不高的问题。提出了一种选择性的案例库动态维护策略。首先定义案例属性重要度的权重阈值,并采用遗传算法对权重阈值进行迭代搜索,在此基础上,将权重小于该阈值的属性删除;其次,通过引入相应的吸收与删除准则,对案例库中满足条件规则的案例进行选择或删除,从而剔除冗余案例而组成新的子案例库,达到动态维护案例库的目的。该维护策略可以有效避免效用问题的产生,减小冗余案例和有害案例对推理可靠性的不利影响,不仅降低了问题求解的时间复杂度,还提高了CBR分类器的分类效率;(4)为验证本文方法的有效性与优越性,分别进行了CBR分类器的实验与应用研究。主要利用UCI机器学习资源库中提供的典型分类数据集,设计了十折交叉验证的实验方案。分别对权重的动态迭代学习方法、动态评价-修正推理策略以及选择性的案例库动态维护策略进行了对比实验。结果表明这些方法提高了CBR分类系统的分类准确率,保障了系统的运行效率。在实验成功的基础上,将具有动态学习能力的CBR模型应用于TE化工过程的故障诊断中,从诊断的准确率、灵敏度、特异度等方面考察了本文方法的有效性。结果表明,与其他诊断方法相比,具有动态学习能力的CBR系统能够显著提高故障诊断的准确率,有效提升了CBR系统的综合性能和学习能力,为CBR在类似工业过程中的应用提供了方法基础。更多还原

【Abstract】 Case-based reasoning(CBR), as a new reasoning technology and machine learning method, has been successfully applied on the pattern recognition, intelligent control, decision support and other fields. Especially in the pattern classification field, the performance of the CBR system, such as the classification accuracy, classification efficiency and the classification reliability are the key factors that affect the success of the application of CBR. Therefore, to improve the performance of CBR and achieve a CBR system with high accuracy, high efficiency is of great significance. However, in the process of building the CBR model, there exists some theoretical problems, such as the limitation of the CBR’s creative thinking ability caused by lazy learning mechanism and the lack of theory support on constructing the dynamic CBR learning model. These problems have been widely concentrated on and still remain unsolved, which restrict the comprehensive improvement of CBR performance. Therefore, further investigation on the CBR is of great importance. Aiming at the four steps that affect the learning ability of CBR, namely case retrieve, case reuse, case revise and case retain, an improved CBR classifier with dynamic learning ability is constructed, where water-filling principle, case evaluation link, selective case maintenance strategy are introduced. The detailed contents are listed as follows:(1) Aiming at the weak learning ability caused by the statistical weight in the retrieval process. A dynamic iterative weight distribution method is proposed, which combines the water-filling principle in the wireless communication field and the thought of fraction programming: firstly, this method regards each attribute as sub channels, utilizes the volatility parameters of each attribute to achieve the initial statistical weight; then the iterative adjustment indexes are constructed according to the information in the nearest neighbors, to get the final weight result, which lays a reliable basis for the similarity calculation and the following reasoning steps.(2) In order to solve the bad reuse performance brought by the traditional reuse mechanism of CBR system, an evaluation link is added into the 4R cycle(after case reuse), where a dynamic evaluation-revise strategy is proposed: firstly, acquire the suggested solution through case reuse, according to which divide the nearest neighbors into the two set, one is a set of which the cases have the same category with the target case, the other one is the set of which cases have different categories with the target case; secondly, trustworthiness indexes are proposed to calculate the specific trustworthiness value of the suggested solution, then a trustworthy threshold is set to evaluate the case as trustworthy or untrustworthy; at last, the cases in trustworthy set could directly reuse the suggested solution and restored in the case base, while the untrustworthy set should conduct the following case revise step. As for the case revise, a strategy based on second retrieval is proposed to obtain the final result of the untrustworthy set. The evaluation link decreases the uncertainty and potential risks that untrustworthy cases may bring, which improves the learning ability and reliability of CBR system.(3) As for the low efficiency problem caused by the incremental learning mechanism of CBR, a dynamic case maintenance strategy based on case selection is proposed: firstly, a feature reduction method based on threshold optimization is proposed, where genetic algorithm is adopted to search an optimized weight threshold, according to which could remove the attribute whose weight is lower than the threshold to accomplish the feature reduction step; then the absorption and removal criteria are introduced to direct the case maintenance: select the cases that satisfy the rules to construct the subset to reduce the case base size. The proposed case maintenance strategy could avoid the utility problem, reduce the bad influence caused by redundant cases and harmful cases, decrease the time complexity, and guarantee the reliability of CBR.(4) In order to test the effectiveness of methods proposed in the dissertation, some experiments are conducted with 10 datasets from UCI data base, where 10-fold cross validation method is deployed in the experiments to test the effectiveness of the dynamic iterative weight distribution learning method based on water-filling, the evaluation-revise strategy based on trustworthiness, and the case maintenance strategy base on selection respectively. The experimental results show that the improved CBR could achieve a better classification performance, which verify that the proposed methods could effectively improve the accuracy, efficiency and the dynamic learning ability of CBR. In the end, CBR model with dynamic learning ability is utilize to build a diagnosis system for the application on TE pro-cess, the results(diagnosis accuracy, diagnosis sensitivity, specificity, ROC) on the faults of TE process indicate that the proposed methods could provide a satisfying diagnosis performance, which proves the effectiveness and superiority of CBR on improving the learning performance and reliability of CBR system, providing the methodology foundation for similar industrial processes.更多还原