【作者】 李锋刚；

【导师】 倪志伟； 杨善林；

【作者基本信息】 合肥工业大学 ， 管理科学与工程， 2007， 博士

【摘要】 如何解决企业和社会组织管理过程的优化，如何提高其智能化水平，是管理理论和实践中一个重要的课题。现代管理实践中系统规模越来越大，约束条件增多，非线性严重，环境更加复杂，致使系统优化难度越来越大。为解决这些问题，本文提出了一种新型的基于优化型案例推理的智能决策技术OCBR—IDT，其涉及管理科学、人工智能、运筹学和决策科学等诸多领域，是优化理论和决策技术的结合。具体的研究内容和主要贡献包括：1．在探讨案例推理技术起源、发展、基本思想和特点的基础上，研究了案例推理技术用于决策过程的各个阶段，包括决策案例的表示、决策案例的组织和索引、决策案例的修正、决策案例的学习及决策案例库的维护。探讨了案例推理中的核心技术k-NN在案例推理中的应用。提出了一种新型的基于优化型案例推理的智能决策技术OCBR—IDT，并将其应用于构建基于OCBR—IDT的智能决策支持系统，采用人机协调技术，集人的智能和机器的智能于一身，并采用基于案例推理和规则推理相结合的推理方法，提高了对决策过程的智能辅助程度和系统的灵活性、适应性。2．在回顾属性选择的策略，包括搜索策略和评价策略，以及属性选择的方法的基础上，对属性选择过程予以形式化。具体研究了基于熵的两种属性优化选择策略，即信息增益法和增益比率法，用层次化k-fold交叉验证和k—近邻(k—NN)相结合的技术考察其性能。结果表明基于熵的属性评价策略是一种有效的评价策略。利用遗传算法特有的遗传算子搜索机制，用基于相关性的启发式作为评价机制，提出一种GA—CFS方法，用于从属性集中选择对给定案例最优的属性子集。对选择出来的属性子集用C4.5算法和k—fold交叉验证相结合评价其分类性能。实验对比结果表明，GA—CFS方法可以确定出与分类和预测最相关的属性子集，同时在几乎不降低分类准确性的情况下，极大地减小属性的表示空间。研究了基于主成分分析(Prime Component Analysis，PCA)的属性选择，从总体主成分到标准化变量主成分，最后给出样本主成分的计算方法。使用Iris数据作为测试，对选出的主成分使用C4.5和k—NN方法考察其分类准确性。结果表明，使用这种方法可以降低数据维度，同时可以达到一定的分类准确率，当然提取出的主成分的实际意义需要根据具体的问题结合专业知识予以解释。3．在综合分析案例检索的类型、任务、启发式方法、案例相似性度量方法后，从不同维度研究了案例相似性的度量，即案例间相似性的度量和案例属性间的相似性度量。对基于几何模型的相似性度量方法和基于特征属性的相似性度量方法予以分析。分析了管理决策中的委托求解法和综合加权法后，提出了一种基于提升和投票的多策略相似性检索技术。其具有融合多个模型、集中群体智慧的特点，可以克服现有检索方法依靠单一模型进行决策所具有的不可靠性，以降低决策风险，增强决策的鲁棒性和智能性。4．详细探讨了禁忌搜索技术在算法设计过程中的关键参数，包括邻域和邻域搜索，禁忌表和禁忌表的大小，短期记忆和长期记忆，停止准则和搜索的效率等。探讨了这些参数设置的方法、实现的技术和不同设置对算法的影响。研究了用禁忌搜索技术求解图结构案例中的TSP问题。首先是对TSP问题的描述，接着研究了用朴素禁忌搜索算法求解随机生成的20城市TSP问题，并测试了算法的收敛速度和案例解的优化过程。还研究了基于启发式案例解决方案，给出了基于“贪婪”搜索的初始解启发式禁忌搜索对20城市TSP问题的求解结果。研究了禁忌搜索算法对图结构案例的检索技术。首先给出了图结构案例的定义，图结构案例相似性的度量，包括图中顶点和弧之间的映射和对应。接着是禁忌搜索各个阶段在图结构案例中的实现，最后给出了一个两阶段案例检索技术，同时使用了朴素禁忌搜索技术和高级禁忌搜索技术，提高了对图一类复杂案例的检索效率。最后探讨了对禁忌搜索算法的多种改进策略。5．综述了智能诊疗系统在医学领域的应用、国内外已开发的各类医学诊疗系统和中医专家系统，分析了案例推理技术在中医领域应用的重大研究价值和前景。将本文提出的OCBR—IDT应用于我国三大显学之一徽学中新安医学的研究中，构建基于案例推理技术的优化型智能新安医学防治中风病诊疗系统。从概念设计到结构设计研究了其实现的技术方法和路线，最后开发了此系统。为从病名诊断、辨证论治到处方用药提供智能化的决策支持，实现了新安医学防治中风病研究的现代化、信息化和智能化。更多还原

【Abstract】 How to optimize the managerial process of enterprises and social organizations and improve its intelligence has been a key issue in terms of both theory and practice. In modern management practice, the system is becoming increasingly larger, with more constraint conditions, more serous non-linear problems, and complexity of the surroundings, which add to the difficulty in the optimization of enterprise systems.In order to deal with these problems, an original Optimized Case-based Reasoning--Intelligent Decision Technique (OCBR-IDT) is presented in this paper, which touches upon Management Science, Artificial Intelligence (AI), Operation Research and Decision Science, et al academic fields, and it combines optimization theory (classic optimization and modern heuristic optimization technique) with decision technique. In detail, this paper includes:1. On the basis of a probe into the origin, development, basic thoughts and characteristics of Case-Based Reasoning (CBR), the dissertation does research on how to apply CBR into each stage of the decision process, including the presentation, establishment and index, as well as the revision and learning, and maintenance of the decision case-base. It also discusses the application of k-NN technique in CBR.The paper presents a new Optimized Case-based Reasoning--Intelligent Decision Technique (OCBR-IDT), and applies it to construct an intelligent decision support system. It adapts the man-machine coordinate techniques, integrates the man’s brainpower and machine intelligence, combining the case-based reasoning with rule-based reasoning. This improves the flexibility、adaptation of the system and its assistance degree to decision process.2. On review of the strategies of the attributes selection, its process is formalized. The thesis investigates two strategies of attributes selection based on entropy, e.g. information gain and gain ratio. And its performance has be examined by using the strata k—fold cross validation and the k-NN techniques.Employing the genetic operator, which characterizes genetic algorithm, as the searching approach and correlation-based heuristic as the evaluating mechanism, the thesis presents a GA-CFS method to select the optimal subset of attributes for a given case library. Then the author combines the C4.5 algorithm with k-fold cross validation to evaluate its classification performance. The compared experimental results indicate that the proposed method can identify the most related subset for classification and prediction, while mostly reducing the representation space of the attributes whereas hardly decreasing the classification precision.Attributes selection based on the Prime Component Analysis (PCA) is considered, from total PCA to standard variance PCA, and algorithms of sample PAC are given. Using Iris data as test case base, the selected Prime Components’ classification performance is checked by C4.5 and k—NN method. The experimental result shows that it can depress the data dimension as well as achieve certain classification precision.3. On analysis of the types, tasks, heuristics and similarity measure of case retrieval, the dissertation considers the measure of the case similarity from various dimensions, namely, between cases and between attributes of two cases. The methods of similarity measure based on the geometry model and the properties attributes are also analyzed.After introducing the methods of the commission problem-solving and syntactic weight sum, a multi-strategies similar case retrieval technique is presented, which is based on the boosting and vote. It can colligate multi model, collect group intelligence, overcome the lack of credibility in single model decision, and decrease the decision risk, while increase the robust and intelligence of decision process.4. The critical parameters in Tabu Search (TS) algorithm designs are discussed in detail, i.e. neighbors and neighbor search, tabu list and size of the tabu list, explicit memory and attributive memory, stop criterion and search efficiency. The author also explores the setting means of these parameters, its realization and influence on algorithm.The TSP problem belonging to map structure case is investigated. First the author describes the TSP problem. Then the author solves the 20 city TSP problem randomly generated using the na(?)ve TS algorithm and tests the convergence of the algorithm and optimization process of the case solution. Finally, the author investigates the case solution, which is based on the heuristics and reaches to the result of 20 city TSP problems using initial key, which are generated by "greedy" search strategy based on the heuristics TS algorithm.The case retrieval technique of map structure using the TS algorithm is presented. First the definition of map structure case and the similarity measure of it are put forward. Then its realization using TS algorithm and a two-stage case retrieval method is presented. It uses the naive and advance TS algorithm as well. Finally, the author investigates the improving strategies of TS.5. The author presents an overview of intelligent diagnosis systems’ applications in medicine in domestic and abroad, especially in Traditional Chinese Medicine (TCM), analyses its important research value and significance using case-based reasoning technique into TCM. The OCBR—IDT technique is applied into the Xin’an medicine belonging to Hui School which is one of the three prominent schools in China, and constructs an intelligent stroke prevention and treatment system based on Xin’an medicine, from the concept design, structure design to its development. It can provide the disease diagnosis, distinguishing evidence and determining treatment, recipes and herb with intelligent decision support. It contributes much to the modernization, information and intelligence of Xin’an TCM.更多还原