【作者】 胡静；

【导师】 高隽；

【作者基本信息】 合肥工业大学 ， 计算机应用技术， 2007， 博士

【摘要】 机器学习以知识的自动获取和产生为研究目标，是人工智能领域研究的热点问题之一。分类器的优化设计一直是机器学习、模式识别和数据挖掘等领域研究的核心问题，它在图像识别、语音理解、自然语言处理、医疗诊断及Web页面的分类等领域具有广泛的应用。如何提高分类器对环境的适应能力，是分类器优化设计的关键问题。应用机器学习方法实现分类器的优化设计，是机器学习和人工智能界的一个重要研究课题。本文以机器学习及其在神经网络分类器设计中的应用为研究内容，重点研究了机器学习在求解神经网络分类器设计中的网络结构优化和样本选择优化问题的方法。本文的主要研究工作如下：1)从分类器设计的角度出发，讨论了机器学习的最新发展方向及面临的具体问题，并对近年来出现的一些新机器学习方法进行了分析和研究。2)研究了流形学习的分类器网络结构优化设计问题，针对利用神经网络对同一对象的非线性结构样本集进行分类和识别时，如何合理地设计网络结构的问题，提出了一个新颖的基于低维参数空间估计的神经网络结构设计的方法。该方法以流形学习为基础，结合Sammon系数有效估计出低维参数空间大小，并将此对应到神经网络结构分组设计的隐节点分组数目上，从而设计出具有一定泛化能力的网络结构。3)研究了主动学习的分类器样本优化选择准则：针对模糊神经网络分类器设计过程中所遇到的样本采样与标记过程耗时长、代价大的问题，提出了一个新颖的模糊神经网络样本主动选择准则，以最小—最大边界法以及确定样本的不确定性阈值两个新概念来定义样本的选择标准，确保选择其中信息量尽可能大的样本进行标记，使得网络设计过程中对未标记样本的标记工作量和时间大为减少。4)未标记样本的模糊神经网络分类器设计：针对已标记和未标记样本的混合分类问题。提出了一种基于非刻度—多维度收缩的、新的排序—模糊神经网络分类器模型。该模型首先利用非刻度—多维度收缩法对输入的所有样本进行了排序，然后获得样本间的相似性测度值，并利用该相似性测度值指导随后的分类器超盒扩张与压缩过程，从而使得该模型不仅提高了对未标记样本进行有效分类的性能，而且无论是在网络结构方面，还是在训练时间方面都有所改进。更多还原

【Abstract】 Machine Learning makes a target of automatic retrieval and produce of knowledge. It has become one of the key areas in artificial intelligence and machine learning. The optimization design of classifier also is nuclear question in field of Machine Learning, Pattern Recognition and Data Mining, it has wide application in image recognition, speech understanding, medical treatment diagnosis and classification of web page. To improve the adaptation to the environment of classifier is the key problem for optimization design of classifier. Making use of machine learning methods to realize optimization design of classifier becomes an important research topic in machine learning and artificial intelligence.The research topic of this paper is machine learning and its applications in design neural network classifier. It has been focused on methods of optimization design of neural network classifier’s structure and samples selection. The research topic of this paper has 4 parts as follow:1) From the point of design classifier view, the questions of the develop way and problem of machine learning have been discussed, some new methods appeared in recent years have been analyzed and researched.2) Research of optimization design of classifier’s structure. Based on manifold learning. A novel approach of designing of neural networks based on parameter space in the low-dimension manifold was proposed to solve the problems about neural networks design rationally, which is used in recognition and classification of congener samples with non-linear configuration. This method based on manifold learning combines Sammon stress in order to estimate the value of parameter space in low-dimension, furthermore this value corresponds with the number of hidden in neural networks.3) Research of optimization selection of classifier’s samples based on active learning. A novel approach of active learning based on fuzzy neural network classifier was proposed to solve the problems of surprisingly time consuming and costly in sample collection and annotation. Two new concepts of Min-Max Margin Based Approach and Uncertainty threshold on samples were introduced as a rule of active sample selecting to guarantee the most informative samples annotated. Therefore, the annotation and time cost were greatly reduced.4) Design of fuzzy neural network classifier of unlabeled sample. A novel kind of Ordination-Fuzzy min-max neural network (OFMM) based on non-metric multidimensional scaling (MDS) was proposed to solve the classification problems of unlabeled input pattern. Firstly, all the input patterns were sorted by MDS to get their similarity measures. Then these measures were used to supervise the following expansion and contraction stage of hyperboxes for classification. OFMM had improvements both in the validity of unlabelled patterns classification, and the network structure and training time.更多还原