样本问题

We firstly summarize the NN system modeling methods and the popular optimization algorithms, and point out the deficiencies, the limitation of noise cancellation modeling and the need of human intelligence participation, in our former optimization work after analysis. An abstract problem is extracted. It is to find out control variant that makes system output optimized using the data comprising a sample set with N x samples and their corresponding y, and it does not explore given system and achieves full automatic computation procedure. Next we propose a multi-variant controlled system optimization method utilizing CASSANDRA-I small type neural computer and program its application software. Moreover systematic experiments are made to demonstrate the method's validity, ability of noisy samples disposal and practical application.

我们总结了神经网络系统建模方法以及得到广泛采用的优化算法，并分析指出我们前阶段的优化工作中存在的去噪声建模有局限性和计算过程需人干预的不足，在此基础上把问题抽象为：已从系统获得含N个控制量样本的样本集S以及每个样本对应的系统输出，如何依据这些数据求出一个使系统输出y得到优化的控制量x，而不研究具体系统，实现不需要人智能参与自动完成的计算过程；提出应用人工神经网络硬件系统——CASSANDRA-I小型神经计算机的多变量控制系统优化方法，设计完成了其应用软件；并进行系统的实验工作，验证算法的有效性以及此方法处理带噪声样本和实际应用能力。

The paper consists of three parts. First, we give the calculation for every order central moment for the Logistic population, mean vector and covariance matrix of the order statistic of random sample from the Logistic population, and means of several special functions for the order statistic of random sample from the Logistic population. Secondly, we discuss the parameter estimation for the Logistic population. Finally, we solve the question about the goodness-of -fit test for the Logistic population.

研究的主要内容分三部分：第一部分是关于Logistic总体各阶中心矩、样本次序统计量期望向量与协方差阵及若干个特殊的样本次序统计量函数期望的计算问题；第二部分是关于Logistic总体分布参数的估计问题；第三部分是关于Logistic总体的拟合优度检验问题。

Aiming at the problems of over-fitting and generalization for neural networks, the locally most entropic colored Gaussian joint input-output probability density function estimate is studied, and a new method by means of Chebyshev inequality is proposed to self-revise respectively according to every cluster.

针对神经网络的过拟合和泛化能力差的问题，研究了样本数据的输入输出混合概率密度函数的局部最大熵密度估计，提出了运用Chebyshev不等式的样本参数按类分批自校正方法，以此估计拉伸样本集，得到新的随机扩充训练集。

The FHS-SVM based on the structural risk minimization criterion can solve the small sample learning problem, which can get a good shallow buried target and clutter classification performance with only shallow buried target training samples to obtain the parameters of hypersphere. Furthermore, the factors of misclassification risk and bury environment diversity are combined into the discriminator study procedure using the fuzzy membership of training samples, which improve the practical value of the FHS-SVM in shallow buried target discrimination.

FHS-SVM基于结构风险最小原理，在有效解决小样本学习问题的同时，只需要浅埋目标训练样本就能优化超球面参数，获得较好的浅埋目标和杂波分类性能；并且利用训练样本的隶属度将误判风险和埋设环境多样性等因素融入鉴别器学习过程，提高了FHS-SVM浅埋目标鉴别算法的实用性。

The method obtained multiple training samples from a single face image by multi channel sampling before two-dimensional independent component analysis transformation. Experimental results on the ORL face database and YALE face database show that the proposed method is feasible and has higher recognition performance compared with GREY, PCA, ICA, 2DICA, projection combined PCA, FLDA, sampled FLDA and other algorithms where only one sample image per person is available for training.

该方法是在2DICA运算之前，首先对单训练样本进行采样，通过多频率采样可以获取多个不同频率下的采样样本，然后对采样样本进行2DICA特征提取，最后采用神经网络分类识别，对人脸库ORL和YEL作了相关实验，将该方法与GREY、PCA、ICA、2DICA、PC PCA、FLDA、Sampled FLDA等传统方法作了比较，最终证明了该方法可以有效地解决单训练样本人脸识别的问题。

The Hausdorff measure for the level sets of multi-parameter symmetric stable processes has been obtained and the inner connection of the measures and local times for it has been demonstrated, which describes the fractal nature for the level sets of this kind of processes. As we know, it is always one of the focuses in the field of random fractal. The reasult of the existence and continuity of the local time for multi-parameter stochastic processes with stable components has been got in the project. Then we gained the measure for the random fractal sets such as the range and graph of this kind of processes. Furthermore, a progress besides the mentioned above in the project is that we have resolved the multifractal analysis of the sample paths for Brownian sheet, that is the multifractal decomposition of white noise in high dimension.

解决了多指标对称稳定过程水平集的Hausdorff测度问题，同时给出N指标d维稳定过程水平集的Hausdorff测度下界的最佳估计，它们揭示了其水平集的测度与局部时之间的内在关系，刻划了多指标对称稳定过程水平集的分形特征，表明了研究此类问题的积极意义，该问题是随机分形研究领域的国内外学者所关注的热点问题之一；解决了多指标稳定分量过程的局部时存在性与连续性问题，进而解决了由多指标稳定分量过程产生的随机分形集--象集、图集的测度问题；此外，本研究项目还解决了布朗单样本轨道的重分形分析，即高维白噪音的重分形分析问题。

To the problem that large-scale labeled samples isn't easy to acquire in the course of SVM training, the active learning strategy is used in the SVM training and an incremental training algorithm of active SVM based on the uncertainty based sampling of distance ratio is proposed in the paper. The experimental results show that the active SVM learning strategy can considerably reduce the labeled samples and costs compared to the passive learning method, and at the same time it can ensure the accurate classification performance is kept as the passive SVM and also expedite the SVM training.

针对 SVM训练学习过程中难以获得大量带有类标注样本的问题，本文将主动学习策略应用于SVM增量训练中，提出了一种基于距离比值不确定性抽样的主动SVM增量训练算法，实验结果表明在保证不影响分类精度的情况下，应用主动学习策略的SVM选择的标记样本数量大大低于随机选择的标记样本数量，大大降低了标记的工作量或代价，而且训练速度同样有所提高。

In chapter 4, basing theories and methods of scientific visualization, and artificial neural network BP algorithm, we integrate the Visual C++, OpenGL graphics library and Excel VBA technique to develop the program of artificial neural network and to make the BP algorithm visually, this program works can be divided into four parts: Using C language to develop program about BP algorithm; Using Visual C++, develop the GUI Interface, make input parameter visually; Using OpenGL graphic technique to display the training sample point in three dimension; at last using Excel DDE technique display the error graphic tables in Excel system In chapter 5, on the view of engineering application, we establish new method of surface reconstruction basing artificial neural network, develop interface program between module and commercial CAD/CAM system, meantime deeply discuss some key problems, for example, setting up the base plane, using the API technique, cutting and editing surface boundary, and also discuss the more compliant problem: how to intersect surface, at end we finish the work of translation from our surface reconstruction module to commercial CAD/CAM system, then make reverse engineering system basing artificial neural network more useful.

第四章基于科学计算可视化理论，依据人工神经网络BP算法理论模型，综合Visual C++，OpenGL图形库以及Excel VBA等多项软件开发技术，编制了人工神经网络程序，实现了BP算法的可视化映射。具体工作分为四部分：利用C语言实现人工神经网络BP算法；利用VisualC++的GUI技术开发图形用户界面，实现参数设置可视化；利用OpenGL图形技术进行三维映射，显示学习样本及训练样本点；利用微软电子表格DDE动态数据交换技术，在Excel上动态显示学习误差曲线图。第五章从工程应用的角度出发，提出了一种新的基于人工神经网络算法的曲面裁剪重构方法，完成了曲面重建模块与通用CAD/CAM系统的接口设计工作，对其中的若干关键问题进行了深入讨论，例如基平面设定、API技术的应用、边界裁剪等问题，同时，对曲面计算中较为困难的曲面相交问题也进行的专门探讨，最终完成了曲面重建模块向CAD/CAM系统的数据传输工作，使人工神经网络逆向工程系统趋向实用。

Based on the analysis on the physical property of complex HRRP samples, we point out that the frame template, classification algorithm and feature extraction method for complex HRRP samples should be unvaried with the initial phases.

在分析复数HRRP样本特性的基础上，指出由于初相敏感性问题，原先适用于实数HRRP样本的方位模板、识别方法和特征提取方法一般都不能直接用于基于复数HRRP的RATR，我们必须重新寻找既与复数HRRP样本的初相无关又能利用其剩余相位信息的方法。

In a database the concept of an example might change along with time which is known as concept drift When the concept drift the classification model built by use of old data is unsuitable for classify new data Therefore concept drift has become a hot issue in data mining in recent years Although many algorithms had been proposed to resolve this problem they can not provide users with the reason of concept drift However a user might be very interested in such rules For example doctors want to find what makes disease change; researchers want to know the reason of the variety of the weather; and decision makers would like to understand why a customer's shopping habit change In this thesis we propose a Concept Drift Rule mining Tree called CDR-Tree to solve this problem CDR-Tree can not only find the rule of concept drift also build the prediction model for both old and new data at the same time

无论在大型资料库或现实生活中，同一资料样本的概念有可能会随著时间的递移而改变，也就是产生所谓的概念漂移。当样本发生概念漂移时，由旧有资料所建构的分类模组将不再适用於预测新获得的资料，因此，近年来概念漂移已成为资料探勘中一项热门的研究议题。虽然已有？多学者提出不同的技术来解决概念漂移的问题，但是这些方法都是利用修正或重建的方式来产生适合新资料的预测模组，并无法提供造成概念漂移的原因。然而对使用者而言，其感兴趣的可能正是这些引起概念漂移的规则，如医生可能想了解引起疾病变化的主因、学者会想要知道气候转变的规则、或是决策者想找出顾客购物习惯改变的因素等。因此，本论文提出概念漂移规则探勘树( Concept Drift Rule mining Tree )，简称CDR-Tree，来解决这个问题。CDR-Tree不但能探测出造成概念漂移的主要原因，亦能同时建立新旧资料的预测模组以供决策者运用使用。

And Pharaoh spoke to Joseph, saying, Your father and your brothers have come to you.

47：5 法老对约瑟说，你父亲和你弟兄们到你这里来了。

Additionally, the approximate flattening of surface strip using lines linking midpoints on perpendicular lines between geodesic curves and the unconditional extreme value method are discussed.

提出了用测地线方程、曲面上两点间短程线来计算膜结构曲面测地线的方法，同时，采用测地线间垂线的中点连线和用无约束极值法进行空间条状曲面近似展开的分析。