【作者】 翁时锋；

【导师】 张长水；

【作者基本信息】 清华大学 ， 控制科学与工程， 2003， 硕士

【摘要】 神经网络是通过模拟生物神经系统的结构功能而构成的一种信息处理系统。经过40余年的发展,神经网络已经成为机器学习、模式识别、信号处理等信息科学领域中一类十分重要的方法。本文首先概括地总结了神经网络的发展历程。在三个不同阶段中,神经网络的发展具有不同的特点。到20世纪90年代,神经网络成为信息科学中一个重要的热点研究领域。文中还讨论了神经网络的特点、常见的分类方法、学习类型和学习算法等内容。其次,文章着重介绍了一种重要非监督学习神经网络:SOM特征映射。Kohonen提出的Self-Organizing Mapping(SOM)特征映射是一种用于高维数据可视化的有力工具。它能够揭示隐藏在高维数据中的复杂的非线性关系,并将其在低维空间中以简单的几何关系展现出来。最后,本文提出一种对SOM网络的改进:自适应SOM特征映射(ASOM)。由于SOM的投影空间通常是一维或二维的固定的网格,因此,它并不能很好地完成非监督数据集的聚类分析任务。本文在已有算法的基础山,提出了一种具有随机进化机制的ASOM网络。在ASOM网络中,每一个神经元具有根据局部环境而进化的能力。这些进化操作包括:消亡、分裂、合并等。微观分析表明:这些操作不仅有利于神经元学习到在SOM输出平面上的合理分布,而且在更一般的意义上,可以克服竞争学习中一些不稳定的情况。我们还从多智能体系统和人体免疫系统的角度对ASOM网络进行分析和讨论,发现了其中一些有意义的联系和有趣的现象。实验表明,ASOM具有很好的自适应性和鲁棒性,可以有效应用于高维数据的内在结构挖掘,是对经典SOM网络的一种重要的改进。更多还原

【Abstract】 Neural networks are the information processing systems constructed bysimulating the structure and function of the biological neural system. Itappeared at 1960s. After the development of more than 40 years, neuralnetworks have become a group of the important methods in the field of theinformation science, such as machine learning, pattern recognition and signalprocessing.This thesis firstly summarized the development experience of neuralnetworks. After three dramatic development stages, neural networks havebecome one of the focus research area in information science. Thecharacteristics, classification, learning types and learning algorithms werealso discussed.Secondly, a famous unsupervised learning neural network, SOM, wasintroduced that is proposed by Kohonen. Essentially, SOM is a featureprojection, which can be applied in revealing the nonlinear statisticalrelationship hidden in the high-dimensional data.Thirdly, this thesis proposed an improved version of SOM to enhance theadaptability, robustness and the ability of mining the dataset’s clusteringtendency. The new neural network was name as Adaptive SOM (ASOM). Themajor algorithmic feature of ASOM was that all the neuron units in the outputplane decided to execute the evolutionary working mechanisms according totheir own local environments. The evolutionary mechanism includes splitting,vanishing, combining and etc. The microcosmic analysis showed that thoseevolutionary mechanisms provided a solution to the problems that faced in theconventional learning of competitive networks. We also discussed ASOMneural network from the viewpoints of multi agent system and human immunesystem, and found some interesting relationships between them.The empirical evidences showed the proposed algorithm was adaptiveand robust. This work could be viewed as an important development of classicKohonen self-organizing feature map.更多还原