中国科学院软件研究所机构知识库
Advanced  
ISCAS OpenIR  > 中科院软件所  > 中科院软件所
题名:
神经网络学习和进化学习的研究
作者: 孟祥武
答辩日期: 1997
专业: 计算机软件
授予单位: 中国科学院软件研究所
授予地点: 中国科学院软件研究所
学位: 博士
关键词: 神经网络 ; 神经网络学习 ; 进化神经网络 ; 遗传算法 ; 进化规划 ; 进化学习 ; 机器学习
摘要: 近年来,神经网络学习和进化学习是人工智能领域研究的热点。学习能力是人类智能行为最为本质的特性之一。所有的智能系统都是进化的。模拟进化是提供产生机器智能的一种方法。本论文综述了神经网络学习和进化学习,讨论了构造优化神经网络结构的一些方法,并对这些方法进行了分类,分析了这些方法的优缺点。评述了联接机制与符号机制相结合的机器学习研究工作。本论文就神经网络学习和进化学习中存在的几个重要问题,进行了较为深入的研究,主要工作如下:1.分析了基于知识的人工神经网络(KBANN)的优缺点,研究了领域理论和KBANN之间的关系。为了学习性能好,一个学习系统必须有效地利用已有的先前知识。KBANN用先前知识初始化一个神经网络。它把初始领域理论转变成神经网络,因此决定了网络拓扑结构和初始权值。理论分析和实验结果表明:初始领域理论极大地影响概念的学习,基于完善领域理论,KBANN能很好地学习。它的训练时间比标准神经网络的要短。基于不完善领域理论,有时KBANN不能学习一个简单的问题。增加隐结点可克服KBANN的这个局限性。KBANN应该并且必须改变它的初始网络拓扑结构。因为初始领域理论往往是近似正确的,不完善的或不正确的。2.讨论了进化神经网络的编码表示机制的分类和特性,分析了它们的优点和缺点。这些将有助于针对不同的应用,设计和选择编码表示。提出了遗传算法的一种图文法编码表示机制,给出了相应的算子定义,模式、模式长度及其阶的定义。图文法表示属文法编码方法。文法编码方法相对而言不受问题大小的影响。证明了一个基于图文法表示机制的遗传算法模式定理,描述了交叉和突变对模式作用的效果,该图文法可用来定义神经网络。神经网络的学习能力映射为适应值。3.证明了Hopfield网与图灵机等价。给出了用Hopfield风计算部分递归函数的构造性证明。由于部分递归函数与图灵机等,故Hopfield风与图灵机等价 4.提出了一个基于遗忘进化规划的Hopfield网学习算法。通过遗忘部分个体,算法能避免局部最小。用极限环表示概念,这很适合于字识别,一个字附有不同的字体。一个字的不同字体可用极限环表示、给定不动点,极限环或迭代序列,通过解不等式组算法能同时获得Hopfield网的拓扑结构和权传值。该算法克服了进化Hopfield网学习的局限性。它还能找到多个优化解,实验也证明了所提算法的有效性。5.提出了一个基于任意给给定训练集的离散型Hopfield网联相记忆学习算法。该算法能增加训练样本的维数,也就是在Hopfield网中增加结点,因而能存储任意给定的训练模式集。该算法克服了传统Hopfield网学习的局限性。实验结果也证明了该方法的有效性。6.提出一种用遗传算法求解文件分配问题的新方法。文件分配问题是计算机网络和分布系统中一个非常重要的问题。该求解算法简单,易于实现。它还能找到多个优化解。此方法可以方便地动态改变文件及其拷贝在各台服务器上的分布,能7较好地解决工种中的文件优化分配问题。该方法还可以应用到其它资源需要分配的领域。最后,第九章总结了全文,并给出了进一步的工作。
英文摘要: Neural network learning and evolutionary learning are the hot topics over the decade in artificial intelligence circle. The ability to learn is one of the most essential characteristic of human intelligent behavior. All intelligent systems are evolutionary. It is natural to simulate evolutionary processes in order to create machine intelligence. This dissertation reviews neural network learning and evolutionary learning. It discusses some methods to construct optimal neural network architectures, formulates the classification of these methods, analyses the advantages and disadvantages of these methods, describes some of the researches on machine learning that combine the symbolic and connectionist learning. This dissertation deals deeply with several important problems existed in neural network learning and evolutionary learning. The following is the main research works: 1. Chapter 3 analyses the advantages and disadvantages of Knowledge-Based Artificial Neural Networks (KBANN), studies the relation between domain theory and KBANN. Theoretic analysis and experimental results indicate that the initial domain theory can greatly affect how well concepts are learned. Based on complete domain theory, KBANN learns well. It's training is faster than "standard" neural networks. Based on incomplete domain theory, KBANN can not learn a simple problem sometimes. The limitation of KBANN can be overcome by adding nodes. 2. Chapter 4 discusses the classification and properties of the encoding representations of evolving neural networks, analyses the advantages and disadvantages of these methods. It presents an encoding representation of genetic algorithms based on graph grammar, gives the corresponding definitions of genetic operators, schema, schema order and length. It proves a schema theorem for genetic algorithms which representation schema is based on graph grammar. The effect of crossover and mutation on schemata is described. The graph grammar is used to define the neural network. The learning ability of the neural network is mapped into a fitness value. 3. Chapter 5 proves the equivalence of Hopfield neural networks and Turing machine. The partial recursive function is constructed by Hopfield neural networks. The partial recursive function is equivalent with Turing machine, the computability of Hopfield neural networks is therefore equivalent with Turing machine. 4. Chapter 6 presents a learning algorithm of Hopfield neural network based on evolutionary programming with forgetting. The algorithm can avoid local minima by forgetting some individuals. Under the constraints of fixed points, and the limit cycles of iteration sequences, the algorithm simultaneously acquires both the topology and weights for Hopfield neural network by solving inequalities. It copes with the limitations of evolving Hopfield learning algorithm. It can also find several optimal solutions. Experimental results also demonstrate the effectiveness of the algorithm. 5. Chapter 7 presents a learning algorithm of Hopfield's discrete-time associative memory with any given training set. The algorithm can add the dimensions of training patterns, so it can store any given training pattern set. It copes with the limitations of canonical Hopfield learning algorithm. Experimental results also demonstrate the effectiveness of the approach. 6. Chapter 8 presents a new method to solve file allocation problem based on genetic algorithms. The method can dynamically change the distribution of files on the servers. It can well solve optimal file allocation problem in engineering. Finally, chapter 9 summarizes the results of this dissertation. It concludes by presenting areas for future work.
语种: 中文
内容类型: 学位论文
URI标识: http://ir.iscas.ac.cn/handle/311060/6222
Appears in Collections:中科院软件所

Files in This Item:
File Name/ File Size Content Type Version Access License
N98829.pdf(2545KB)----限制开放-- 联系获取全文

Recommended Citation:
孟祥武. 神经网络学习和进化学习的研究[D]. 中国科学院软件研究所. 中国科学院软件研究所. 1997-01-01.
Service
Recommend this item
Sava as my favorate item
Show this item's statistics
Export Endnote File
Google Scholar
Similar articles in Google Scholar
[孟祥武]'s Articles
CSDL cross search
Similar articles in CSDL Cross Search
[孟祥武]‘s Articles
Related Copyright Policies
Null
Social Bookmarking
Add to CiteULike Add to Connotea Add to Del.icio.us Add to Digg Add to Reddit
所有评论 (0)
暂无评论
 
评注功能仅针对注册用户开放,请您登录
您对该条目有什么异议,请填写以下表单,管理员会尽快联系您。
内 容:
Email:  *
单位:
验证码:   刷新
您在IR的使用过程中有什么好的想法或者建议可以反馈给我们。
标 题:
 *
内 容:
Email:  *
验证码:   刷新

Items in IR are protected by copyright, with all rights reserved, unless otherwise indicated.

 

 

Valid XHTML 1.0!
Copyright © 2007-2017  中国科学院软件研究所 - Feedback
Powered by CSpace