【作者】 陈捷；

【导师】 杨开英；

【作者基本信息】 武汉理工大学 ， 计算机应用， 2006， 硕士

【摘要】 模块化神经网络(Modular Neural Networks)是一个由多个模块组成的神经网络,每个模块承担神经网络的全局任务的一个子任务,通过各个网络之间的竞争或协作来提高系统的整体性能。大量的实例研究表明,模块化神经网络在泛化能力和可靠性上比单一神经网络都有所提高,并且能有效地解决单个神经网络应用和实现时出现的问题,为研究者提供了一条问题求解的新途径。由于认识到模块化神经网络所蕴含的巨大潜力和应用前景,大量研究者涌入该领域,理论和成果不断涌现,使得模块化神经网络成为机器学习和神经网络领域的研究热点。 本文以回归问题为例就模块化神经网络系统中各个网络的组合方法展开研究,提出了基于“一专多能”的模块化神经网络子网集成的3种方法。该方法对训练样本充分利用,通过适度缩放训练集,训练出一批神经网络个体,这样产生的个体神经网络既拥有相对的准确性,又不失其对其他样本的适应性;将这样的网络进行个体的策略集结,可以使得系统的逼近速度、抗干扰能力、适应能力都得到了显著的提高。文章以8个回归实例进行仿真测试,其结果通过4个评价指标从不同的侧面反应其性能。从结果中看到,“一专多能”的方法在多个试验中获得了优胜,无论对于较难学习的问题还是较容易学习的问题上,始终保持着很好的适应性和误差要求,对于复杂的问题其精度优势表现得更为明显。 本文还描述了作者所开发的一套适合于模块化神经网络研究的《模块化神经网络仿真系统》,详细介绍了系统的总体设计和各模块的功能(包括:输入输出、算法、接口、程序逻辑、存储分配)以及系统界面设计等。该系统力求在原有的算法基础上有所突破、创新,并以此搭建起一套模块化神经网络的试验和研究平台。由于各模块具有开放、通用、可拓展性的特点,也为后期模块化神经网络的教学实践、科研探索提供了有利的支持。更多还原

【Abstract】 A Modular Neural Network (MNN) is a Neural Network (NN) that consists of several modules, each module carrying out one sub-task of the neural network’s global task. Networks are combined in a fashion of competition and cooperation to improve the performance of the overall system. A great deal of successful applications demonstrates that modular neural network outperforms single neural network in terms of generalization and reliability. And modular neural network can be effectively used to solve the single neural network’s problem in its application and implementation for us with a new tool for problem-solving. Having recognized that modular neural network has an enormous potential and bright prospect in application, a large number of researchers plunge them into the field, yielding a lot of relevant theories and application achievements. Also it has become a hot topic in both machine learning and neural network fields.In this dissertation, the first aspect is mainly studied in the context of regression problems. Three methods for integrating the component networks base on thinking of "Knowing something of everything and everything of something" are presented. These methods take full advantage of the training samples, and educate a series of training units by use of moderation zoom training pack. Not only are these neural network units in possession of corresponding veracity, but also have the flexibility to other samples. To make a tactics collection for the unit in this kind of network will improve the speed of getting closer, the ability of anti-jamming, flexibility and make a great progress in the system. This dissertation make an emulation test by using 8 regression problems instances, and the results reflect the capability in other sides analyzed by 4 evaluation standard. Compared the result with other results, the method "Knowing something of everything and everything of something" wins the best in most of experiments. No matter what degree of difficulty the problems are, this method could retain good flexibility and error requirements. For complicated problems, the characteristic of precision advantage could be represented obviously.This dissertation also presents the Simulation System of Modular Neural Network. The system, designed by author, is propitious to the research of modular neural network, it describes the design in total for system implementation, each module’s functions including input/output, algorithms, interface, program logic, memory distribution, and the design of system interface, etc. In order to get somebreakthroughs and innovation on the base of the original algorithm, this system makes progress on some sections and builds up a platform for the experiment and research in modular neural network. Because each module has opening, general, and extensibility characteristics, in upper period, the system could provide advantaged support for the practice teaching and science exploring.更多还原