【作者】 潘翔；

【导师】 姚明海；

【作者基本信息】 浙江工业大学 ， 控制理论与控制工程， 2002， 硕士

【摘要】 在人工智能领域，模式识别是一个非常重要的方面。本文就如何提高识别性能对模式识别系统进行了研究。 本文首先对基于BP学习算法的神经网络模式识别方法进行分析和研究，指出了其存在的缺点，利用模糊逻辑的知识表达能力，在神经网络中引入模糊逻辑，构成模糊神经网络来改善网络性能，针对模糊神经网络的参数难以选择问题，提出了基于遗传算法的模糊神经网络学习模型。但是单一分类器由于其采用的特征类型单一性以及自身的局限性，改进的网络模型虽然在一定程度上提高了识别性能，但是这种提高有限。 研究发现，不同分类器在识别性能上有互补作用，因此如何把各种分类器结合在一起，从而能够集成各个分类器的优点，而抑制它们的缺点，是提高识别性能的关键，对该问题的研究，虽然已经取得了一些成果，但是研究主要集中于抽象级信息的集成，而忽略了表征分类器性能的完整信息——度量级信息，因此不能从根本上解决分类器性能的互补问题。 针对上述问题，本文对多分类器集成进行了深入研究。首先总结了用于模式识别的多分类器集成机理以及存在的问题，并对各种基于抽象级信息的建模方法进行分析和比较，指出了基于抽象级信息的分类器集成存在的缺点，提出了基于度量级信息的分类器集成，并采用模糊积分进行集成建模，对于模型中的模糊积分密度难以确定问题，采用了一种动态评价方法——贝叶斯方法，仿真结果表明了该建模方法的有效性。更多还原

【Abstract】 As an important aspect in the domain of artificial intelligence, pattern recognition can extend the application range of computer and improve the ability of computers to perceive outside information. In this paper, research about how to improve recognition performance is presented.In the thesis, neural network pattern recognition based on BP algorithm is analysised and researched , and its disadvantage is pointed out. With the knowledge representation ability of fuzzy logic, we use fuzzy logic in neural network, and build up fuzzy neural network to improve network performance. Furthermore, We develop a kind of fuzzy neural network learning model based on Generic algorithm to solve the problem that how to decide the parameters of network and better performance can be achieved, while single classifier use single feature and has its limitation, improved network can’t be get expected result.It’s suggested that different classifier offered complementary information, which motivated the interest in combining classifiers to harness their strength. Some achievements are obtained, however most research focused on combination based abstract information and ignored the combination based on measured information, so complementary problem of classifiers hasn’t been solved completely.for the above problem, a deep research about the combination of multiple classifier is presented, at first, we conclude the principle of combination multiple classifier and existed problem, and compare the different methods based on abstract information, also their drawback is pointed out. Then combination of measure information is described with fuzzy integral, and a dynamic evaluated method, bayes method, is to decide fuzzy integral density of model. Simulation shows efficiency of the method更多还原