【作者】 秦中广；

【导师】 毛宗源；

【作者基本信息】 华南理工大学 ， 控制理论与控制工程， 2002， 博士

【摘要】 粗糙集作为一种分析和处理不完整、不确定数据的有效数学工具，已在理论和实际应用上取得长足的发展。本文主要针对粗糙集中的约简和规则抽取进行交叉研究，并结合国家中医药管理局基金项目(A97105)进行实际应用 首先，提出一种新型的属性约简策略ARS(Attributes Reduction Strategy)。该策略把属性值的个数应用到属性约简上，大大降低进行属性约简的计算复杂度和空间复杂度，在约简大量数据更能显示其优越性。并利用属性约简判定ARS策略是有效的属性约简策略。 其次，在属性约简策略ARS基础上优化BP网络结构，约简BP网络结构输入层和隐层单元。在约简BP网络结构过程中，根据属性本身信息的重要程度不一样的情况，提出基于权重的属性约简策略ARS-W，在BP网络结构的隐层单元约简取得良好效果；而且把模糊综合评价引入BP网络结构评价，为约简BP网络结构提供评判标准。 再次利用属性约简策略ARS优化模糊神经网络，简化输入层结构和第三、四层结构，解决传统模糊神经网络初始设计中，输入空间划分可能出现“维数灾难”的问题和系统的初始规则数会随输入的维数呈指数增长的问题，从而为模糊神经网络结构的优化提供了新的途径。 本文还对粗糙集在分散控制的应用进行初步探讨，针对复杂系统中由于各子系统相互关联，其分散控制器的控制规则难以人工提取的问题，充分利用子系统的输入、输出数据，抽取分散控制器的控制规则。该方法无需进行模糊化和去模糊化，控制直观、简单、方便。 最后结合国家中医药基金项目(A97105)，实现中医类风湿关节炎智能诊断。利用粗糙集从类风湿关节炎病例数据中自动抽取诊断知识，构建中医类风湿关节炎智能诊断系统(TCMRAIDS)和在线诊断(RAONLINE网站)，并在广州中医药大学得到成功应用。更多还原

【Abstract】 Rough Set works as an effective mathematic tool for dealing with uncertain and unholonomic data. Its theory and application get quick development in past ten years. The crossing researches on reduction and rule extraction are the main points of this dissertation, which are funded by State Administration of Traditional Chinese Medicine (A97105).Firstly, a new attributes reduction strategy (ARS) is discussed. This method combines the number of attributes values into attributes reduction, which reduces the time complexity and the space complexity. It’s advantage can be showed more clearly when dealing with a great number of data.Secondly, ARS is used to reduct the neurons in input layer and hidden layer for optimizing BP Neural Network structure. During reduction process, attributes reduction strategy based on weight (ARS-W), when the significance of every attribute itself is different, is used in reducting the neurons in hidden layer. This method takes effect; Fuzzy Multi-Factorial Evaluation is used in BP Network reduction structure evaluation, which offers a criterion for BP Network structure.Thirdly, ARS is used to reduct the structure of input layer 、 third layer and fourth layer to optimize Fuzzy Neural Network. This method offers a good solution to dimension disaster in input space and original rules exponential growth that may occur in traditional Fuzzy Neural Network original structure.Fourthly, an exploration of Rough Set in decentralized control is discussed. This method makes full use of sample data, extract the control rules of the decentralized controller. This method doesn’t need fuzzification and defuzzification. So the control is simple and convenient.更多还原