基本信息
作者: 铁锦程
学生学号:
班级:
学校:
指导老师:
写作日期:
字数: 25,000
编号: 45
类别: 工科〉人工智能
级别: 博士论文

[摘要]　
智能系统的研究对象通常存在严重的不确定性，用常规方法无法建立其精确 的模型。按实现的功能不同，这些对象的模型可分为决策型和映射型。模糊 逻辑和神经网络分别是建立这两类型模型的有力工具，同时遗传算法在以上 建模过程的参数优化中显示邮较强的能力。近来兴起的办计算是研究这三种 方法以及一些期货方法互相协同的学科分支。本文在软计算的架构下以模糊 逻辑、神经网络、遗传算法为工具，对智能系统建模进行了较深入的研究， 提出了一些新的智能系统建模方法。这些方法充分利用了以上三种方法的特 点：模糊ｉｆ．．．ｔｈｅｎ．．．形式模糊规则表达人类知识中的模糊概念和推理的能 力；神经网络具有的能够从训练数据中不断学习以实现一个期望映射关系的 强大学习能力；遗传算法的全局和局部搜索能力，并特别注意到使它们在应 用中能够互相配合和补充，以求最大限度地利用已有信息构造智能系统的模 型。 本文的主要成果有：
１．模糊系统设计中模糊规则的提取是关键，同时隶属函数形状的合理取值 也对模糊系统的性能影响极大。本文提出了一种基于模糊ｃ－均值算法和Ｂ Ｐ网 络的模糊规则提取方法。这种算法针对Ｔａｋａｇｉ方法中因为硬限聚类而使数据 利用率不高的情况，引入了模糊ｃ－均值聚类算法对采样数据进行预处理，以 得到规则数和规则前件之隶属函数，同时利用ＢＰ网络来实现训练数据对之间 的映射。采用这种方法更有效地利用了采样数据所包含的信息，且结构易于 实现。仿真的结果说明了它的有效性。
２．提出了一种基于模糊Ｋｏｈｏｎｅｎ网络和ＢＰ算法的模糊规则提取方法。这种 方法首先针对Ｋｏｈｏｎｅｎ网络的缺陷，采用模糊ｃ－性簇聚类算法予以改进，构 成了基于模糊ｃ－线性簇聚类算法的模糊Ｋｏｈｏｎｅｎ网络。这种网ＧＡ的最大特点 是对某些特定分布的对象具有较高的聚类速度，因此非常适用ＧＡ于训练数据的 预处理。用这种网络作为前件，ＢＰ网络作为后件所形成的模糊规则提取方法 是一种完全基于神经网络的模糊规则提取形式，且仿真说明其具有较好的效 果。
３．对遗传算法用于模糊规则优化进行了较系统的分析和研究，分别从ＧＡ算 法用于模糊集合的优化和规则集的优化两方面进行了阐述。对ＧＡ算法用于模 糊集的优化从编码方法、隶属函数的选择、交叉算子和变异算子的选择等方 面均给出了较一般的算法。对ＧＡ算法用于模糊规则集的优化给出了几个很典 型的实例。
４．传统的遗传算法其编码串是位置决定的且定长度。本文提出了一种新的 遗传算法编码形式－关联编码，这种编码不固定长度且不是位置决定的，在 实现交叉时不需执行对齐，从而简化了相应的执行步骤。其特殊的废码能够 有效地吸收交叉和变异操作带来的码混乱。仿真结果说明这种遗传算法比常 规的遗传算法要有效的多。
５．四足机器人动步态行走控制系统是一个多变量非线性系统。本文引入一 种分层递阶模糊控制器的设计方法，不仅可以极大地减少其模糊控制规则， 而且可以减化模糊控制逻辑的设计复杂度，从而为四足机器人提供了一种简 单实用的模糊控制器。本文不仅从理论的角度验证了该设计方法在多变量系 统中的有效性，而且针对四足机器人的仿真研究也证明了该方法的实用性。
６．提出了一种具有普适性的模糊规则优化算法。从原始模糊规则集中迭代 辨识出模糊关系，并利用Ｋｏｈｏｎｅｎ特征映射对模糊关系表中元素较集中的点进 行聚类，提取模糊关系的特征值，从而变换出反映控制系统特征的模糊控制 规则。该算法不仅可以应用于四足机器人模糊控制器的设计，而且对其它复 杂大系统的模糊规则优化问题也具有普适性。

The models of intelligent systems are often imprecise and difficult to build with the conventional control methods. According to their different functions, these models can be divided into two types, decision type and mapping type. Fuzzy logic and neural networks are useful tools to construct these two type models respectively. At the same time, genetic algorithms also show its power in the parameter optimization during the modeling process. Soft computing is a booming research field that studies the collaboration of the above three and some other technologies. Under the framework of soft computing, this dissertation proposes some new methods for intelligent systems modeling using fuzzy logic, neural networks and genetic algorithms, which are the main components of soft computing. These methods take full use of the characteristics of the three technologies, that is, explicit knowledge expression capability of human knowledge in fuzzy if...then... style rules of fuzzy systems, learning capability to realize a mapping relation from training data of neural networks, global and local search approach capability of genetic algorithms. The aim of these methods is to achieve cooperative effect by combining the above different characteristics to construct models of intelligent systems from given information. The main contributions of this dissertation are as follows: 1.In order to solve the key problem in the design of fuzzy systems, the rule extraction problem, a new algorithm based on the fuzzy c-means (FCM) algorithm and back propagation (BP) network is proposed. This algorithm uses fuzzy clustering algorithm as the rule's premise part to preprocess the sampled data, while a BP network is used to realize the mapping relation between I/O pairs. Simulation results show its high efficiency. 2.The Kohonen network is firstly improved by the fuzzy c-linear family clustering algorithm. The main characteristic of this network lies in its high efficiency in clustering some types of data. Then a rule-extracting method based on this network and BP is given. 3.A comprehensive survey on the optimizaton of fuzzy systems using genetic algorithms is given. Some general methods of coding the parameters of fuzzy sets and fuzzy rules are proposed. Studies on operator choosing and fitness function building are also presented. 4.A new coding scheme of GA, which we call Guanlian coding GA, is proposed. The superiority of the algorithm to conventional GAs is demonstrated by the utilization of various measures used to compare them. The application chosen for this is the inverted pendulum stabilizing controllers. 5.In order to decrease the size of the fuzzy rule bases and to simplify the fuzzy control system of the quadruped robot, this paper leads into a method of designing a hierarchical fuzzy control architecture. The method clusters the most important system variables as the first layer variables and the other system variables as the second layer variables. 6.The dissertation also proposes an optimization algorithm of fuzzy control rules which identifies the fuzzy relation's distribution over an area. The algorithm can cluster the fuzzy relation and transform it into optimized fuzzy rules. Besides the quadruped robot application, the method are also show to be applicable in many other complicated systems.