【作者】 周志坚；

【导师】 毛宗源；

【作者基本信息】 华南理工大学 ， 控制理论及应用， 1999， 博士

【摘要】 模糊技术、神经网络与遗传算法是计算智能的三大信息科学，是智能控制领域的三个重要基础工具。如何将模糊技术、神经网络与遗传算法互相交叉，有机地结合起来，已成为众多研究者关注的焦点。本论文围绕这一热点进行了两方面的研究。 一方面，在研究了遗传算法的基本原理、理论基础以及模糊逻辑与神经网络融合的基础与途径之后，详细研究了基于联接机制的模糊神经网络构成方法。围绕这一模糊神经网络结构，首先提出了一种基于遗传算法的三阶段优化策略。在给定初始参数基础上，利用基于十进制编码的遗传算法实现模糊神经网络的结构优化，用基于二进制编码的遗传算法实现模糊神经网络的参数优化。仿真结果表明该优化策略是有效的且简单易用。 上述优化策略是利用专家经验数据来完成网络结构与参数优化的，网络的性能很大程度取决于专家经验的准确程度，为了补偿或减少人为因数的影响，针对控制系统，本论文又提出了一种基于去模糊优化的模糊神经网络控制器。在利用上述三阶段优化策略获得网络的结构和参数后，重构网络控制器的去模糊化部分，进一步细调控制变量的隶属函数，实现控制器性能的优化。从仿真结果可以看出，引入去模糊优化后，控制器的性能有所改善。 上述两种优化算法均是基于经验数据对的，当未能有效获取经验数据时，上述方法很难实现。基于最优控制的思想，—种基于遗传算法的模糊神经网络最优控制方法被提出。在测得被控对象输入输出的基础上，通过对控制系统的过程模拟，利用遗传算法优化包含控制器性能的指标来离线寻找最优的模糊神经网络控制器结构和参数。经过遗传算法训练的模糊神经网络控制器被接入模糊神经网络智能控制系统中。仿真结果表明，利用此方法实现的控制，系统的动态性能和静态性能都优于用常规模糊控制器实现的控制。 另一方面，探讨了遗传算法、神经模糊技术在中医分型诊断中应用的可行性。针对具体的中医疾病一类风湿性关节炎，提出一基于模糊神经网络的分型诊断系统。在利用基于互信息的遗传算法实现了临床症状的压缩之后，以此为基础利用更多还原

【Abstract】 As three information sciences in computational intelligence, fuzzy Logic, neural networks and genetic algorithms are three important basic tools in intelligent control. Many research workers have concerned on the integration of techniques in these three sciences recently. This paper’s work focus on the application of the integration in two ways.On the one hand, after studying the basic principles and theories of genetic algorithm, and studying the base and way of the integration of fuzzy logic and neural networks, the structure of connectionist fuzzy neural network is studied at some length. Around this structure, an optimization strategy with three steps is presented firstly. Based on the initial parameters that have been determined in the first step, the structure of a fuzzy neural network is optimized by using a genetic algorithm with the decimal coding scheme, and the parameters are optimized with the binary coding scheme. Numerical simulations show that the optimization strategy mentioned above is available and easy to use.Though the optimization strategy mentioned above use experiment data from expert to optimize the network structure and parameters, the performance of the network depends on the accuracy of expert experience to a great extent. In order to compensate and eliminate the influence of manmade factors, a fuzzy neural network controller based on the optimization of defuzzification is presented in this paper again. After the fuzzy neural network is got through using the strategy mentioned above, the defuzzification part of it is reconstructed, and the membership functions of the control action is further refined. Then the performance of the controller is optimized. The results of computer simulation更多还原