【作者】 何春梅；

【导师】 徐蔚鸿；

【作者基本信息】 长沙理工大学 ， 计算机应用技术， 2006， 硕士

【摘要】 不确定性普遍存在于主观和客观世界中,模糊性是它最重要的形式之一。不确定性人工智能是人工智能的深化和发展,现已经成为人工智能研究的热点和重大的前沿课题。而模糊逻辑系统和神经网络系统是不确定性人工智能的重要研究内容,在构建许多不确定性智能系统时,通常要为系统准备先验知识,比如模糊逻辑系统的已知规则和神经网络系统的训练模式。先验知识的不确定性引起了事物的复杂性和处理难度,这种不确定性从数学角度可以认为是一种误差,从动力学角度可以认为是一种摄动,有效获取先验知识以及评估和控制先验知识的误差对系统性能的影响已经成为不确定性智能系统的重要研究内容。有效获取先验知识已有很多工作,但始终没有关键突破;评估和控制先验知识的误差对系统性能的影响在传统模糊推理和模糊神经网络方面取得了一定的进展,但在一般神经网络还未开展,为此本文主要做了如下工作:(1)研究了训练模式摄动对高斯型径向基函数网络性能的影响,指出了训练模式摄动的常见情形,建立了相关的定义和引理,接着从理论上严格证明了对某些模式摄动该网络的输出摄动幅度不放大,并用MATLAB编写了一个仿真实验应证了该理论证明的结果,该部分工作为径向基函数网络的性能分析提供了一个新的研究点,对训练模式的获取过程能提供警示和指导。(2)提出了一种基于遗传算法的根据已知系统控制目标自动获取模糊控制规则的方法。该方法与其他多种方法相比较,具有能根据系统的控制目标,自动搜索获得最优(次优)模糊控制规则集的优点,并在倒立摆控制系统进行了仿真应用,具有较好的仿真效果。最后本文对不确定性智能系统的一些研究问题进行了展望。更多还原

【Abstract】 Uncertainty, one of whose most important forms is fuzziness, exists universally in the objective and subjective world. Artificial Intelligence with uncertainty, the advanced development of artificial intelligence, is a hot and important research topic now. And the fuzzy logic system and neural network system are the important research topics of artificial Intelligence with uncertainty. When we create many uncertain intelligent systems the transcendental knowledge, for example the known rules of the fuzzy logic systems and the training patterns of the neural network systems, should usually be prepared for the systems. The uncertainty of the transcendental knowledge, which can be thought of as errors from the viewpoint of mathematics and perturbation from dynamics, causes the complexity and solving difficulty of the things. Effective acquisition the transcendental knowledge, in which there has been much work but there is not pivotal breakthrough, and evaluation and controlling the errors of the transcendental knowledge impacting the performance of the systems, which has gotten some achievement in traditional fuzzy inference and fuzzy neural network but has not begun in universal neural network, have been two important research topics of the uncertain intelligent systems. The main work in this paper is as follows:1. This paper studies the influence of the uncertainty of patterns perturbation on the performance of Gauss Radial Basis Function Neural Network. It gives interrelated definitions and lemmas, points the usual instances of the training patterns perturbation and then theoretically and strictly proves that the outputs perturbation of some patterns perturbation of the Gauss Radial Basis Function Neural Network is no bigger than the training patterns perturbation. Lastly it gives an experiment in MATLAB validating the validity of the theoretically inferred result. This part of work gives a new viewpoint in studying the performance of the Gauss Radial Basis Function Neural Network.2. A method, which can automatically searches for the most or secondary optimized fuzzy control rules set according the controlling object of the systems, automatically acquiring fuzzy rules based on the genetic algorithm is proposed in this paper. An experiment illustrates the procedure of acquiring rules by genetic algorithm in pendulum system. The simulation results demonstrate the feasibility and effective-更多还原