一种新的神经模糊系统及其学习算法

A NEW NEURO-FUZZY SYSTEM AND ITS LEARNING ALGORITHM

本文提出一种规则结论部分的语言变量具有离散隶属度函数的、基于Mamdani形规则的新神经模糊系统,并描述了它的学习算法。新神经模糊系统由模糊推理系统及其一对应的神经网络系统构成。在只有训练数据的情况下,首先提出了一种基于RBF神经网络的模糊建模方法。而在模糊推理系统由模糊建模或者直接由专家经验知识确定后,应用梯度下降法优化神经网络系统参数。倒立摆控制和时间序列预测的仿真试验体现了本文提出的新的神经模糊系统的可用性和优越性。

In this paper, a new Mamdani type neuro-fuzzy system, which possesses discrete membership functions in the consequent parts of IF-THEN rules, is proposed and the learning algorithm of the new system is also detailed. The new neuro-fuzzy system includes a fuzzy inference system and the corresponding neural network system. The RBF neural network based fuzzy modeling algorithm is provided for the situation in which only the input-output data are available. After the fuzzy inference system has been built by fuzzy modeling algorithm or by knowledge of experts, a descent gradient algorithm is applied to optimize the parameters in the system. The performance and superiority of the new neuro-fuzzy system is illustrated in the inverted pendulum control and chaotic time series prediction problem.