基于高斯函数与分级学习的D-FNN算法研究

Research on D-FNN Algorithm Based on Gauss Function and Classification Learning

D-FNN基本思想是构造一个基于扩展的RBF神经网络,它可以看成是一个TSK模糊系统,也可以看做是基于归一化的高斯RBF神经网络。该文提出的算法,学习前,模糊神经网络不需要预先确定,在学习的过程中,参数估计与结构辨识同时进行,并根据系统精度要求及模糊规则的重要性,自动地产生或者删除一条模糊规则。在学习速度、系统结构和泛化能力方面进行了仿真实验,仿真结果表明D-FNN具有更简洁的结构和优良的性能。

Dynamic fuzzy neural network （D-FNN）, whose basic idea is to construct a RBF neural network based on extension, could be seen as a TSK fuzzy system, as well as a Gaussian RBF neural net work based on normalized. In the algorithm proposed, fuzzy neural network does not need to be predeter- mined before learning. During the process of learning, parameter estimation and structure identification are done simultaneously, and a fuzzy rule would be automatically generated or deleted, according to the system accuracy requirement and importance of fuzzy rules, Simulated experiments are performed in terms of learning speed, system structure and the generalization ability. The results show that D-FNN has more concise structure and more excellent performance.