Abstract: Aimed at the problem that the low recognition rate and the poor generalization ability in traditional neural networks, an improved learning algorithm of Self-Organizing Fuzzy Neural Network（SOFNN） is presented. In this algorithm, it is as a basis for modifying, deleting and adding neurons that each neuron output and the sum of all these neurons output in the Ellipsoidal Basis Function（EBF） layer are stored. Then it can obtain effective neurons of the network and reduce the training time of samples. In order to ensure the network convergence, it uses the least square method（RLSE） to estimate the parameters and uses gradient descent method to modify the parameters. Compared with other fuzzy neural networks, the superiority in the accuracy, structure complexity and anti-jamming is effectively verified in the real data set.

Key words: fuzzy neural network, self-organizing, prediction model, Ellipsoidal Basis Function（EBF）

摘要： 针对传统神经网络识别率低和泛化能力差的问题，提出了一种改进的自组织模糊神经网络（SOFNN）学习算法。以保存椭球基函数（EBF）层各个神经元的输出及输出之和为依据进行神经元的修改，删除和增加，进而得到网络的有效神经元，并减少样本训练的时间。用最小二乘法（RLSE）估计参数，用梯度下降法修改参数，保证网络收敛。与其他的模糊神经网络相比，在精确度、结构复杂性和抗干扰性方面的优越性，在真实数据集上得到了有效的验证。

关键词: 模糊神经网络, 自组织, 预测模型, 椭球基函数（EBF）