Abstract: During the construction process of Extreme Learning Machine (ELM), its input weights are randomly generated, and these parameters are non-optimized and contain no prior knowledge of the inputs. To solve these problems, combining the clustering method of Extension Neural Network type 2 (ENN-2), an extension clustering based extreme learning machine (EC-ELM) neural network was proposed. In EC-ELM neural network, the radial basis function centers of hidden neurons were firstly taken as the input weights, then extension clustering method was used to adaptively adjust the hidden neurons number and center vectors, and this well-adjusted information was trained by Moore-Penrose generalized inverse to obtain the output weights. Meanwhile, the effectiveness of this network was tested by the Friedman#1 dataset and the Wine dataset. The results indicate that EC-ELM provides a simple and convenient way to train the structure and parameters of neural network, and it is of higher modeling accuracy and faster learning speed than Extension theory based Radial Basis Function (ERBF) or ELM, which will provide a new way to apply the EC-ELM to complex process modeling.

Key words: extension clustering, extreme learning machine, radial basis function, regression, classification

摘要： 针对极限学习机(ELM)算法随机选择输入层权值的问题，借鉴第2类型可拓神经网络(ENN-2)聚类的思想，提出了一种基于可拓聚类的ELM(EC-ELM)神经网络。该神经网络是以隐含层神经元的径向基中心向量作为输入层权值，采用可拓聚类算法动态调整隐含层节点数目和径向基中心，并根据所确定的输入层权值，利用Moore-Penrose广义逆快速完成输出层权值的求解。同时，对标准的Friedman#1回归数据集和Wine分类数据集进行测试，结果表明，EC-ELM提供了一种简便的神经网络结构和参数学习方法，并且比基于可拓理论的径向基函数(ERBF)、ELM神经网络具有更高的建模精度和更快的学习速度，为复杂过程的建模提供了新思路。

关键词: 可拓聚类, 极限学习机, 径向基函数, 回归, 分类