在线增量极限学习机及其性能研究  

Research on online incremental extreme learning machine and its performance

作　　者：马致远[1] 罗光春[1] 秦科[1] 汪楠 Ma Zhiyuan;Luo Guangehun;Qin Ke;Wang Nan(School of Computer Science ＆ Engineering,University of Electronic Science ＆ Technology of China,Chengdu 611731,China;School of Information Science ＆ Engineering,East China University of Science ＆ Technology,Shanghai 200237,China)

机构地区：[1]电子科技大学计算机科学与工程学院,成都611731 [2]华东理工大学信息科学与工程学院,上海200237

出　　处：《计算机应用研究》2018年第12期3533-3537,共5页Application Research of Computers

基　　金：国家自然科学基金青年科学基金项目(61604054);中央高校基本科研业务费项目(ZYGX2016J083);四川省科技厅应用基础项目(2017JY0027)

摘　　要：针对在线学习中极限学习机需要事先确定模型结构的问题,提出了兼顾数据增量和结构变化的在线极限学习机算法。算法以在线序列化极限学习机为基础,通过误差变化判断是否新增节点,并利用分块矩阵的广义逆矩阵对新增节点后的模型进行更新,使模型保持较高的正确率。通过在不同类型和大小的数据集上的实验表明,所提算法相较于经典极限学习机及其在线和增量学习版本都具有较好的分类和回归准确率,能够适应不同类型的数据分析任务。Aiming at the problem that extreme learning machine（ ELM） requires to determine the structure of the model in online learning,the paper proposed an on-line incremental ELM algorithm that supports both data increments and structural changes. On the basis of on-line sequential extreme ELM,the algorithm determined whether to add a new node or not by monitoring the changes of error. It used generalized inverse of block matrix to help update the model after new node was included so that the model maintains high accuracy. Experiments on different datasets with diverse volumes show that,the proposed algorithm has better classification and regression accuracies compared with classic ELM and its on-line and incremental learning versions,and it is suitable for various types of data analytical assignments.

关 键 词：极限学习机 增量学习 在线学习 广义逆 在线增量极限学习机 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]