一种基于增量加权平均的在线序贯极限学习机算法  

An incremental weighted average based online sequential extreme learning machine algorithm

作　　者：张明洋 闻英友 杨晓陶 赵宏 ZHANG Ming-yang WEN Ying-you YANG Xiao-tao ZHAO Hong(School of Computer Science and Engineering, Northeastern University, Shenyang 110819, China State Key Laboratory of Software Architecture, Neusoft Corporation, Shenyang 110179, China)

机构地区：[1]东北大学计算机科学与工程学院,沈阳110819 [2]东软公司软件架构新技术国家重点实验室,沈阳110179

出　　处：《控制与决策》2017年第10期1887-1893,共7页Control and Decision

基　　金：国家863计划项目(2015AA016005);国家自然科学基金项目(61402096;61173153;61300196)

摘　　要：针对在线序贯极限学习机(OS-ELM)对增量数据学习效率低、准确性差的问题,提出一种基于增量加权平均的在线序贯极限学习机(WOS-ELM)算法.将算法的原始数据训练模型残差与增量数据训练模型残差进行加权作为代价函数,推导出用于均衡原始数据与增量数据的训练模型,利用原始数据来弱化增量数据的波动,使在线极限学习机具有较好的稳定性,从而提高算法的学习效率和准确性.仿真实验结果表明,所提出的WOS-ELM算法对增量数据具有较好的预测精度和泛化能力.Considering the problem that online sequential extreme learning machine（OS-ELM） has low efficiency and accuracy when processing incremental data, an online sequential extreme learning machine algorithm based on incremental weighted average（WOS-ELM） is proposed. The principle of this algorithm is weighting model residuals trained with raw data and ones trained with incremental data, using the function got as the cost function of this algorithm, deducing a training model which is able to balance raw data and incremental data, and using raw data to weaken the fluctuation of incremental data. With this principle, the algorithm can raise stability, learning efficiency and accuracy of the OS-ELM.The result of a simulation experiment shows that the proposed algorithm has a higher predicting accuracy and better ability of generalization compared with other algorithms.

关 键 词：单隐层前馈型神经网络 在线序贯极限学习机 加权平均 增量 代价函数 

分 类 号：TP273[自动化与计算机技术—检测技术与自动化装置]