具有遗忘机制的在线宽度学习算法  

Online Broad Learning System with Forgetting Mechanism

作　　者：包洋 郭威 BAO Yang;GUO Wei(College of Computer Science and Technology,Nanjing Tech University,Nanjing 211816,China;College of Information Engineering,Yancheng Teachers University,Yancheng 224007,China)

机构地区：[1]南京工业大学计算机科学与技术学院,南京211816 [2]盐城师范学院信息工程学院,江苏盐城224007

出　　处：《吉林大学学报（信息科学版）》2022年第6期1017-1025,共9页Journal of Jilin University（Information Science Edition）

基　　金：国家自然科学基金资助项目(61603326)。

摘　　要：对动态数据流的在线学习问题,传统的在线BLS(Broad Learning System)算法无法准确地捕捉数据最新的变化趋势。为此提出两种具有遗忘机制的在线BLS算法——基于遗忘因子的在线BLS算法(FF-OBLS:Online Broad Learning System based on Forgetting Factor)和基于滑动窗口的在线BLS算法(SW-OBLS:Online Broad Learning System based on Sliding Window)。FF-OBLS在在线学习过程中通过为旧样本添加遗忘因子以体现新旧样本对学习模型的不同贡献,SW-OBLS在在线学习过程中通过删除旧样本以消除旧样本对学习模型的影响,从而使学习模型对动态数据流的后续趋势做出更准确的分析和预测。为验证提出的两种在线BLS算法的有效性,使用动态回归数据集进行实验。实验结果表明,具有遗忘机制的在线BLS模型在预测精度和时间开销上均优于传统在线BLS模型,更适合处理动态数据流问题。For the online learning problem of dynamic data flow, the traditional online BLS(Broad Learning System) algorithm can not accurately capture the latest change trend of the data. Therefore, two online BLS algorithms with forgetting mechanism, one is based on forgetting factor(FF-OBLS: Online Broad Learning System based on Forgetting Factor) and other is based on sliding window(SW-OBLS: Online Broad Learning System based on Sliding Window), are proposed. FF-OBLS reflects the different contributions of old and new samples to the learning model by adding forgetting factors to old samples in the online learning process, SW-OBLS eliminates the impact of old samples on the learning model by deleting old samples in the online learning process, so as to enable the learning model to accurately analyze and predict the subsequent trend of dynamic data flow. In order to verify the effectiveness of the proposed two algorithms, dynamic regression data sets are used in the experiment. The experimental results show that the online BLS models with forgetting mechanism are better than the traditional online BLS model in the perspective of prediction accuracy and time cost, therefore they are more suitable to deal with dynamic data flow problems.

关 键 词：宽度学习系统 动态数据流 遗忘机制 遗忘因子 滑动窗口 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]