计算机工程与应用 ›› 2019, Vol. 55 ›› Issue (3): 120-125.DOI: 10.3778/j.issn.1002-8331.1710-0172

• 模式识别与人工智能 • 上一篇    下一篇

多任务正则极限学习机的研究与应用

睢璐璐,韩东升,闫  飞,阎高伟   

  1. 太原理工大学 信息工程学院,太原 030024
  • 出版日期:2019-02-01 发布日期:2019-01-24

Research and Application of Multi-Task Regularized Extreme Learning Machine

SUI Lulu, HAN Dongsheng, YAN Fei, YAN Gaowei   

  1. College of Information Engineering, Taiyuan University of Technology, Taiyuan 030024, China
  • Online:2019-02-01 Published:2019-01-24

摘要: MT-ELM通过隐含层共享不同任务间的数据特性实现多任务学习,但MT-ELM忽略任务间关联程度的差异以及存在的过拟合问题,为此提出基于MT-RELM软测量建模方法。首先,利用RELM解决过拟合问题;其次,考虑任务之间关联度的差异,基于相关性较强的任务其权值向量也较相似的假设,在每个任务输出权值的基础上加入约束条件,利用此约束条件表示任务间的相关程度;最后,利用ADMM算法迭代求解得到MT-RELM的模型参数。基于合成数据集与湿式球磨机数据集的结果表明,此算法可有效地提高模型的预测精度以及泛化能力。

关键词: 正则极限学习机, 多任务, 交替乘子法, 过拟合

Abstract: MT-ELM enables multi-task learning using hidden layers to share data characteristics among different tasks. However, MT-ELM ignores correlation differences between tasks and over-fitting problem. Therefore, MT-RELM is proposed. Firstly, RELM is used to solve the over-fitting problem. Secondly, considering correlation differences between tasks, the constraints are added to output weights based on the assumptions that similar task having similar weight, so this paper uses this constraint to indicate relevance level between tasks. Finally, ADMM algorithm are used to solve MT-RELM model parameters. The results based on the synthetic dataset and wet ball mill dataset show that this algorithm can effectively improve the prediction accuracy and generalization ability.

Key words: regularized extreme learning machine, multi-task, alternating direction method of multipliers, over-fitting