用于鲁棒性建模的概率权重极限学习机-计算机科学.PDF

Vo1. 44 No.8 第 44 卷第8 期 计算机科学 2017 年 8 月 COMPUTER SCIENCE Aug.2017 用于鲁棒性建模的概率权重极限学习机 周闻范彬朱曹陆新江 (中南大学机电工程学院 长沙 410083) 摘 要极限学习机(ELM)在机器学习领域获得了很多的关注，并在应用方面取得了极大的成功。然而，极限学习 机对训练数据中的异常值点和非高斯噪声非常敏感，ρ、而大大阻碍了 ELM 的应用。概率权重 ELM 方法主要对含有 异常值和非高斯噪声数据集进行建模，首先建立概率局部 ELM模型，并在此基础上利用 Parzen 窗方法建立局部模型 的概率分布，然后将概率分布作为权重来融合所有的局部模型以建立全局鲁棒性模型。该方法成功地应用了数学例 子和UCI 实例，并与传统ELM、正则化 ELM 和鲁棒ELM进行了比较分析，结果表明概率权重ELM表现出了较好的 建模性能。 关键词 极限学习机，鲁棒性，噪声，概率分布，异常值 中图法分类号 TP181 文献标识码 A 001 10. 11896/j. issn. 1002-137叉 2017.08.041 Probabilistic Weighted Extreme Leaming Machine for Robust Modeling ZHOU Chuang FAN Bin ZHU Lei LU Xin-jiang (College of Mechanical and Electrical Engineering ,Central South University ,Changsha 410083 ,China) Abstract Extreme 1earning machine (ELM) has attracted a 10t of attention in the machine 1earning fie1d and gained great success in application. However , it is sensitive to outliers and non-Gaussian noise in the training dataset ,which greatly hinder the application of EL鸟也 Probabilistic weighted ELM was proposed to model the dataset with the outliers and non-Gaussian noise. First ,a distributed local ELM modeling is developed ,upon which the probability distribution function CPDF) of multiple local models is estimated by the Parzen window method. Th凹， the distribution function is further used as weights to integrate alllocal models to construct a global robust ELM mode l. The successful application of this robust probabilistic weighted ELM method to both artificial case and rea1 1ife case as well as its comparison to