Probabilistic Regularized Extreme Learning Machine for Robust Modeling of Noise Data
提出一种概率正则化极限学习机，通过最小化建模误差的均值和方差来提升对非高斯噪声和异常值数据的建模性能，实验证明其比传统极限学习机更鲁棒。
The extreme learning machine (ELM) has been extensively studied in the machine learning field and has been widely implemented due to its simplified algorithm and reduced computational costs. However, it is less effective for modeling data with non-Gaussian noise or data containing outliers. Here, a probabilistic regularized ELM is proposed to improve modeling performance with data containing non-Gaussian noise and/or outliers. While traditional ELM minimizes modeling error by using a worst-case scenario principle, the proposed method constructs a new objective function to minimize both mean and variance of this modeling error. Thus, the proposed method considers the modeling error distribution. A solution method is then developed for this new objective function and the proposed method is further proved to be more robust when compared with traditional ELM, even when subject to noise or outliers. Several experimental cases demonstrate that the proposed method has better modeling performance for problems with non-Gaussian noise or outliers.