Mitigating Outlier Effect in Online Regression: An Efficient Usage of Error Correntropy Criterion

In this paper, a modified version of maximum correntropy criterion (MCC) with application in online regression (or adaptive filtering) is proposed. It is well known that information theoretic criteria such as error correntropy criterion (ECC) and error entropy criterion (EEC) have the advantage of better performance in supervised learning problems like regression and adaptive filtering when the error between system output and labels (sometimes called desired signals) contains outliers and/or does not follow a Gaussian distribution. Specifically, we improve the existing adaptive maximum correntropy criterion algorithm (known as AMCC) by simply eliminating major outliers during learning process. This elimination leads to better steady state performance than previously known algorithms.