Smart initialisation and approximating loss function for robust regression

Thomas Servotte, Jakob Raymaekers, Tim Verdonck

Abstract: Highlights•We investigate the benefit of using the wrapping technique to get smart initialisation weights for LTS and LMS.•We investigate the benefit of using approximating loss functions using soft sorting for LTS and LMS.•We demonstrate the potential application of Soft LTS and wrapping initialisation to non-linear regression with neural networks.•Extensive empirical experiments on simulated and real datasets demonstrate the potential of the suggested improvements.