A note on robust descent in differentiable optimization

In this note, we recall two solutions to alleviate the catastrophic cancellations that occur when comparing function values in descent algorithms. The automatic finite differencing approach (Dussault and Hamelin) was shown useful to trust region and line search variants. The main original contribution is to successfully adapt the line search strategy (Hager and Zhang) for use within trust region like algorithms.

Article

Download

View A note on robust descent in differentiable optimization