Worst-case example using linesearch methods for numerical optimization with inexact gradient evaluations

摘要

Two approaches often used to improve the robustness of numerical optimization algorithms are linesearch and trust region methods. Trust region methods have previously been shown to be extremely forgiving of high levels of noise and inaccuracy in gradient evaluations. The authors present a worst-case example demonstrating that linesearch methods can be very fragile with respect to such noise.