A smoothing trust-region Newton-CG method for minimax problem (Q928081)

The authors discuss a method for minimax problems by changing the original formulation and solving it in with a standard method for smooth unconstrained optimization. Therefore the underlying nonsmooth problem is written such that only one kink is left which smoothes out for the applicability of a gradient method as it has been done by \textit{I. Zang} [Math. Program. 19, 61--77 (1980; Zbl 0468.90064)]. The smoothing function in the neighborhood of the kink is slightly different. The new problem is solved with a standard trust-region solver and tested on several simple minimax problems. Its performance is compared to other, but not the fastest, algorithms for these problems and appear to be slightly superior. From the runtimes of the compared routines one can conclude that the presented method has the same order in the variables of the underlying optimization problem.