2017. May 25. 14:15
H building, room 306
Department of differential equations
Lecturers: Philipp Hungerlaender and Franz Rendl Alpen-Adria Universitaet Klagenfurt
The semismooth Newton method of Kunisch et al for bound constrained convex quadratic programming is extremely efficient, if it converges.Unfortunately, global convergence may fail in general.
We first present two variants to make it globally convergent, one uses recursion, the other a type of combinatorial line search. Both variants maintain the positive features of the SN-Method, and there does not seem to be a clear champion among the two.
Finally, we address modifications to make the SN-Method applicable to general convex quadratic problems, including linear equality constraints. First computational experiments look very encouraging.