Crash Techniquees for Large-Scale Complementarity Problems

Abstract

Most Newton-based solver for complementarity problems converge rapidly to a solution once they are close to the solution point and the correct active set has been found. We discuss the design and implementation of crash techniques that compute a good active set quickly based on projected gradient and projected Newton directions. Computational results obtained using these crash techniques with PATH and SMOOTH, state-of-the-art complementarity solvers, are given, demonstrating in particular the value of the projected Newton technique in this context.