Andreas Grothey and Ken McKinnon
Technical Report MS 98-015
Bundle methods for the minimization of non-smooth functions have been around for almost 20 years. Numerous variations have been proposed. But until very recently they all suffered from the drawback of only linear convergence. The aim of this paper is to show how exploiting an analogy with SQP gives rise to a superlinearly convergent bundle method. Our algorithm features a trust region philosophy and is expected to converge superlinearly even for non-convex problems.
Bundle methods, optimization, SQP
Text (with minor corrections made on 25th June 1999)
Postscript MS 98-015.ps (~550Kb).
Compressed postscript MS 98-015.ps.Z (~140Kb).
G-Zipped postscript MS 98-015.ps.gz (~100Kb).