ePubs

The open archive for STFC research publications

Full Record Details

Persistent URL http://purl.org/net/epubs/work/66086
Record Status Checked
Record Id 66086
Title A practical dual gradient-projection method for large-scale, strictly-convex quadratic programming
Contributors
Abstract We consider solving a given large-scale strictly convex quadratic program by applying the well-known accelerated gradient-projection method to its dual. While this might seem at first sight to be inadvisable since the dual Hessian is defined in terms of the inverse of the primal one, it turns out that all operations may be performed very efficiently so long as a sparse Cholesky factorization of the primal Hessian may be found. In particular, the gradient-projection part of each iteration requires a sequence of "Cholesky forward solves" with sparse right-hand sides, while the acceleration part may be achieved using, for example, a suitably preconditioned conjugate gradient method. Much use is made of the Fredholm alternative. We illustrate performance of this approach on standard large-scale QP examples, and highlight the methods' use for warm-starting. A new fortran package DQP will shortly be available as part of GALAHAD.
Organisation STFC , SCI-COMP , SCI-COMP-CM
Keywords optimization , quadratic programming
Funding Information
Related Research Object(s):
Licence Information:
Language English (EN)
Type Details URI(s) Local file(s) Year
Presentation Presented at International Conference on Continuous Optimization 2013, Lisbon, Portugal, July 29th - August 1st 2013. talk.pdf 2013