MAGMA: Multilevel Accelerated Gradient Mirror Descent Algorithm for Large-Scale Convex Composite Minimization
- Submitting institution
-
Imperial College of Science, Technology and Medicine
- Unit of assessment
- 11 - Computer Science and Informatics
- Output identifier
- 2286
- Type
- D - Journal article
- DOI
-
10.1137/15M104013X
- Title of journal
- SIAM Journal on Imaging Sciences
- Article number
- -
- First page
- 1829
- Volume
- 9
- Issue
- 4
- ISSN
- 1936-4954
- Open access status
- Compliant
- Month of publication
- November
- Year of publication
- 2016
- URL
-
-
- Supplementary information
-
10.1137/15M104013X
- Request cross-referral to
- -
- Output has been delayed by COVID-19
- No
- COVID-19 affected output statement
- -
- Forensic science
- No
- Criminology
- No
- Interdisciplinary
- No
- Number of additional authors
-
2
- Research group(s)
-
-
- Citation count
- 7
- Proposed double-weighted
- No
- Reserve for an output with double weighting
- No
- Additional information
- Multi-level algorithms use models at different fidelity to approximate solutions to optimisation problems. Previously, the best available convergence rate for a first-order multilevel algorithm was O(1/k) (k = number of iterations), but we propose an algorithm whose rate is O(1/k^2), which cannot be improved. The research was presented at an invited session of the 2016 SIAM Imaging Conference, and at an invited lecture at the Isaac Newton Institute Cambridge (UK) as part of "The Mathematics of Machine Learning" theme. It led author Hovhannisyan to co-found a start-up company that exploits such optimisation methods in the construction industry (https://nplan.io).
- Author contribution statement
- -
- Non-English
- No
- English abstract
- -
