Gradient based hyper-parameter optimisation for well conditioned kriging metamodels
- Submitting institution
-
Queen Mary University of London
- Unit of assessment
- 12 - Engineering
- Output identifier
- 595
- Type
- D - Journal article
- DOI
-
10.1007/s00158-016-1626-8
- Title of journal
- Structural and Multidisciplinary Optimization
- Article number
- -
- First page
- 1
- Volume
- 55
- Issue
- 6
- ISSN
- 1615-147X
- Open access status
- Compliant
- Month of publication
- December
- Year of publication
- 2016
- URL
-
-
- Supplementary information
-
-
- 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
-
4
- Research group(s)
-
-
- Proposed double-weighted
- No
- Reserve for an output with double weighting
- No
- Additional information
- Kriging is the most used tool for industrial optimisation problems with computationally expensive simulations (CFD, FEA). Its weakness is poor scalability (currently, low tens of variables), due to poor scalability of kriging hyper-parameter finding, typically by Simulated Annealing or Genetic Algorithm. This fundamental work (funded by EU-FP7-PEOPLE-2012-ITN AMEDEO grant) on deriving adjoint gradients for the condensed likelihood function, condition number, correlation matrix and regularisation parameters enables fast gradient-based hyper-parameter optimisation. Demonstrated by a large-scale aircraft wingbox optimisation with 126 variables, was implemented in Altair’s optimisation software (Dr Royston Jones, royston.jones@uk.altair.com) and led to employment of Ollar and Mortished by Altair.
- Author contribution statement
- -
- Non-English
- No
- English abstract
- -