Ticket #2341 (closed enhancement: fixed)
Fine-grain management of evaluations
|Reported by:||wehart||Owned by:||jdsiiro|
Description (last modified by wehart) (diff)
Control evaluations of objectives/constraints/gradients in Coliny solvers at a finer granluarity to better manage the use of expensive/cheap evaluations.
Here's a note from Mike Eldred:
If you want to fully support the application-specific nature of combined vs. separated objective/constraint evals and combined vs. separated val/grad/Hessian evals, then you could go so far as to flexibly support each possibility in your algorithm design (although the separated/separated and combined/combined options seem the most useful) and let the user select the more efficient approach. For DAKOTA, my default would be combined/combined, but there are definitely cases where other settings would be more efficient by avoiding unneeded calculations. This design may be unrealistic if you have lots of algorithms from different sources, but would seem to have good payoff if practical.
Pam Williams is working on supporting exactly this capability within
OPT++ for combined/separated val/grad/Hessians (am internal OPT++ mode
override option). For objective/constraint combinations, OPT++ can't readily support this so we're doing combinations in the wrapper.
- Description modified (diff)
- Component changed from Coliny to Colin
- Priority changed from high to normal
- Version changed from 2.0 to Colin 2.0
- Milestone set to Colin 3.1
- Owner changed from wehart to jdsiiro
- Type changed from defect to enhancement