Open
Description
As a follow-up to the QP scaling work, see #1463, we should implement some QP solver tolerances strategies.
I suggest doing the following
- move NLP tolerances to common options
- implement different strategies for QP solver tolerances:
- fixed:
qp_tol_* = additional_qp_tol_factor * nlp_tol_*, similar to what is currently implemented, but more systematic - adaptive scaling based:
qp_tol_ineq = additional_qp_tol_factor * nlp_tol_ineq * min(constraint_scaling_factors) - adaptive based on current residual: Could allow to solve first QPs within QP-based NLP solver to a lower tolerance compared to later ones.
Metadata
Metadata
Assignees
Labels
No labels