search

POptUS / IBCDFO

Interpolation-Based Composite Derivative-Free Optimization
Other
7 stars 2 forks source link

Challenging interpolation sets. #166

Open wildsm opened 2 weeks ago

wildsm commented 2 weeks ago

In rare cases, after a subproblem solution has been evaluated, pounders will produce an interpolation set with with fewer than n+1 points. This is because the order in which points are examined for sufficient affine independence is reverse lexicographic.

A test case illustrating the issue on the main branch is in #149.

This is not a bug in formquad, but an issue with pounders assuming that, since the trust region radius and center have not changed and the model is not valid, a model-improving point will be evaluated; after no more than n-1 such iterations, the model will be valid and progress will be made (either through accepting a solution in the mean time, or having a valid model, which then allows the radius to shrink).

Multiple fixes are given below.

wildsm commented 2 weeks ago

One simple fix is shown here: https://github.com/POptUS/IBCDFO/compare/main...bugfix/safeguarding_Mind_1

wildsm commented 2 weeks ago

The alternative version is here: https://github.com/POptUS/IBCDFO/compare/main...bugfix/safeguarding_Mind_2