Enable LP subproblem in julia/runtests.jl

[odow]

x-ref #72

[odow]

So I guess we need some other options?

https://github.com/cvanaret/Uno/actions/runs/11731547352/job/32681840126?pr=77#step:8:23697

[cvanaret]

I fixed some stuff in #82.

You can then try preset=filtersqp and subproblem=LP, which will overwrite the subproblem strategy in the filtersqp preset. We still may have troubles with infeasible instances, I need to fix something there.

[odow]

@cvanaret have you tried to solve problems with HiGHS locally? You might want to cancel the CI run, because it seems to be taking forever.

[odow]

The log is full of runs that look like:

and

[cvanaret]

@cvanaret have you tried to solve problems with HiGHS locally?

Yes, I ran the SLP method with HiGHS on our subset of 429 CUTEst instances (it can solve around 60% - that's the curse of subproblems without curvature...) as well as a few infeasible instances (all solved).
I will add a couple of instances as unit tests to avoid cluttering up the CI.

The log is full of runs that look like:

Can you post the JuMP models of these two instances?

[odow]

I just picked two random screenshots. Here are a few that failed tests. The cvx ones are probably good to start with, because they should have a unique global solution.

https://github.com/jump-dev/MINLPTests.jl/blob/master/src/nlp-cvx-expr/001_011.jl
https://github.com/jump-dev/MINLPTests.jl/blob/master/src/nlp-cvx-expr/002_011.jl
https://github.com/jump-dev/MINLPTests.jl/blob/master/src/nlp-cvx-expr/101_010.jl
https://github.com/jump-dev/MINLPTests.jl/blob/master/src/nlp-cvx-expr/102_012.jl
https://github.com/jump-dev/MINLPTests.jl/blob/master/src/nlp-cvx-expr/102_013.jl

https://github.com/jump-dev/MINLPTests.jl/blob/master/src/nlp-expr/001_010.jl
https://github.com/jump-dev/MINLPTests.jl/blob/master/src/nlp-expr/002_010.jl
https://github.com/jump-dev/MINLPTests.jl/blob/master/src/nlp-expr/003_014.jl
https://github.com/jump-dev/MINLPTests.jl/blob/master/src/nlp-expr/003_015.jl
https://github.com/jump-dev/MINLPTests.jl/blob/master/src/nlp-expr/008_010.jl
https://github.com/jump-dev/MINLPTests.jl/blob/master/src/nlp-expr/009_010.jl
https://github.com/jump-dev/MINLPTests.jl/blob/master/src/nlp-expr/009_011.jl

[cvanaret]

@odow Thanks!
Locally, my trust-region SLP + HiGHS converges on all instances and gets the same result as my trust-region SQP + BQPD (except for nlp-expr/001_010 but the respective minimizers are both global, see jump-dev/MINLPTests.jl#50). For the maximization problem nlp-expr/003_015.jl, both methods find the local maximum (0, 2).

PRs #83 and #84 should have fixed a lot of issues.

[cvanaret]

@odow at least this was a quick fail :D
The problem here is that it's taking the default preset (ipopt) and replacing the subproblem with LP. Therefore it's not using a trust region and subproblems get unbounded. Can you try with the long list of options I wrote in the other message?

[odow]

Ugh. When I rebased on your latest stuff I dropped that commit 😢

[cvanaret]

OK looking at the first fails in the log:

• /tmp/jl_PWcf5p/model.nl: fail is documented in Periodic function in nlp-expr/001_010 implies infinitely many global minimizers jump-dev/MINLPTests.jl#50.
• /tmp/jl_NYSwin/model.nl: local maximum.
• /tmp/jl_buq6o8/model.nl and /tmp/jl_SZ6RRC/model.nl: cannot reduce stationarity (I guess the duals don't settle fast enough). We may have to reduce the tolerance (or find out if there's a bug).

I will look at the complete list and try and figure out what went wrong.

[odow]

cannot reduce stationarity (I guess the duals don't settle fast enough). We may have to reduce the tolerance

Open to different parameter settings

[cvanaret]

@odow I implemented the precedence rules for setting options (#85) and I added a filterslp preset (trust-region filter SLP) with similar options to filtersqp (#88).
Can you rebase and replace the option list with simply preset=filterslp? This should pass more tests.

[cvanaret]

Thanks for looking at the failures. It was a good call to relax the tolerance of the tests.

Here's some feedback:

• nlp_expr/008_010: local solution.
• nlp_cvx_expr/501_011: I haven't tested all 20 dimensions, but the largest instance (n = 20) is solved locally in 3248 iterations (SLP methods are just sloooow), which is over the default limit 2000. So perhaps we can bump it (max_iterations=10000).
• nlp_cvx_expr/501_010: same as nlp_cvx_expr/501_011 but requires fewer iterations.

[odow]

Once CI finishes this is good to go from my end.

[cvanaret]

It's a wrap. Thanks a lot!

[odow]

Merge #78 and then tag [email protected]?

[cvanaret]

Uno 1.3.0 is out!