ERGO-Code/HiGHS

Wrongly claimed infeasible model

Closed this issue · 3 comments

apfelix commented

Hi Highs-Team,

I have a model that I'm pretty sure should be feasible, but Highs claims it's infeasible.

Model (I added the .txt extension, so github allows the upload):
highs171_problem.lp.txt

I tried to solve the model via the python interface (not sure if that's relevant). Version highspy==1.7.1_dev1 as well as the version build from the current latest (commit 51dea53f) claim infeasible, while version highspy==1.5.3 gives the following solution (again added the .txt extension, printed using writeSolution and style=1 from version 1.5.3)
solution.sol.txt

I'm pretty sure about version 1.5.3 being right, a) from the way I build the model (I solved for another objective [maximizing] and creating this model as copy with an additional constraint <other objective> >= <other optimal solution value> - tolerance_epsilon, so the solution of the other model should be feasible here as well), and b) since CBC and SCIP agree with version 1.5.3.

PS: Thanks for the great job in general in providing a well-performing freely usable solver 🙂

jajhall commented

I've reproduced the behaviour with latest, and note that with presolve=off HiGHS gets a solution with the optimal objective value of -4052.7422.

There are bugs in presolve to be fixed. I'd just started work on #1710 when your message came in. The fact that v1.5.3 gives the right answer (with presolve) isn't surprising, as bugs have been fixed since v1.5.3, leading to presolve running differently, with the possibility of other bugs being exposed....

#1710 is a MIP with only 5 variables and 3 constraints, so you'll forgive me if that's more attractive than your 806 rows; 544 cols; 2298 nonzeros; 54 integer variables

Thanks for your kind words: HiGHS is so satisfying to have developed

fwesselm commented

This seems to be fixed by #1730 as well.

  • 👍2
apfelix commented

FYI: the problem no longer consists in latest

  • 🚀2