WriteMPS vs WriteLP

[aabbas90]

Hi, Thanks for the extensive study!

I just observed that there is some issue in exporting to .mps format at the following location. Specifically the generated ILP is already tight at the LP root relaxation with 0 objective value:

mis-benchmark-framework/solvers/gurobi/main.py

Line 50 in 66211bb

opt_model.writeMPS(args.output / (graph_name_stub + ".mps"))

When I change it to writeLP instead then it is not the case and the LP relaxations are not tight. I tried on SATLIB and ER graphs.

[MaxiBoether]

Hi Ahmad,

Thank you very much for your comment! Please bear with me, it has been quite a while since working on this. If I remember correctly, we used the mps file to run grbtune.

I am certainly no expert in how to best use Gurobi's API. Mainly we followed the Gurobi user guide [1] which states both MPS and LP files can be used. Consulting [2], it seems to me like MPS files are more "precise" than LP files, which probably why we defaulted to them at the time.

Do I understand correctly that the mps file generated by pulp is not correct, compared to the LP file? If so, do you think this is a bug in our usage/framework or within PuLP?

Thank you so much for your help.

[1] https://www.gurobi.com/documentation/9.5/refman/command_line_tuning.html
[2] https://support.gurobi.com/hc/en-us/articles/360013420131-What-are-the-differences-between-LP-and-MPS-file-formats-

[aabbas90]

Ok looks like issue is because when gurobi reads those .mps files it is minimizing the objective instead of maximizing. I guess issue is in pulp or mps format in general.
Reference and solution: https://support.gurobi.com/hc/en-us/articles/360013420371-Why-does-my-MPS-file-produce-different-objectives-with-different-solvers-

Looks like an existing issue with pulp:
coin-or/pulp#459

Overall I think a potential user of this benchmark can have a pitfall if he/she exports the instances as .mps and then uses gurobi to solve them. For me atleast, the problem is resolved :)