Different solutions between Gurobi 7.5.2 and 8.1.1

Answered

Oscar Dowson

• May 28, 2019 19:36

Dear Gurobi,

I have a model that now gets a different solution on Gurobi 8.1.1 compared to Gurobi 7.5.2:

Here is the MPS file

https://gist.github.com/odow/b00626dad00eba2a00bc93e5bcc03647#file-gurobi_bug-mps

The logs for 7.5.2 and 8.1.1 are also in the gist.

Is this a tolerance issue? Expected behavior? Or was a bug fixed?

We're about to submit a paper with results from 7.5.2 and wondering if we have to re-do all the numerical experiments...

Cheers,

Oscar

0

• 

  Oscar Dowson

  • May 28, 2019 19:46

  

  

  

  To follow up, it appears to be an integer feasibility tolerance issue. It seems that v8.1.1 has a tighter tolerance by default, and the "better" solution found by v7.5.2 is actually integer infeasible with a tighter tolerance.

  0
• 

  Daniel Espinoza

  • May 31, 2019 14:56

  Hi Oscar,

   

  Indeed, it seems that it is a tolerance issue, if you tighten primal tolerances in v752 you get the same value as in v811.

  Now, it seems that you are modeling some implication constraints (I see lots of 1e6 * binary in the constraints and very large right hand sides). If accuracy is an issue for you, you might want to try changing the numeric focus, or using implication constraints and let gurobi do the reformulation as a MIP....

  Hope this helps,

  Daniel

  0

Please sign in to leave a comment.