Skip to main content

Curious behaviour in crossover

Answered

Comments

2 comments

  • Silke Horn
    Gurobi Staff Gurobi Staff

    Hi,

    There are some hints of possible numerical issues, but other than that, I see no obvious issues in your logs.

    The matrix range is a little bigger than recommended (and it might get worse during presolve; we don't see that in the logs). Also, some of the RHSs are pretty big. We recommend a range of matrix coefficients of at most seven orders of magnitude (for your model, it is a bit bigger) and maximal objective, bound, and RHS value of at most 1e4 (here, we care about the maximal value not the range).

      Matrix range     [8e-05, 1e+03]
      Objective range  [1e+00, 1e+03]
      Bounds range     [1e-02, 5e+04]
      RHS range        [3e-07, 6e+07]

    There are also warnings about "Sub-optimal termination", but those may be caused by your extremely tight BarConvTol. What happens when you set this to the default or 1e-9 instead?

    Finally, there are warnings about the Markowitz tolerance being tightened. Those may also indicate some sort of numerical issue (but this need not be the case).

    To summarize, to me this is all looking okay. Is there anything else about the solution that concerns you?

    Silke

    1
  • Bruno Colonetti
    Gurobi-versary
    First Question
    First Comment

    Hey, Silke

    There are also warnings about "Sub-optimal termination", but those may be caused by your extremely tight BarConvTol. What happens when you set this to the default or 1e-9 instead?

    A: I chose this tight tolerance exactly to reduce the time spent on the crossover. It works well on average.

     

    To summarize, to me this is all looking okay. Is there anything else about the solution that concerns you?

    Not really. 

     

    I'll try to keep improving the numerics of the model then. The large RHss that you see in the log are really just from a handful of constraints. (The 1e+07 RHS is really from a single constraint, all other RHSs are pretty okay.) 


    Thanks for taking the time, Silke

    Bruno

    0

Please sign in to leave a comment.