Strange crossover behavior
Awaiting user input
Dear colleagues,
I'm solving the root relaxation of my MILP model. Since the model has a large number of rows and columns (about 2-E6 and 2-E6, resp.) and is rather sparse (less than 8-E6 nonzeroes), we use the barrier algorithm + crossover (Method=2 Crossover=4)
The model is solved by the barrier method very efficiently (about 100 sec on my desktop), the first and the second phases are passed by the crossover quickly as well.
But, in the third phase, the algorithm behaves very strangely (see the screenshot enclosed). My main concern is about the first line of the Root simplex log. As it follows from the listing, at iteration 1149261 (710s), both the primal and the dual infeasibilities are vanishing. Therefore, this solution should be optimal. Why the algorithm did not stop here? Perhaps, I should set up some additional parameters?
With best regards,
Michael
-
Hi Michael,
At a first glance, this looks like numerical issues may be present. Could you share this model, cf. Posting to the Community Forum? If not, could you share the coefficient statistics of your model, i.e., the first 20-30 lines of the Gurobi Log. You could try to experiment with the NumericFocus parameter at see whether it improves the behavior. You should also try setting Crossover=-1 to check the automatic setting. It might be better to to do a primal instead of a dual clean up.
Best regards,
Jaromił0
Please sign in to leave a comment.
Comments
1 comment