[Bonmin] Integer Variables being varied continuously by Ipopt of Bonmin

[Stefan Vigerske]

Hi,

not sure what exactly you mean by "iteration", but probably you just 
look at the solution of a relaxation, not the incumbemt.

Setting the integer tolerance to 1 means that Bonmin can completely 
ignore integrality constraints.

Stefan

On 12/30/2014 01:25 AM, Ramesh, Ujwal (rameshul) wrote:
> Hi All,
>
>
> I have modeled a convex nonlinear problem using Bonmin's c++ interface.
>
>
> The problem also has absolute value constraints of the following format,
>
>
>      abs(x[i]-x[j]) > threshold
>
>
> I have modeled these absolute value constraints using the Big M approach and have introduced integer variables as below:
>
>
>      x[i]-x[j] + M*B > threshold and
>
>      -x[i] +x[j] +M*(1-B) > threshold
>
>
> where M is a large value (based on my problem and not random) and B is the newly introduced binary/integer variable ( Have tried both approaches).
>
>
>
> While using a branch and bound algorithm, I printed the variable values for a particular iteration and see that the integer/binary variables have the values like below.
>
>
>      x[20]= 0.229109
>
>    x[109]= 0.254611
>
>
> due to which the trial point which seems to be the best is being rejected due to in- feasibility  and the algorithm exits stating that my problem is infeasible ( which is obvious if the integer/binary variables are varied continuously)
>
>
>
> I even tried changing the option integer_tolerance as below
>
>
>   bonmin.options()->SetIntegerValue("bonmin.integer_tolerance",1);
>
>
> which by default is 1e-06 but is not respected by the solver.
>
>
> Could anyone comment if I am missing something or using a faulty model.
>
>
> Thanks in advance for your help.
>
>
> I can furnish more details of my problem if required.
>
>
> Thanks and Regards,
>
> Ujwal Ramesh
> Graduate Student,
> University Of Cincinnati
> Ph: +1-513-302-4771
>
>
>
> _______________________________________________
> Bonmin mailing list
> Bonmin at list.coin-or.org
> http://list.coin-or.org/mailman/listinfo/bonmin
>