[Bonmin] Bonmin failing to deliver solution to GAMS (or failing to calculate solution?) in convex MINLP

[Stefan Vigerske]

Hi,

I think I can reproduce your problem with Bonmin 1.4 (gams 23.7)
Bonmin claims to have found improving solutions
  OA0003I New best feasible of -8.9867 found after 10.3964 sec.
  OA0003I New best feasible of -9.40927 found after 15.4497 sec.
and then stops because the gap is closed sufficiently.

Unfortunately, when GAMS evaluates the objective function w.r.t. the 
best solution reported by Bonmin, it has only a value of 18.60159.
(here and above I report the negated objective function values, since 
Bonmin can handle only minimization problems)

However, with a current Bonmin 1.5, it seems to work correctly.
So I guess this bug has been fixed in Bonmin. (Pierre?)
It will also be available in the next GAMS release, which I hope will be 
out soon.

Stefan


>
> Hi,
>
> I'm sending you a slightly modified model. I'm solving it with the
> following options:
>
> bonmin.milp_solver Cplex
> bonmin.allowable_fraction_gap 0.1
> bonmin.algorithm B-QG
>
> - This is the output I get:
>
> Bonmin finished. Found feasible point. Objective function value = -9.40927.
> Resolve with fixed discrete variables to get dual values.
> NLP0014I 4 OPT -9.4093 20 1.963
> Warning: Optimal value of NLP subproblem differ from best MINLP value
> reported b
> y Bonmin.
> Will not replace solution, dual values will not be available.
> Best solution: -1.859887e+001 (41 nodes, 12.074 seconds)
> Best possible: 9.409270e+000 (only reliable for convex models)
>
> Absolute gap: 2.800814e+001 (absolute tolerance optca: 0)
> Relative gap: 297.665415% (relative tolerance optcr: 0.1%)
>
>
> - If I solve with B-OA, I get the following:
>
> Bonmin finished. Found feasible point. Objective function value = -9.40927.
> Resolve with fixed discrete variables to get dual values.
> NLP0014I 3 OPT -9.4093 25 2.488
>
> Best solution: 9.409269e+000 (0 nodes, 16.239 seconds)
> Best possible: 9.409268e+000 (only reliable for convex models)
>
> Absolute gap: 1.460884e-006 (absolute tolerance optca: 0)
> Relative gap: 0.000016% (relative tolerance optcr: 0.1%)
>
>
> Hope this helps,
> Juan
>
>>
>> Hi,
>>
>>> Hi,
>>>
>>>> This is an example output I'm getting from a model solved by bonmin:
>>>>
>>>> Bonmin finished. Found feasible point. Objective function value =
>>>> -8.477.
>>>> Resolve with fixed discrete variables to get dual values.
>>>> NLP0014I 4 OPT -8.477 18 1.887
>>>> Warning: Optimal value of NLP subproblem differ from best MINLP value
>>>> reported by Bonmin.
>>>> Will not replace solution, dual values will not be available.
>>>>
>>>>
>>>> Best solution: -9.701475e+000 (36 nodes, 12.637 seconds)
>>>> Best possible: 8.477002e+000 (only reliable for convex models)
>>>>
>>>> Absolute gap: 1.817848e+001 (absolute tolerance optca: 0)
>>>> Relative gap: 214.444651% (relative tolerance optcr: 0.05%)
>>>>
>>>> However, the primal solution given to GAMS corresponds to objective
>>>> value -9.701475, and not 8.477002.
>>>
>>> The best feasible solution that Bonmin found has value -9.701475.
>>> The 8.477002 is only an upper bound that tells you that there exists
>>> no solution with a value above 8.477002.
>>
>>
>> I thought the output line:
>> "Bonmin finished. Found feasible point. Objective function value =
>> -8.477."
>>
>> meant that bonmin has found a feasible point with objective value
>> 8.477. I don't understand why it subsequently says that the best
>> solution is -9.701475, which is much worse. In fact, I thought bonmin
>> is exact for convex models. The solution reported, however, is way
>> outside the required tolerance levels, as you can see in the output.
>> On the other hand, if I try with a different algorithm (OA), the best
>> solution found *is* 8.477.
>>
>> I'll send you a model later to see if you can reproduce the problem.
>>
>> Juan
>>
>>>
>>> Stefan
>>>
>>>>
>>>> Juan
>>>>
>>>>> Hi,
>>>>>
>>>>>> Hello,
>>>>>>
>>>>>> I'm trying to solver a number of convex MINLP instances of a problem
>>>>>> using Bonmin through GAMS and am frequently getting this error
>>>>>> message
>>>>>> when bonmin finishes:
>>>>>>
>>>>>> Warning: Optimal value of NLP subproblem differ from best MINLP value
>>>>>> reported by Bonmin.
>>>>>> Will not replace solution, dual values will not be available.
>>>>>>
>>>>>> When this happens, GAMS does not receive the last feasible point (and
>>>>>> supposedly optimal solution) found by bonmin. So first of all I'm not
>>>>>> sure if bonmin is correctly calculating the solution in these
>>>>>> cases, and
>>>>>> if it is, it's not sending the solution to GAMS. Is there any way to
>>>>>> solve this or get around it?
>>>>>
>>>>> When Bonmin finishes and has found a feasible solution, then the
>>>>> Gams/Bonmin link solves the NLP obtained from the MINLP by fixing the
>>>>> discrete variables to the values in the solution. It's doing this to
>>>>> compute dual values.
>>>>> The warning says, that this final NLP solve ended with a solution that
>>>>> has a worse objective function value than the solution value reported
>>>>> by Bonmin.
>>>>> However, in this case, you should still get the primal solution that
>>>>> Bonmin reported, only a dual solution will be unavailable.
>>>>>
>>>>> If you don't get Bonmin's solution, it would be great if you could
>>>>> send me a model to reproduce this behaviour.
>>>>>
>>>>> In the next GAMS release, there will be an option to disable the final
>>>>> solve in the Gams/Bonmin link.
>>>>>
>>>>> Stefan
>>>>>
>>>>>
>>>>>>
>>>>>> Regards,
>>>>>> Juan
>>>>>>
>>>>>>
>>>>>> _______________________________________________
>>>>>> Bonmin mailing list
>>>>>> Bonmin at list.coin-or.org
>>>>>> http://list.coin-or.org/mailman/listinfo/bonmin
>>>>>>
>>>>>
>>>>>
>>>>
>>>>
>>>>
>>>
>>>
>>
>>
>>
>> --
>> Juan José Gálvez García, Ph.D.
>> Dept. Ingeniería de la Información y las Comunicaciones
>> Facultad de Informática
>> Universidad de Murcia
>> Campus de Espinardo, 30100 E-mail: jjgalvez at um.es
>> Murcia, SPAIN Phone: +34868887882
>>
>>
>> _______________________________________________
>> Bonmin mailing list
>> Bonmin at list.coin-or.org
>> http://list.coin-or.org/mailman/listinfo/bonmin
>>
>