[Cbc] Cbc returns invalid result on simple problem unless presolve is enabled

[John Forrest]

Giacomo (and Stefan),

I agree that improving the numerical properties would be a good idea.

Allan Wechsler has pointed out to me that there is an integer tolerance 
which could be set to a smaller value.  This can not be set from the 
standalone cbc, but can be from the example drivers - 
model.setIntegerTolerance(1.0e-10).

Just doing that does not not fix the problem, but also turning off 
scaling does.

John Forrest
On 17/11/16 15:17, Giacomo Nannicini wrote:
> Hi John and Stefan,
> I would hardly call this a bug. The model is ill-conditioned
> (coefficients ranging from 10^-2 to 10^7) and it almost seems
> specifically built to 'trick" a solver into problems. The commercial
> solver CPLEX has exactly the same behavior: if preprocessing is
> disabled, it returns a solution with value 1000. So I do not think
> there is much that can be done, other than trying to improve the
> numerical properties of the model.
>
> GiacomoAllan Wechsler
>
> On Thu, Nov 17, 2016 at 6:36 AM, John Forrest
> <john.forrest at fastercoin.com> wrote:
>> Stefan,
>>
>> It is a bug, but I am not sure how to fix.
>>
>> The optimal continuous solution is 0.0,1.0e-9,0.0.  These values are all
>> within integer tolerances so it sets to integers i.e. 0,0,0 and thinks that
>> as the continuous solution is integral then the search is over and fixes
>> integer variable to 0 and leaves others .  However while the continuous
>> objective value was 9.0e-10, the  new solution will be 1.0e-3,0.0,0.0 -
>> giving an objective value of 1000.
>>
>> The fixes that I can think of would affect performance too much and/or be
>> complicated.
>> On 16/11/16 09:09, Stefan Hacker wrote:
>>
>> Hi,
>>
>> I am trying to use Cbc to solve what I thought to be a simple problem.
>> However the behavior I see makes me wonder if I am encountering a Cbc
>> bug or if there is something wrong in the way I am using it.
>>
>> If I run the attached model using Cbc 2.9.7 (version isn't that
>> important I think) from the console like this:
>> cbc model.lp -solve
>>
>> I get the expected objective value of 0.9. However without presolving
>> enabled
>> cbc model.lp -preprocess off -solve
>>
>> Cbc returns an objective value of 1000 which is incorrect.
>>
>> Originally I was using Cbc as a library (close to the minimum.cpp
>> example from the documentation) and didn't consider presolve to be
>> relevant for such a small problem. Activating it feels like I am just
>> suppressing a symptom without understanding or fixing the cause. Even
>> more confusing to me is that after the call to branchAndBound the
>> CbcModel returns 0 for primary and secondary status and
>> isProvenOptimal returns true.
>>
>> Is the result being wrong without presolve a bug or am I trying
>> something stupid here without noticing? Also Cbc returning "optimal"
>> in all ways I know to check but still returning an incorrect result is
>> unsettling. Did I miss some indicator Cbc gave me?
>>
>> Thanks a lot.
>>
>> Regards,
>> Stefan
>>
>>
>>
>> _______________________________________________
>> Cbc mailing list
>> Cbc at list.coin-or.org
>> https://urldefense.proofpoint.com/v2/url?u=http-3A__list.coin-2Dor.org_mailman_listinfo_cbc&d=CwICAg&c=Ngd-ta5yRYsqeUsEDgxhcqsYYY1Xs5ogLxWPA_2Wlc4&r=js2M0T-3OIMIVDvokcKjokJbk0F8QOCd0mT4FsVFE88&m=FlCDUkCFyl6LUVRWa6Mu7ehYk1QuoUvgFPJh_u4AQ-4&s=sSI8pGvVrbw0LJd5V9yjX8yTc8fpKvA0GSiLH77xu2c&e=
>>
>>
>>
>> _______________________________________________
>> Cbc mailing list
>> Cbc at list.coin-or.org
>> https://urldefense.proofpoint.com/v2/url?u=http-3A__list.coin-2Dor.org_mailman_listinfo_cbc&d=CwICAg&c=Ngd-ta5yRYsqeUsEDgxhcqsYYY1Xs5ogLxWPA_2Wlc4&r=8FaVxIx99H879Ebkij38NpcRZoYLK-7QMl1LqyvBccA&m=fJ6JENZYx2k6pJo0f5S4KxtatK98Kji6H7IG4z1t5rk&s=GZXssk61y1TSxW4GAKGZzmDUkWHg__fPQFQgvzZbV1c&e=
>>
>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://list.coin-or.org/pipermail/cbc/attachments/20161118/782fbc45/attachment.html>