[SCIP] Too long solving time
Ambros Gleixner
gleixner at zib.de
Fri Aug 17 16:39:40 CEST 2018
PS: Actually, it also works without the trick of changing fpfeastol and
fpopttol, but then you really need to turn off presolving and use least
squares scaling "-s0 -g5".
PPS: Despite the fact that SoPlex can solve it in exact solving mode,
Matthias' comment is still valid: be extra sure that this is the right
and best formulation for your problem. :)
Am 17.08.2018 um 15:21 schrieb Ambros Gleixner:
> Hi Yu,
>
> This is actually a nice example where SoPlex shines with its exact
> rational solving capability. Have a look at
>
> http://soplex.zib.de/doc-4.0.0/html/EXACT.php
>
> for details and the mentioned papers on SoPlex' LP iterative refinement.
>
> However, because of the bad conditioning of your LP, one trick is
> needed: you need to relax the termination tolerance for the intermediate
> floating-point solves. Then SoPlex can even solve your model to zero
> violation with
>
> soplex -s0 -g5 -f0 -o0 -q -c --readmode=1 --solvemode=2
> --int:checkmode=2 --real:fpfeastol=1e-3 --real:fpopttol=1e-3 Simulation.lp
>
> in about one minute. For your termination tolerance of 1e-10 it solved
> in 10 seconds with
>
> soplex -s0 -g5 -f1e-10 -o1e-10 -q -c --readmode=1 --solvemode=2
> --int:checkmode=2 --real:fpfeastol=1e-3 --real:fpopttol=1e-3 Simulation.lp
>
>
> I have not played around with all parameters, so maybe it can be tuned
> even more. But no preprocessing and least squares scaling seem to a
> good idea.
>
> Best,
> Ambros
>
>
>
> Am 17.08.2018 um 12:35 schrieb Yu Chen:
>> Hi Matthias
>>
>> Thanks for your help!
>>
>> Currently I cannot change my model.
>>
>> Besides, the reason why I enforced 1e-10 is that I got another problem
>> before. At that time, the feasible solution was actually not
>> constrained by my model. For example there was an equation ‘x+y=0’,
>> but the solution showed ‘x=1e-6’ and ‘y=0’, and I was suggested to use
>> very low tolerance and then I got like ‘x=1e-6’ ‘y=-1e-6’.
>>
>> Considering this, do you think I can adjust some parameters to save
>> time when running SoPlex instead of changing the model?
>>
>> Best regards
>> Yu
>>
>> Sent from my phone. Excuse the brevity.
>>
>>
>>
>> On Fri, Aug 17, 2018 at 12:00 PM +0200, "Matthias Miltenberger"
>> <miltenberger at zib.de <mailto:miltenberger at zib.de>> wrote:
>>
>> Dear Yu Chen,
>>
>> your model has some numerical issues. The matrix coefficients are in
>> the of 1e-05 to 1e+11. Such a large range can easily cause
>> difficulties and violations in the final result.
>> To make matters worse, you are disabling presolving/simplification
>> (-s0) and enforce very small tolerances (1e-10).
>>
>> Testing other solvers on this instances also yielded warnings and
>> violations due to the bad numerics.
>>
>> My advice: Try to improve the model and verify that your result is
>> actually meaningful with regards to the problem. I don't think that
>> such a small tolerance is really required.
>>
>> Oh and by the way: On my machine the model was solved within a
>> second using the standard tolerances and default settings.
>>
>> Cheers
>> Matthias
>>
>> On 17.08.2018 10:59, Yu Chen wrote:
>>>
>>> Hi,
>>>
>>> I am new to Soplex but I encountered a trouble when using it.
>>>
>>> I made a LP file and used the code in terminal:
>>>
>>> /build/bin/soplex -f1e-10 -o1e-10 -l1e-10 -s0 -x -q -c
>>>
>>> But the Soplex took nearly one day to solve it. Do you know how to
>>> decrease the solving time?
>>>
>>> Best regards,
>>>
>>> Yu
>>>
>>>
>>>
>>> _______________________________________________
>>> Scip mailing list
>>> Scip at zib.de
>>> https://listserv.zib.de/mailman/listinfo/scip
>>
>> -- \__________________
>>
>> Matthias Miltenberger
>> Zuse Institute Berlin
>> Takustr. 7, 14195 Berlin
>> www.zib.de/miltenberger
>> miltenberger at zib.de
>> +49 (30) 841 85-245
>>
>>
>>
>> _______________________________________________
>> Scip mailing list
>> Scip at zib.de
>> https://listserv.zib.de/mailman/listinfo/scip
>>
>
--
Ambros Gleixner, Research Group Mathematical Optimization Methods at
Zuse Institute Berlin, http://www.zib.de/gleixner
More information about the Scip
mailing list