[Scip] getting LP dual solutions

[Tobias Achterberg]

One more comment from me:

I have the impression that you want to do something like this:

createmyproblem(scip);
SCIPsetIntParam(scip, "limits/nodes", 1);
SCIPsolve(scip);
accessdualsolutionvalues(scip);


This is not possible. The reason is that SCIP is not an LP solver but a CIP or MIP solver.
Thus, the concept of a dual is not available for a CIP solution.

What you want to do is to access the dual solution vector of the root LP relaxation
(before or after cutting planes). Actually, you might have luck, and the LP solution is
still present in SCIP's internal data structures such that accessing it via
SCIProwGetDualsol() and similar methods is possible, but this is not the way it is
indendet to work.

In order to access the solution of any LP relaxation, you need to install a plugin. For
example, you could add a heuristic plugin with timing SCIP_HEURTIMING_AFTERLPLOOP. There,
you can fetch the dual solution values.


In any case, you need to be aware of presolving. This means, your original proboem
formulation is different from the LP relaxation that is solved by SCIP. Hence, you cannot
map the dual solution of the presolved LP relaxation to your original formulation. Even if
you turn off presolving, the ordering of the rows may change.


Best,

Tobias

Alberto Ceselli wrote:
> I see ... so your suggestion is to work with rows and cols instead of
> constraints and variables whenever possible, and to check the name given
> by the constraint handler just as a last resort (e.g. you have to check
> some property of a linear constraint which depends on the dual value,
> and you really know that the constraint is in the LP), right?
> 
> Ok, many thanks!
> 
> Alberto
> 
> On Thu, 2009-01-22 at 17:13 +0100, Timo Berthold wrote: 
>> Hi Alberto.
>>
>>> Is there a method for getting the dual solution for generic constraints,
>> There is no such method, since there is no matching constraints <=> dual 
>> variable.
>> SCIP is able to handle more general constraints then just linearizable ones, 
>> therefore it may happen that a constraint is not represented in the LP (e.g. 
>> bounddisjunction) or even that there are several rows in the LP for a single 
>> constraint (e.g. AND).
>>
>>> or to get the type of a constraint given its pointer?
>> What you may do to determine the type of a constraint is to ask for the name 
>> of its corresponding cons handler:
>> strcmp(SCIPconshdlrGetName(SCIPconsGetHdlr(cons)), "linear" )) == 0
>>
>>> . Is there a way of getting the value of the dual variable of each constraint 
>> in the last LP solution?
>> However, the method described above is a bit ugly.
>> If you are just interested in the duals of the LP rows, you may better use
>> SCIP_Real SCIProwGetDualsol( ) 
>> The LP rows can be obtained by
>> SCIPgetLPRowsData()
>>
>> Best regards,
>> Kati and Timo
>>
>>
>> Am Donnerstag 22 Januar 2009 12:18:47 schrieb Alberto Ceselli:
>>> Hi all!
>>>
>>> I'm using SCIP as a general purpose solver, that is I read a problem
>>> in .mps format and try to optimize it. Let's say I set a nodes limit and
>>> stop optimization at the root node. Is there a way of getting the value
>>> of the dual variable of each constraint in the last LP solution?
>>>
>>> I've tried to get the vector of problem constraints with
>>>
>>> SCIP_CONS ** conss = SCIPgetConss(scip);
>>>
>>> guessing there's a method like
>>>
>>> SCIPgetDualsol(scip, conss[i])
>>>
>>> but all I've been able to find are methods like
>>>
>>> SCIPgetDualsolLinear(scip, conss[i]) or
>>> SCIPgetDualsolSetppc(scip, conss[i])
>>>
>>> However, I don't know in advance which type the constraint would be.
>>>
>>> Is there a method for getting the dual solution for generic constraints,
>>> or to get the type of a constraint given its pointer?
>>>
>>> Many thanks,
>>>
>>> Alberto
>>>
>>> _______________________________________________
>>> Scip mailing list
>>> Scip at zib.de
>>> http://listserv.zib.de/mailman/listinfo/scip
>>
> 
> _______________________________________________
> Scip mailing list
> Scip at zib.de
> http://listserv.zib.de/mailman/listinfo/scip