[SCIP] creating and checking solution in the original space

[HORVÁTH Markó]

Dear Rolf and Ambros,

It seems it was my sneaky mistake. I set my primal heuristic to 
cloneable, and use inappropriate scip pointers afterward.

Sorry for the false alarm, but thanks for your help!

Best regards,
Marko

2022. 10. 20. 9:23 keltezéssel, Rolf van der Hulst írta:
> Dear Marko,
>
> I noticed a potential point of confusion; you mention that you 'get' 
> the solution using SCIPlinkLPSol().
> However, this function does not set the the values of the variables in 
> the solution, and only semantically 'links' the solution to the 
> current LP solution, so that the solution can be associated with a 
> certain lp solution. Instead, one should use SCIPgetSolVal(), 
> SCIPgetSolVals() or SCIPgetSolVarsData() with a NULL ptr for the 
> SCIP_SOL * argument in order to get the solution values of the current 
> LP solution.
>
> You also mention using SCIPgetSolVal(), but if one uses this on the 
> SCIP_SOL * on which SCIPlinkLPSol() was called and which is not NULL, 
> then the output should be bad because you are not actually working 
> with the LP solution.
>
>
> Best,
>
> Rolf
>
> On 20-10-2022 00:35, Ambros Gleixner wrote:
>> Dear Marko,
>>
>> Sorry for the delayed reply.  From your description I have no good 
>> idea what could go wrong.  Your workflow should be fine. Could it be 
>> that you use SCIP_VAR* pointers for transformed problem variables 
>> when calling SCIPsetSolVal() on the original solution? I suppose you 
>> have run the code in debug mode and no asserts came up?
>>
>> Best,
>> Ambros
>>
>>
>>
>> Am 13.10.22 um 09:31 schrieb HORVÁTH Markó:
>>> Hi all,
>>>
>>> I'm working on a project, where we need to solve a task 
>>> assignment-like problem. Briefly,
>>> - we have x-variables to assign quantities to operators in a 
>>> planning horizon,
>>> - we have y-variables to aggregate certain x-variables (i.e., y = x 
>>> + x + ... + x),
>>> - we have surplus/backlog variables to handle stock level (i.e., 
>>> y(t) + s(t-1) + b(t) = s(t) + b(t-1) + demand(t)),
>>> - and we have a lot of knapsack constraints on x-variables which 
>>> make the problem hard.
>>>
>>> Because on bigger instances SCIP hardly finds an integer feasible 
>>> solution, I'm about to create a primal heuristic method. That is, I 
>>> want to round down the LP values of the x-variables, and adjust the 
>>> other variables to satisfy the aggregation and balance equations.
>>>
>>> I want to work on original variables, because preprocessing may 
>>> change variable bounds, and the created solution would be infeasible 
>>> (and I also should deal with multi-agregated variables, etc.). It is 
>>> not problem for me, if I cannot store the solution in SCIP, but I 
>>> want to check its feasiblity.
>>> 1) I created a primal heuristic class from *scip::ObjHeur*.
>>> 2) I got the current LP solution via *SCIPlinkLPSol*.
>>> 3) I created a new primal solution in the original space via 
>>> *SCIPcreateOrigSol*.
>>> 4) Then, for (original) x-variables I do the rounding: 
>>> *SCIPgetSolVal -> SCIPfloor -> SCIPsetSolVal*. (And for all the 
>>> other variables I also set the adjusted value via SCIPsetSolVal).
>>> 5) Unfortunately (for the second time), *SCIPcheckSolOrig**raises an 
>>> error*: solution violates original bounds of variable <t_x_47_128_2> 
>>> [-0,0] solution value <15>
>>>
>>> But, when I check the solution value of the original/transformed 
>>> variables x_47_128_2/t_x_47_128_2 in the new solution, these are not 
>>> 15, but 0.
>>>
>>> So, could you please help me if my approach is good or not, and if 
>>> yes, what can cause this checking error?
>>>
>>> Best regards,
>>> Marko
>>>
>>>
>>>
>>> _______________________________________________
>>> Scip mailing list
>>> Scip at zib.de
>>> https://listserv.zib.de/mailman/listinfo/scip
>> _______________________________________________
>> Scip mailing list
>> Scip at zib.de
>> https://listserv.zib.de/mailman/listinfo/scip
> _______________________________________________
> Scip mailing list
> Scip at zib.de
> https://listserv.zib.de/mailman/listinfo/scip