[Scip] Fwd: Fwd: stop condition of pricing routine

[Mattia Barbieri]

Hi Gerald, hi all.
 Did you found any solution to the stabilization issue? I tried the method
mentioned above, the convex combination with previous dual solutions, and it
seems to work fine for certain type of istances (the problem that was very
difficult to solve still remain difficult, and no improvement is achieved).
Hence I would like to try the method with the 5-piecewise linear penalty
function. This should be implemented using a modified relaxator (as we
discussed previously), but create and solve a new LP at each pricing round
seems to be an enormous waste of time. How can I get information from the
last LP to warmstart the new one?

  Thanks a lot,
  best regards

On 29 July 2010 15:08, Gerald Gamrath <gamrath at zib.de> wrote:

>  Hi Mattia,
>
> sorry about the idea with the probing node, I think, this won't work either
> because variables added at a probing node remain in the problem. Besides,
> you would again increase the feasible region in the probing node (adding new
> primal variables which would not neccessarily have been 0 at previous
> nodes).
>
> I think there might be another possibility, but I have to think about it
> some more lest I again propose some idea that will not work.
>
> Best regards,
> Gerald
>
> Am 28.07.2010 20:50, schrieb Mattia Barbieri:
>
> Technically speaking, can you give me more details about the probing mode?
> I saw in the documentation methods like SCIPstartProbing()
> [SCIPendProbing()], SCIPsolveProbingLP(). But how add new variables and
> constraints to this LP? Just SCIPaddVar and SCIPaddCons? Do they affect only
> the current opened probing node? I mean, when I end probing, the variables
> and constraints added since probing start are neglected?
> The order should be approximately this:
>
> REPEAT {
>   SCIPstartProbing()
>   SCIPnewProbingNode()
>
>  <- here add penalty terms and relatives constraints, according to the last
> dual values of the unstabilized-LP (i.e. re-center the trust-region)->
>  <- what about the "extreme" solution of the unstabilized-LP?. Is this not
> already included?->
>
>  <-take the dual values of this stabilized LP->
>
>  SCIPbacktrackProbing(SCIPgetProbingDepth -1?)
>  SCIPendProbing()
>
>  <-find new priced variable with above dual values->
> }
> UNTIL no priced var found AND trust-region doesn't change.
>
>  Thanks in advance.
>
> On 28 July 2010 18:45, Gerald Gamrath <gamrath at zib.de> wrote:
>
>>  Hi Mattia,
>>
>> Am 28.07.2010 16:51, schrieb Mattia Barbieri:
>>
>>
>> Also this idea sounds good. Perhaps the overhead introduced by the LP
>> solving is too heavy, but this has to be tested experimentally. If I
>> understood correctly, the primal solution given by the LP solver will be
>> used to start the LP probing node, and this hopefully would speed up the
>> solving process. Is that right?
>>
>>>
>>>     yes, this should be right. At least, if you only add dual
>> constraints to the LP for the stabilization and no primal constraints/dual
>> variables. Then the current basis should stay primal feasible and SCIP would
>> try to use the primal simplex. If you add both dual constraints as well as
>> dual variables, then you could perhaps first add the dual variables,
>> reoptimze with the dual simplex, then add the dual constraints and
>> reoptimize with the primal simplex. However, I think you would have similar
>> problems even if you would implement your own stabilized B&P-Solver, so I do
>> not think that you loose much by doing it this way in SCIP.
>>
>> Best regards,
>> Gerald
>>
>> _______________________________________________
>> Scip mailing list
>> Scip at zib.de
>> http://listserv.zib.de/mailman/listinfo/scip
>>
>>
>
>
> --
> Mattia Barbieri
> barbieri.mattia at gmail.com
>
>
>


-- 
Mattia Barbieri
barbieri.mattia at gmail.com
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://listserv.zib.de/mailman/private/scip/attachments/20100803/cb8f9b06/attachment.html