[Scip] Pricer regenerates the same column

[Ambros Gleixner]

Dear Marco,

I think changing numerics/epsilon is not the best idea.  If I understand 
correctly, you are using SCIPisPositive() and friends to check the sign 
of your reduced costs.

Instead you should use the primal and dual feasibility tolerance of the 
LP solver, which is returned by the methods SCIPlpfeastol() and 
SCIPdualfeastol().  (The corresponding parameters are numerics/lpfeastol 
and numerics/dualfeastol.)  I believe you want to compare your reduced 
costs to SCIPdualfeastol().

Changing numerics/epsilon has many other side effects, so I would really 
not recommend it.

ambros


On 05/21/2014 10:41 AM, Marco Casula wrote:
> Dear Gerald,
>
> thanks for the input (also, thanks to Martin for ccing me his answer to
> Cristina).
> In the end, I think it was just a matter of solver tolerance: I am using
> SoPlex 1.7.2, and though I believed its tolerance to be e-12, by means
> of display/lpinfo I noticed a number of tolerance parameters set to
> e-06. Since I am not sure  how to change the value of all of them, I
> lowered scip "numerics/epsilon" from the default e-09 to e-06, and so
> far the problem has not reappeared.
>
> Best,
> Marco
>
>
> 2014-05-16 19:07 GMT+02:00 Gerald Gamrath <gamrath at zib.de
> <mailto:gamrath at zib.de>>:
>
>     Dear Marco,
>
>     sorry for the late reply. What you write sounds quite reasonable and
>     you already tried most things that I would normally aks you for. So
>     I don't see an obvious thing that might have gone wrong, and we
>     probably need to dig a bit more into what happens.
>
>     Therefore, please check the solution status of the LP when your
>     pricer is called by calling SCIPgetLPSolstat(). Are you already in
>     reduced cost pricing when the pricer keeps generating the same
>     column? Also, please enable LP solver output by setting the
>     parameter "display/lpinfo" to TRUE. When you checked the reduced
>     cost of the previously generated variables, they were negative? What
>     is the LP solution value of the "old" variable and what are its bounds?
>
>     Another thing: did you mark all constraints to be modifiable?
>
>     Best,
>     Gerald
>
>     Am 06.05.2014 18:59, schrieb Marco Casula:
>>     Dear all,
>>
>>     I am working on a branch and price algorithm for the VRP with time
>>     windows and stochastic travel times. While on many instances the
>>     algorithm terminates correctly, with a number of
>>     instances/parameters (delay distributions, penalties)
>>     combinations, the algorithm gets stuck as the pricer keeps
>>     regenerating the same column.
>>
>>     I have found a similar case in the mailing list archives, and I
>>     tried every suggestion I found there, without success:
>>     the delay flag of the pricer is set to TRUE, as is the initial
>>     flag for the new variables (I had previously set this to FALSE,
>>     with the same result);
>>     the variables are added via SCIPaddPricedVar;
>>     I checked SCIPvarIsInLP for the variable added in the previous
>>     iteration of the pricer, and it is true;
>>     the reduced cost of the variable added in the previous iteration
>>     (via SCIPgetVarRedCost) coincides with the one i compute for the
>>     variable in the current iteration;
>>     I changed the upper bound on the variables from 1.0 to
>>     SCIPinfinity, with no effect.
>>
>>     I was hoping for some more insight on the matter.
>>     Please find attached a logfile for a non working instance
>>     (terminated early), and the Pricer code.
>>     For completeness, here's a link to a full-working folder
>>     (including all source files)
>>     https://www.dropbox.com/sh/664n10xssiicd4o/G0xtWoLH2T
>>
>>     Thanks for any help
>>     Marco Casula
>>
>>
>>
>>     _______________________________________________
>>     Scip mailing list
>>     Scip at zib.de  <mailto: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
>