[Cbc] Simple problem found infeasible... but it's not

[John Forrest]

Marco,

Thanks.  I wish I knew why Gomory cuts have started giving trouble.

If I run with debug I get

Cut with 7 coefficients, cuts off known solutions by 1.403e-05, 
lo=-1.79769e+308, ub=9.75001
( 3 , -9.75 ) ( 4 , -9.75 ) ( 6 , 9.75001 ) ( 7 , 9.75001 ) ( 8 , 
9.75001 ) ( 9 , 9.75001 ) ( 10 , 9.75001 )
Non zero solution values are
( 3 , 1 ) ( 4 , 1 ) ( 7 , 1 ) ( 9 , 1 ) ( 10 , 1 )

so you can see it is an accuracy problem.

Yesterday I had changed stable so that CglGomory cuts were not default, 
but CglGMI was default - as CglGMI is more stable.  I then decided that 
was a bit drastic so changed it back unless a Cbc configure option is set.

So in stable (svn) you now have several options.

In configure use -DSWAP_GOMORY to use CglGMI as default (you can still 
switch back by using gomory ifmove etc.).

In configure use -DWEAKEN_CUTS=1.  This weakens cuts a bit if they look 
slightly odd.

If performance does not suffer, I would suggest first option - or just 
change to using -gomory off -gmi ifmove.

John Forrest

On 24/03/15 08:23, mg wrote:
> Hi All,
> I'm currently solving some very small multiknapsack-like problems and 
> I got an infeasibility on the attached problem (CBC version release 2.9.3)
>
> It seems related to some wrong cut, since disabling the cut generators 
> (with cutsonoff off) it works.
>
> I don't know if it's important, but if you look at the attached LP, 
> you can see a constraint called "PrimaryObjAsConstr" that comes from a 
> previous solve where this constraint was actually the objective function.
> Well, the RHS of this contraint is the objective value of the previous 
> solve (4) plus a tolerance (0.1). If I put it to 4, the model is 
> solved without any problem even with the cut generators enabled. Does 
> this lead to some integer-related stuff ?
>
> For the moment I can solve it by disabling the cut generation, but I 
> wanted to report the issue that might rise again in other cases.
>
> Thanks in advance,
> Marco
>
>
> _______________________________________________
> Cbc mailing list
> Cbc at list.coin-or.org
> http://list.coin-or.org/mailman/listinfo/cbc

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://list.coin-or.org/pipermail/cbc/attachments/20150324/a6b0a5e1/attachment.html>