[Clp] Incorrect solutions from Clp

[John Forrest]

Ted,

2. - as described - is not a bug - you have to say solve or dual or 
similar before -solution $.  Clp is just giving you the solution before 
solving it.

1. probably is a bug

What version of SYMPHONY should I check out to debug this - and how do I 
reproduce bugs?

John


On 18/08/14 17:27, Ted Ralphs wrote:
> Hi all,
>
> I've been puzzling over the fact that Clp sporadically gives incorrect 
> solutions when used within SYMPHONY to solve LP relaxations of an 
> MILP. It's difficult to create a minimal example of the issue because 
> it happens when the problem has already been modified and warm solved 
> a number of times. There are two different seemingly unrelated 
> scenarios in which issues arise:
>
> 1. When doing strong branching with a dual iteration limit, Clp 
> sometimes returns an incorrect objective value. I have debugged this 
> issue and I can see that when the solver first stops in 
> hitMaximumIterations(), the value seems correct, but it is then later 
> re-calculated before the solver exist in computeObjectiveValue() and 
> this computed value is incorrect (if I dump the LP to a file and solve 
> with stand-alone Clp, the corrected value exceeds the reported optimum 
> by a substantial amount, while the uncorrected is a correct lower 
> bound) . It seems to compute this value from a primal solution, though 
> I would think that this would be done from a dual solution in this 
> case. This problem can be fixed, by the way, by solving from a hot 
> start rather than a warm start, which perhaps gives a clue as to what 
> the issue is. I have not been able to create an isolated example yet, 
> though it happens almost immediately when branching in the root node 
> of fast0507 from MIPLIB 3 and can be easily reproduced there within 
> SYMPHONY with a short computational run.
>
> 2. The second problem is that Clp sometimes gets a (very) wrong 
> solution when solving an LP to full optimality. I have managed to 
> isolate one LP relaxation where this can be replicated on the command 
> line with stable version 1.15, revision 2046. If you solve the 
> attached MPS file with
>
> clp danoint1.MPS
>
> the result is
>
> Optimal - objective value 63.062691
> After Postsolve, objective 63.062691, infeasibilities - dual 0 (0), 
> primal 0 (0)
> Optimal objective 63.06269136 - 703 iterations time 0.252, Presolve 0.01
>
> However, solving it with
>
> clp danoint1.MPS -solution $
>
> gives the wrong objective value and the incorrect solution that I am 
> seeing in SYMPHONY:
>
> Model was imported from ./danoint1.MPS in 0.016001 seconds
>
> status unknown
> Objective value               0
>      31 C0000031               1                      0
>      39 C0000039               1                      0
>     505 C0000505              80                      0
>     506 C0000506              90                      0
>     507 C0000507              68                      0
>     508 C0000508              69                      0
>     509 C0000509              83                      0
>     510 C0000510              75                      0
>     511 C0000511              71                      0
>     512 C0000512              66                      0
>     513 C0000513              74                      0
>     514 C0000514              81                      0
>     515 C0000515              63                      0
>     516 C0000516              63                      0
>     517 C0000517              75                      0
>     518 C0000518              70                      0
>     519 C0000519              65                      0
>     520 C0000520              61                      0
>
> Note that Clp says "status unknown" in this case.
>
> John, any idea on either of these two problems?
>
> Cheers,
>
> Ted
> -- 
> Dr. Ted Ralphs
> Associate Professor, Lehigh University
> (610) 628-1280
> ted 'at' lehigh 'dot' edu
> coral.ie.lehigh.edu/~ted <http://coral.ie.lehigh.edu/%7Eted>
>
>
> _______________________________________________
> Clp mailing list
> Clp at list.coin-or.org
> http://list.coin-or.org/mailman/listinfo/clp

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://list.coin-or.org/pipermail/clp/attachments/20140818/80db27cd/attachment.html>