Closed svigerske closed 5 years ago
Attachment driver.cpp by jwatson created at 2009-12-11 05:44:19
Attachment driver.output by jwatson created at 2009-12-11 05:44:34
Comment by @rlougee created at 2009-12-11 10:51:43
Forwarded to CglTwomir project manager, Oktay Gunluk.
Still a problem with current Cgl/master and CPLEX 12.8.0.0. Could this be related to #36?
Indeed, when using
Index: src/CglTwomir/CglTwomir.cpp
===================================================================
--- src/CglTwomir/CglTwomir.cpp (revision 1463)
+++ src/CglTwomir/CglTwomir.cpp (working copy)
@@ -740,9 +740,9 @@
DGG_setEqualityConstraint(data,j);
/* check if the row is bounded above/below and define variable bounds */
- if ( rowUpper[i] < COIN_DBL_MAX )
+ if ( rowUpper[i] < si->getInfinity() )
DGG_setIsConstraintBoundedAbove(data,j);
- if ( rowLower[i] > -1*COIN_DBL_MAX )
+ if ( rowLower[i] > -si->getInfinity() )
DGG_setIsConstraintBoundedBelow(data,j);
data->lb[j] = 0.0;
then 22 cuts are generated instead of only 3 and the LP relax value only increases to 10576.4.
Issue created by migration from Trac.
Original creator: jwatson
Original creation time: 2009-12-11 05:43:05
Assignee: somebody
Version: release 0.5.1
I have isolated a situation in which I believe CglTwomir is returning invalid cuts. I have attached a test driver (driver.cpp), a test input file (mas74.mps, from miplib2003), and the output trace I obtain on our rather vanilla RedHat linux server.
The “catch” is that I’m using CPLEX 11.2, and the OsiCpxSolverInterface. Although this technically shouldn’t matter, I suppose (the behavior also occurs with other versions of CPLEX).
A quick summary of the unexpected behavior: 1) The LP relaxation for mas74 solves fine – relaxation is around 10K. Validated with various other solvers. 2) 3 two-mir cuts are found and successfully applied. 3) Upon re-solve after cut addition, the LP relaxation is something like 21K, which is far above the integer optimal of ~11K. Obviously incorrect behavior.
Any help is greatly appreciated – I am seeing similar behavior on a handful of other test instances.
Thanks!
Jean-Paul