[Scip] Infeasible results in scip, feasible in exported problem

Martin Bergner mbergner at mathematik.tu-darmstadt.de
Fri Jan 29 12:22:01 MET 2010


Hello all,

I get infeasible solutions, again. This time, however, it is at least
independent of the used LP solver.

I have a linear constraint \sum_i x_i = 1, the x_i \in [0, 1] are
continuous but implicitly integer. It doesn't make a difference if I
model them as such or as continuous. The integer value results from
additional constraints which are not initial, but checked, propagated,
separated and global. The explicit upper and lower bound is added to the
variable.

The underlying problem is a routing problem where one path has to be
chosen per order. The set partitioning constraint (modelled as linear
because of the continuity) is the constraint enforcing this choice of
exactly one order. After creating the problem with
SCIPcreateVar/consLinear and adding the constraints and variables, I
solve the problem. After solving, get the best solution with

SCIP_SOL* sol = SCIPgetBestSol(scip);

I get the value of my x variables with:

SCIPgetSolVal(scip, sol, model->getVariableX(i))

Now, I get various forms of infeasible solutions:

1. more than 1 variable X of the order has value 1 (I checked the
exported model, the set partitioning constraint is violated!)

2. one or more of the X variables has a value > 1 (e.g. 5,6,7) which can
not be feasible, as I explicitly set the upper bound of the created X
variable to 1 (in SCIPcreateVar)

Solving the exported Model in cplex or scip provides the correct
value ... .

This problem has appeared only recently but I don't know exactly when.
Attached the exported model, which may not be of much benefit, since it
is solved correctly after importing. 

Regards,
Martin
-------------- next part --------------
A non-text attachment was scrubbed...
Name: problem.lp.gz
Type: application/x-gzip
Size: 18521 bytes
Desc: not available
Url : http://listserv.zib.de/mailman/private/scip/attachments/20100129/9a219dfa/problem.lp.bin


More information about the Scip mailing list