[SCIP] Pricer keeps generating single-vertex segments for partitioning problem
Robert Schütz
robert.schuetz at stud.uni-heidelberg.de
Thu Jul 13 12:25:54 CEST 2017
Hey all,
I have implemented an image partitioning problem in SCIP, involving a
pricer that generates segments with negative reduced costs.
Each possible segment has a fitting error calculated by the sum of the
fitting errors of each pixel in it and the fitting value of a segment is
given by the color of one "super" node in that segment and therefore
constant when solving the pricing problem.
The pricing problem is of itself a MIP involving constraints on the
connectivity of the generated segment and solved using SCIP.
Now the problem is that after a while, the pricer keeps generating
single-pixel segments consisting only of a super node (which are
therefore added multiple times, even though they should have positive
reduced costs after adding them once).
This happens shortly after I get the following output:
(node 1) solution of LP 43 not optimal (pfeas=1, dfeas=0) -- solving again with tighter feasibility tolerance
34.6s| 1 | 0 | 844 | - |2444k| 0 | 0 | 215 | 56 | 215 | 56 | 0 | 0 | 0 | -- | 5.000000e+04 | Inf
r34.6s| 1 | 0 | 844 | - |2444k| 0 | 0 | 215 | 56 | 215 | 56 | 0 | 0 | 0 | -- | 1.955779e+04 | Inf
and after another round of the pricer:
time | node | left |LP iter|LP it/n| mem |mdpt |frac |vars |cons |cols |rows |cuts |confs|strbr| dualbound | primalbound | gap
36.6s| 1 | 0 | 877 | - |2583k| 0 | 0 | 220 | 56 | 220 | 56 | 0 | 0 | 0 | -- | 1.955779e+04 | Inf
r36.6s| 1 | 0 | 877 | - |2583k| 0 | 0 | 220 | 56 | 220 | 56 | 0 | 0 | 0 | -- | 1.912082e+04 | Inf
After doing some research on the mailing list etc, I tried these two
- SCIPchgVarUbLazy(scip, var, 1.0)
but the behaviour does not change regardless of the segment variables
being binary or contiunous and this setting being set.
- SCIPsetIntParam(scip, "presolving/maxrestarts", 0)
which did obviously not hinder SCIP from solving again (see SCIP's
output above) so I guess this does not count as restarting :)
Do you have an idea what the problem could be?
Yours, Robert
More information about the Scip
mailing list