Background:
I am solving a non-convex MINLP with SCIP. It contains about 100 binary variables, 8000 constraints and 8000 variables. I am initializing the problem with a heuristic based on "engineering insight". The initial point is feasible and probably close to the optimal solution.
Problem:
The optimality gap starts around 75%. Unfortunately SCIP stalls out^ with a gap around 60% after 5 minutes. With display/verblevel = 5
in scip.set
, I see the following message repeated many times (with different LP numbers):
(node 75) solution of LP 32563 not optimal (pfeas=1, dfeas=0) -- solving again with tighter feasibility tolerance
I recompiled SCIP with CPLEX as the linear solver. This didn't help.
Questions:
Any recommendations for improving the LP solver robustness? More specifically,
- There are probably parts of the model that are poorly scaled. Is there any diagnostic information I can use from SCIP (or the LP solver) to track down these scaling issues?
- Does anyone have recommendations for different solver tolerances to try?
- Is this a red herring, and is the lack of progress from SCIP likely do to something else.
Thank you, Alex.
^ By "stalls out" I mean that SCIP does not print any additional information with display/verblevel = 4
.
Edit: Fix typo.