SetCoverGenerator keeps generating infeasible instances.

[yuandong-tian]

Minimal program:

from pyscipopt import Model
import ecole

generator = ecole.instance.SetCoverGenerator(n_rows=500, n_cols=1000, density=0.05)
m = next(generator)
model = m.as_pyscipopt()
print("Infinity value: ", model.infinity())
print("SCIP version: ", model.version())
print("Ecole version: ", ecole.__version__)

model.presolve()
print(model.getStatus())
model.optimize()
print(model.getStatus())
print(model.getPrimalbound(), model.getDualbound())

In Ecole 0.8 + pyscipopt 4.0 + SCIP 8.0, it gives

Infinity value:  1e+20
SCIP version:  8.0
Ecole version:  0.8.0
infeasible
infeasible
1e+20 1e+20

In Ecole 0.7.3 + SCIP 7.0.3 + pyscipopt 3.2.2, it gives

Infinity value:  1e+20
SCIP version:  7.02
Ecole version:  0.7.3
unknown
optimal
124.0 124.0

Saving to .cip shows that there is a weird constraint in the generated program specification.

  [linear] <c_490>: 1 <= <x_3>[B] +<x_24>[B] +<x_35>[B] +<x_37>[B] +<x_39>[B] +<x_40>[B] +<x_42>[B] +<x_87>[B] +<x_98>[B] +<x_105>[B] +<x_123>[B] +<x_127>[B] +<x_143>[B] +<x_236>[B] +<x_255>[B] +<x_297>[B] +<x_303>[B] +<x_311>[B] +<x_315>[B] +<x_316>[B] +<x_335>[B] +<x_400>[B] +<x_410>[B] +<x_419>[B] +<x_435>[B] +<x_452>[B] +<x_460>[B] +<x_489>[B] +<x_496>[B] +<x_629>[B] +<x_639>[B] +<x_667>[B] +<x_684>[B] +<x_716>[B] +<x_728>[B] +<x_730>[B] +<x_734>[B] +<x_798>[B] +<x_838>[B] +<x_860>[B] +<x_889>[B] +<x_907>[B] +<x_930>[B] +<x_941>[B] +<x_954>[B] <= 2.12199579145934e-314;

Maybe SCIP Infinity is converted to 2.12199579145934e-314 (or a memory overflow error).

[AntoinePrv]

Hi @yuandong-tian,

We did make a change in the SetCoverGenrator in version 0.8.0 in #301, but it should only change the objective, not the constraints. It would be worth testing this with Ecole 0.8.0 and Scip 7

Can you reproduce you example with a fixed seed (.i.e making it deterministic for us).

[yuandong-tian]

Thanks @AntoinePrv for answering. It doesn't seem to be related to random seed. So any seed should just have the same issues. Also how to install Ecole 0.8.0 with SCIP 7?

[AntoinePrv]

@yuandong-tian, using Ecole 0.8 + pyscipopt 4.0 + SCIP 8.0, and running your example about ten times, I only got feasible instances (reading optimal).

[gasse]

Hi all,

@yuandong-tian I don't think you can install Ecole 0.8 with SCIP 7. I think @AntoinePrv meant SCIP 8.

I've tried also your minimal program on my side with both Ecole 0.7.3 and Ecole 0.8.0, I can't reproduce the issue. Given the weird numerical value you get in the constraint, my bet would be for a numerical rounding that occurred somewhere during the instance generation process. That's annoying, but since we can't reproduce it's hard to look into it. @yuandong-tian could you give us additional info about the machine you ran the code on ? Maybe also the python version, the version of the numpy library etc. ? If you're using conda, a conda list should give you that.

Best, Maxime