The strange behavior of the disjunction constraint

[abb-omidi]

@Joao-Dionisio,

Hi,
I hope you are doing well. I tried to analyze the logical capability of the SCIP solver through the PySCIPOpt wrapper to solve a scheduling problem. Also, as an extension of question #1055. The logical constraint I was trying to run is of the form:

$$ (x_{i,m} \land x_{j,m}) \implies (start_{j} \geq finish_{i} \lor start_{i} \geq finish_{j}) $$

It says that if jobs $i$ and $j$ are being assigned to the same machine, they do not overlap. However, it is a series of $\lor$ clauses of the form:

$$ (\lnot x_{i,m}) \lor (\lnot x_{j,m}) \lor (start_{j} \geq finish_{i}) \lor (start_{i} \geq finish_{j}) $$

    for i, j in mapping:
        for m in nb_m:
            model.addConsDisjunction(
                [
                    z[i, m] == 0,
                    z[j, m] == 0,
                    finish[i] <= start[j],
                    finish[j] <= start[i],
                ],
                name=f"disj_{i}_{j}_{m}",
            )

As long as I was running the example with a few tasks/jobs, the model works well without throwing any error, while I extend the jobs, the model solves to optimality with the wrong solution!!!

# The optimal solution
SCIP Status        : problem is solved [optimal solution found]
Solving Time (sec) : 54.66
Solving Nodes      : 729
Primal Bound       : +7.40000000000000e+01 (20 solutions)
Dual Bound         : +7.40000000000000e+01
Gap                : 0.00 %

# The wrong solution
SCIP Status        : problem is solved [optimal solution found]
Solving Time (sec) : 2.06
Solving Nodes      : 2659 (total of 3555 nodes in 2 runs)
Primal Bound       : +8.10000000000000e+01 (1 solutions)
Dual Bound         : +8.10000000000000e+01
Gap                : 0.00 %

I am running scip = 10.0.2, and pyscipopt = 6.2.1.

If you need any further information, just let me know.
All the best
Abbas

PMSP_Logical.txt

[abb-omidi]

Dear @Joao-Dionisio,

A strange observation is that when I set the disjunction constraint attribute 'local' to 'True' the solver works with the normal behavior and every thing goes well.

The constraint itself would be globally valid and I doubt why we need to make it only locally valid!!!

I will check it again to see if I would find any reasonable way to interpret it. Also, I am being glad if I have your feedback.

Best regards