`MOI.get` on attribute where `is_set_by_optimizer = true` and `TerminationStatus = OPTIMIZE_NOT_CALLED`

[pedromxavier]

On the series of querying attributes with our meta-solver/reformulator I encountered a corner case when retrieving attributes after reformulation without an inner optimizer.

Basically, we have an operation mode in psrenergy/ToQUBO.jl where a call to optimize! without an optimizer will just reformulate the problem to the QUBO form, e.g.,

model = Model(ToQUBO.Optimizer)
...
optimize!(model)

instead of the usual

model = Model(
    () -> ToQUBO.Optimizer(QuantumSolver.Optimizer)
)
...
optimize!(model) # Returns termination status from `QuantumSolver`

The problem is that the OPTIMIZE_NOT_CALLED status prevents attribute retrieval for cases where is_set_by_optimizer = true.

I couldn't find any other suitable status for this case besides OTHER_ERROR which is kind of misleading since there was no error to begin with.

Note: Using optimize! is not necessary, we could be using something like reformulate! or compile! but this makes no big difference in the issue.

[odow]

I don't know if I fully understand the issue. Do you have a reproducible example? What attribute do you want to query if the problem hasn't been set?

If you don't call optimize!, then by definition you can't query attributes that is_set_by_optimizer = true.

If you want ToQUBO.Optimizer to reformulate the problem on optimize! and make some attributes available, then it should implement MOI.optimize! and return an appropriate termination status. I guess you could use LOCALLY_SOLVED?

[pedromxavier]

I think the question boils down to choosing the appropriate status. The problem I see with LOCALLY_SOLVED is that there will be no solutions available. Is it enough to set ResultCount = 0 and NO_SOLUTION for PrimalStatus and DualStatus?

OTHER_LIMIT seems be a good wildcard for this.

PS: An example would look like:

# Create the model
model = Model(
    # Use ToQUBO in compilation mode
    () -> ToQUBO.Optimizer(nothing)
)

# Add the variables
@variable(model, 0 <= x[1:2] <= 1)
# Add the objective
@objective(model, Min, sum(x .^ 2))

# Add the global constraint
@constraint(model, alpha, sum(x) >= 0.5)

# Compile the QUBO model
optimize!(model)

get_attribute(alpha, ToQUBO.Attributes.ConstraintEncodingPenalty()) # Error!

[odow]

Is it enough to set ResultCount = 0 and NO_SOLUTION for PrimalStatus and DualStatus?

Yes, this should work.

[blegat]

I couldn't find any other suitable status for this case besides OTHER_ERROR which is kind of misleading since there was no error to begin with.

You could also consider that the error is that there was no solver attached. I think it's more important to have the RawStatusString detail clearly what happens so that solution_summary is clear and the user that forgot to give an optimizer is not confused.

[odow]

Is there anything actionable left to do here? This seems like it can be fixed in ToQUBO, not in MOI.

[pedromxavier]

@odow I think this is indeed very specific to our use case, so I don't think it's even something that is missing from the docs. Maybe, for ToQUBO and other reformulators/meta-solvers a REFORMULATION_COMPLETE or REFORMULATION_ONLY status would be the best fit, ideally speaking.

And yes, the question was answered and this issue looks ready to be closed, thanks!

[odow]

I think LOCALLY_SOLVED + NO_RESULT + an informative description in RawStatusString is enough. The solver terminated normally, and it did not find a primal solution.