Consider renaming fix_variables function to roof_duality

[arcondello]

Since we're releasing a new package, perhaps we should consider using a more explicit name. The deprecated dimod function can still be called fix_variables.

[hhtong]

I'm good with roof_duality() (would be clear to users that it does the same thing as RoofDualityComposite)

[JoelPasvolsky]

Given that it's a method of roof_duality module, is it worth considering a name that is clearer to users with less familiarity with roof duality? Perhaps something like reduce_variables()?

[randomir]

For the uninitiated, it's much more useful to have a function name that tells them what the function is doing, and not how.

I dare you all, try googling "roof duality" and quickly figuring out what a function in Ocean called roof_duality does.

[hhtong]

I'm wondering if reduce_variables() would be confusing though: I would think that reducing means that it would return the variables that still need to be solved for, but it returns what has already been fixed.

[arcondello]

Well, based on its location, I know it's a preprocessing algorithm for BQMs. There is also a docstring. And there are MANY ways to fix or reduce variables. And likely that we'll want to add some of them in the relatively near future.

I would rather err on the side of explicit.

determine_variable_assigments_with_roofduality works 😆

[JoelPasvolsky]

@arcondello, why would the func name need to specify that it's method is roof duality when method is imported from the roof duality module? You can have many fix_variables and reduce_variables methods that come from different modules and users will know which technique based on the module.

[randomir]

Exactly. Like fix_variables_roofduality, or dwave.preprocessing.fix_variables.roof_duality. Or dwave.preprocessing.roof_duality.fix_variables? 😆

[arcondello]

We could move the function into a more informative namespace?

from dwave.preprocessing.reductions import roof_duality`

[arcondello]

jinx.

[arcondello]

I prefer

library.catagory.algorithm

because it is most easily extensible. So

from dwave.preprocessing.reductions import roof_duality
from dwave.preprocessing.lower_bounds import roof_duality  # this is actually the most accurate but that's another story
from dwave.preprocessing.fix_variables import roof_duality
from dwave.preprocessing.partial_assignments import roof_duality

or something like that?

[amahmudwave]

My opinion

Reduction is too common of a term, reduce_variables is okay though.
partial_assignments --- Roof duality sometimes assigns all the variables.

I prefer the remaining two.

[arcondello]

If we merge this question with #2, then maybe it makes sense to move it into the lower_bounds space

from dwave.preprocessing.lower_bounds import roof_duality

lb, assignment = roof_duality(bqm)

make the lower_bound the more important value. Which is where roof duality actually comes from in the literature.

[arcondello]

Now, does the composite also live there?

# A
from dwave.preprocessing.lower_bounds import RoofDualityComposite
# B
from dwave.preprocessing.lower_bounds.roof_duality import RoofDualityComposite
# C
from dwave.preprocessing.composites import RoofDualityComposite

This is the kind of place that dwavesystems/dwave-ocean-sdk#21 would really pay off. Or a mono package 😄 .

[randomir]

I agree we should shift focus to lower_bounds, for the function.

But the composite is really about partially fixing variables, not finding lower bound, right? So, it probably warrants a clearer name.

That's even more so if we stick it in the dwave.composites.* flat namespace.

[arcondello]

Thinking a bit out of the box, what about

sampler = FixedVariableComposite(ExactSolver())

sampler.sample(bqm, fixed_variables= {'a': -1, 'b': 1})
sampler.sample(bqm, fixed_variables='roof_duality')

or something like that?

We could probably migrate the FixedVariableComposite from dimod while we're at it.

Alternative:

sampler = FixedVariableComposite(ExactSolver(), method='roof_duality')

sampler.sample(bqm, fixed_variables= {'a': -1, 'b': 1})
sampler.sample(bqm)