High-level interface

[mtanneau]

(I'm trying out this new GH feature)
cc @joehuchette @frapac

Q: What should the high-level interface look like?

Some directions:

• Having an interface to MOI
• How will the user interact with pre/post crush?
• What low-level components do we want to expose?

[joehuchette]

I would prefer an MOI interface. I would propose that we: keep the existing interface, suitably altered to support IPs and/or to cater to the needs of folks who intend to use it directly. Then, we can build an MOI wrapper directly on top of this.

Why would a user want to interact with the crushes?

Also, having an interface to configure the presolve routines is a priority for me.

Also, cc @dpo

[mtanneau]

Not everybody works with MOI, including me within Tulip.
I think @dpo mentioned he would be interested in a lower-level interface.

Why would a user want to interact with the crushes?

"interact with crushes" may be as simple as having a precrush and postcrush function.
The question is more about their signature, whether we separate primal & dual crush, etc.
In any case, how the presolved problem is solved (if solved at all) is up to the user. IMO, so should be interacting with crushes.

[mtanneau]

For starters:

MOI interface

I'm thinking we could have a thin wrapper, which would do the following

1. Extract problem data from original MOI model
2. Preform presolve
3. Instantiate a presolved_model::MOI.ModelLike that holds the presolved model
4. Return presolved_model along with pre- and post-crush functions

The user side may look like

my_model = generate_some_model()

# Perform presolve; a new MOI.ModelLike is returned
presolved_model, precrush, postcrush = MOP.presolve(my_model)

# Copy presolve_model into an actual optimizer
my_optimizer = XXX.Optimizer()
MOI.optimize!(my_optimizer, presolved_model)

# Perform post-crush: open question

Some remarks:

• Calling MOI.optimize! on presolved_model should error, because in general it does not have an optimizer attached.
  We could relax this, e.g., when an OptimizerWithAttributes is provided:

  presolved_model, precrush, postcrush = MOP.presolve(my_model, MOI.OptimizerWithAttributes(Gurobi.Optimizer))
  MOI.optimize!(presolved_model)  # solve with Gurobi directly

• Pre- and post-crush is performed by the user. I should be possible to have precrush and postcrush work with MOI indices directly, to make things more convenient. Note that precrush and postcrush should operate on the full solution vector.

• If my_model is modified, then the presolved model is no longer valid.