fully integrated model only solves in debug mode

[mayork]

unsure of why this is, needs to be solved ASAP. It appears no variables are at the bound and no constants are relaxed.

[bqpd]

What part of debug does it need; the Bounded constraintset or the Relaxed constraintset?

[mayork]

well...at the optimal point nothing is at the bound and the relax values are all 1...

[bqpd]

Right, but we already know that something's going on because it doesn't solve without debug, so if you try just Bounding it instead of using debug we can see if that's the necessary component.

[mayork]

I edited my local debug as shown below and it solves, so I think it only needs relaxed constants right?

        if self.substitutions:
            constsrelaxed = ConstantsRelaxed(self)
            feas = Model(constsrelaxed.relaxvars.prod()**30 * self.cost,
                         constsrelaxed)
            # NOTE: It hasn't yet been seen but might be possible that
            #       the self.cost component above could cause infeasibility
        else:
            feas = Model(self.cost, self)

[bqpd]

Interesting! That would appear to be the case...

[mayork]

It seems like running debug the way I have it configured is like a penalty convex-concave procedure...thoughts? Since relax is 1 at the optimal point I think that I might actually be finding a valid solution we can trust.

@1ozturkbe

[mayork]

woody and ned think the relaxes constants is a valid solution method. I will write a precondition method to edit the model so we don't need to use debug and close the issue after pushing that.

[1ozturkbe]

Coolio, super down to get this issue resolved! Plus, @bqpd the implementation of constraint violation penalties in the solve I expect would resolve the issue where you get in a feasibility solve loop that never resolves. What do you think?

[bqpd]

I think it will 'resolve' that in the sense that it'll converge, but I think the convergent result will have relax variables with a value greater than one.

[1ozturkbe]

It is interesting that we have models that will never break out of the feasibility solve loop and will return UNKNOWN, but will give relax=1 when implemented with penalty functions. So the implementation does matter in this case; wouldn't you agree?

[bqpd]

Yeah, I'm thinking of changing the implementation, if it works for this with RelaxAllConstraints. Could save a bit of time, even.

[mayork]

I think this is solved by the relaxed constants method, closing the issue now but anyone can feel free to reopen if they disagree