NL writers linear presolver fails when one variable is left about pyomo HOT 4 CLOSED

Thanks for testing this out!

from pyomo.

OK - this is an interesting edge case: you have a model with a non-square subsystem that can be completely presolved out of the model. That is, there is a set of n constraints that act on m (with m>n) variables, and those variables do not appear in any other constraints / objectives. In the original writer (or the new writer with presolve off), those variables and constraints are sent to the solver and the solver will pick (random) values for the m-n degrees of freedom. After presolve, the n constraints are removed from the model - so when the writer collects / categorizes the model variables, none of the m variables are found. This is generally fine - but causes issues when we attempt to find values for the eliminated variables (as the m-n degrees of freedom are not in the var_map).

What do you think the expected behavior should be?

1. leave the variables removed from the model alone (and they will be marked stale)
2. arbitrarily pick a value for the DOF and propagate the values through; but which value to pick?
   a. pick a value as close to 0 as allowed by the bounds
   b. pick the lower (or upper) bound
   c. pick the midpoint between the bounds

I think I would start with proposing 2.a.

from pyomo.

I tried out a few different LP solvers, and their own preprocessing techniques tend to pick a bound -- and this is what an active set method without presolve does too. But if you use an interior point method without presolve you get something closer to 2.c. If the free variable(s) don't have any bounds you seem to get 0.

So I agree that 2.a is the best of the options available. It conforms to existing practice while not being onerous to implement and maintain.

from pyomo.

2.a was the consensus when this was discussed at the Pyomo dev call but we also think this case should log a warning about a modeling issue.

from pyomo.