Presolve Refactor and Distributed OBBT #562
Conversation
refactor into two classes; add externally driven options
|
We should add a test to the examples which exercises this functionality. Maybe SSLP? |
There was a problem hiding this comment.
This is pretty great. I had a few minor comments embedded.
You should add a one paragraph section to the bottom of generic_cylinders.rst and in that paragraph you should mention that the presolve does not work with persistent solvers. (aside: I just put in my (long) todo list an item to add a pickle-after-presolve option that would pickle the presolved scenarios and shut down (I think it might be too difficult to allow one solver for presolve and another for the rest))
mpisppy/utils/config.py
Outdated
| @@ -278,6 +278,28 @@ def popular_args(self): | |||
| domain=str, | |||
| default='') | |||
|
|
|||
| def presolve_args(self): | |||
| self.add_to_config("obbt", | |||
| description="Use OBBT presolver", | |||
There was a problem hiding this comment.
I think this description should be changed to "Use the optimization based bounds tightening as part of the presolver" and maybe in all other descriptions, "OBBT" should be replaced with "OBBT/FBBT"
Also, look at this in config.py:
self.add_to_config("presolve",
description="Run the distributed presolver. "
"Currently only does distributed feasibility-based bounds tightening.",
There was a problem hiding this comment.
Other descriptions are OBBT specific, did not make changes to those.
mpisppy/opt/presolve.py
Outdated
| @@ -112,7 +98,7 @@ def presolve(self): | |||
|
|
|||
| # update the bounds if changed | |||
| for sub_n, _, k, s in self.opt.subproblem_scenario_generator(): | |||
| feas_tol = self.subproblem_tighteners[sub_n].config.feasibility_tol | |||
| feas_tol = 1e-8 #self.subproblem_tighteners[sub_n].config.feasibility_tol | |||
There was a problem hiding this comment.
uncool (but you already knew that)
There was a problem hiding this comment.
I added subproblem_tighteners as a property so that we are able to pickup feasibility_tol from the object itself, since SPOBBT uses SPFBBT as an object.
| # TODO: this will create the solver twice, once here and again | ||
| # before the first solve ... need to resolve this issue | ||
| self.solver = obbt_options.get("solver_name", self.opt.options["solver_name"]) | ||
| self.solver_options = obbt_options.get("solver_options") |
There was a problem hiding this comment.
Sigh. Through no fault of this PR, solver options are not what they should be. I'm pretty sure that cfg.max_solver_threads needs to be explicitly picked up here.
There was a problem hiding this comment.
This was done to allow the use of GAMS solver options, which is done differently.
|
I have added documentation for presolve (FBBT and OBBT) in |
This PR refactors the presolve interval tightener into a base class (
_SPIntervalTightenerBase) and two concrete classes:SPFBBTfor FBBT andSPOBBTfor OBBT, thus adding the OBBT functionality to "Distributed Subproblem Presolve" (#373).To enable OBBT, the user needs to use the
--obbtcommand line option in addition to the existing--presolveoption.Leveraging the existing optimization-based bounds tightening (OBBT) available in Pyomo, this presolver will tighten the bounds on the non-anticipative variables by default, however, the user can specify the
--full-obbtcommand line option to run OBBT on all variables.If the non-anticipative variables have different bounds, the bounds among the non-anticipative variables will be synchronized to utilize the tightest available bound.
Further, there are options to specify solver and solver-specific options in
presolve_argsfor the user.