GDP rewrite #354

jsiirola · 2018-02-16T19:56:49Z

Fixes #318, #241

Summary/Motivation:

This is a rewrite of the pyomo.gdp package.

Changes proposed in this PR:

Rewrite of the Disjunct and Disjunction components to bring them in line with current Component conventions.
Adding more examples
Rewrites of the bigm and chull transformations
Adding a cuttingplane reformulation
Adding GDPopt solver (to pyomo.contrib)

Legal Acknowledgement

By contributing to this software project, I agree to the following terms and conditions for my contribution:

I agree my contributions are submitted under the BSD license.
I represent I am authorized to make the contributions and grant the license. If my employer has rights to intellectual property that includes these contributions, I represent that I have received permission to make contributions and grant the required license on behalf of that employer.

…tionary

…variables like in gams

…d troubles), fixes variables in stickies to validate model, changes add slacks transformation to modify the body of the constraint directly

…esults from first layout on minlp.org (and it's rightgit add stickies.pygit add stickies.py)

… am not sure they are right yet.

… begins to add bounds in the batch processing model

…s debugging stuff in add slacks

…ing model, and notes Francisco's suggestion for second term of disjunction for whether or not there is a storage tank (but no conclusion yet)

…essions by digging into the _SumExpression objects

…ction components

…ession tests pass, adding unit test file for bigm

…ew structure of transformed model

…ansformation block and xor constraints added to parent block and using original indicator variables

…ion tests

…tions, indexed constraints in the disjunction)

GDPopt solver initial implementation

Substituting he disaggregated variables into a constraint should preserve the "equality nature" of the constraint. In addition, catch the special case of "x == 0" in a disjunct so we can fix the disaggregated variable to 0.

…ature.

This supports declaring "empty" disjunctions; fixes #241.

Refactor the gdp.cuttingplane transformation - avoid the use of CUIDs - improved initialization of separation problems - graceful exit when subproblem solves fail

Most of the changes are due to: - equality constraints are no longer split into inequalities - BigM constraints on disaggregated vars that are redundant to bounds are omitted.

qtothec · 2018-02-17T05:40:40Z

pyomo/gdp/plugins/bigm.py

+        return suffix_list
+
+    def _apply_to(self, instance, targets=None, **kwds):
+        config = self.CONFIG().set_value(kwds.pop('options', {}))


I'm very confused how this is passing tests, because on my system, set_value() has no return value (so it's always None), which causes config = None, which causes issues when you try to do config.set_value(kwds) below.

You need the current pyutilib master.

Ah, yes. That seems to have done the trick.

qtothec · 2018-02-17T05:42:12Z

pyomo/gdp/plugins/bigm.py

+    CONFIG = ConfigBlock("gdp.bigm")
+    CONFIG.declare('bigM', ConfigValue(
+        default=None,
+        domain=_to_dict,


What does domain mean with respect to a ConfigBlock?

Note that the domain is passed to the ConfigValue constructor. domain is a callable object (could be a type, function, or functor). When a value is set, the incoming value is passed as the single argument to __call__ and the return value is what is stored in the ConfigValue. At it's simplest, it can be a type like int so that all incoming values are automatically cast to int. In this case it is a little more complicated, as it accepts dicts or "values" (int/float/tuple). The former is preserved and the latter is cast to {None: val}.

To answer your question, ConfigBlocks can not have domains, although they accept an implicit_domain, which is used when the ConfigBlock accepts undeclared (implicit) keys - in which case, any implicit values are passed through the implicit_domain callable.

qtothec

A few questions.
Also, I should note that bigM doesn't currently support big-E Expression objects attached to disjuncts. I assume that chull does not as well.

I got tired of looking through all the examples.

qtothec · 2018-02-27T17:28:51Z

examples/gdp/ConstrainedLayout.py

+
+# bounds for box coordinates
+# TODO: What are they trying to accomplish here? Why is it always index
+# 2 for the first two??


This is an error in the original GAMS file. The intent is to find the min and max x and y coordinates that the boxes can have, based on the circle locations and radii. The code as written works because circle 2 has the higher x and y coordinate values.

qtothec · 2018-02-27T17:30:29Z

examples/gdp/ConstrainedLayout.py

+def no_overlap_rule(model, i, j):
+    return [model.no_overlap_disjunct[i, j, direction] for direction in \
+            model.BoxRelations]
+model.no_overlap_disjunction = Disjunction(model.BoxPairs, rule=no_overlap_rule)


Are we going to support the shortcut notation?

@model.Disjunction(model.BoxPairs) def no_overlap_rule(model, i, j): ...

Similarly with Disjunct.

Yes. That notation should (already) be supported.

qtothec · 2018-02-27T17:33:25Z