Constraint aliasing #743

robfalck · 2022-04-25T14:59:52Z

Summary

Implementation of Dymos without separate pass-thru components for path and boundary constraints.

Boundary and Path constraints are all nonlinear by default. This works around an issue with the way OpenMDAO and pyOptSparse currently interact with linear constraints. Users may override the linearity of a constraint by passing linear=True to add_boundary_constraint or add_path_constraint, but if the constrained output is actually a nonlinear function of the design variables this will result in a failure of the optimization.

Adding multiple initial, final, or path constraints to the same index of a variable will now result in an exception. If possible, this condition is detected in add_boundary_constraint or add_path_constraint. If different but equivalent indices are specified on the constraint then this situation is detected during configure.

Outputs which are non-scalar can now have the constrained elements passed by slice objects.

Related Issues

Backwards incompatibilities

In certain cases, treating a constraint which had been marked linear as a nonlinear constraint may pose convergence issues. Please provide feedback if you encounter such cases.

New Dependencies

None

…t_radau

…raint_aliasing

Dymos now raises an error if a parameter is subject to more than one boundary or path constraint.

…aint_src is now _get_objective_src.

… indices

…undary constraints, but getting closer to being like numpy

…raint_aliasing

The following tests failed: test_cannonball_matrix_state.py:TestCannonballMatrixStateExplicitShape.test_cannonball_matrix_state_gl test_cannonball_matrix_state.py:TestCannonballMatrixStateExplicitShape.test_cannonball_matrix_state_gl_solve_segments test_cannonball_matrix_state.py:TestCannonballMatrixStateExplicitShape.test_cannonball_matrix_state_radau test_check_partials.py:TestCheckPartials.test_check_partials_no test_check_partials.py:TestCheckPartials.test_check_partials_yes test_explicit_shooting.py:TestExplicitShooting.test_brachistochrone_explicit_shooting test_explicit_shooting.py:TestExplicitShooting.test_brachistochrone_explicit_shooting_path_constraint test_explicit_shooting.py:TestExplicitShooting.test_brachistochrone_explicit_shooting_path_constraint_invalid_renamed test_explicit_shooting.py:TestExplicitShooting.test_brachistochrone_explicit_shooting_path_constraint_polynomial_control test_explicit_shooting.py:TestExplicitShooting.test_brachistochrone_explicit_shooting_path_constraint_renamed test_explicit_shooting.py:TestExplicitShooting.test_brachistochrone_explicit_shooting_polynomial_control test_explicit_shooting.py:TestExplicitShooting.test_brachistochrone_static_gravity_explicit_shooting test_explicit_shooting.py:TestExplicitShooting.test_explicit_shooting_timeseries_ode_output test_jupyter_output_linting.py:LintJupyterOutputsTestCase.test_header test_jupyter_output_linting.py:LintJupyterOutputsTestCase.test_output test_lint_docstrings.py:DocstringLintTestCase.test_docstrings test_multi_phase_restart.py:TestExampleTwoBurnOrbitRaiseConnected.test_ex_two_burn_orbit_raise_connected test_phase.py:TestPhaseBase.test_parameter_final_boundary_constraint test_phase.py:TestPhaseBase.test_parameter_initial_boundary_constraint test_phase.py:TestPhaseBase.test_parameter_path_constraint test_run_problem.py:TestRunProblem.test_run_brachistochrone_problem_refine_case_driver_case_prefix test_run_problem.py:TestRunProblem.test_run_brachistochrone_problem_refine_case_prefix test_time_targets.py:TestPhaseTimeTargets.test_explicit_shooting

Allow user to override linear setting of a constraint. Constraints may always be nonlinear. Those that introspection finds definitely may not be linear will result in an exception if set to linear.

test_refine_with_shaped_parameter.py:TestRefineShapedStaticParam.test_refine_shaped_static_param_gl test_refine_with_shaped_parameter.py:TestRefineShapedStaticParam.test_refine_shaped_static_param_radau test_two_burn_orbit_raise_linkages.py:TestTwoBurnOrbitRaiseLinkages.test_two_burn_orbit_raise_link_control_to_param

…raint_aliasing

…e treated as linear is complex and requires a lot of instrospection code. Combined with recent issues with linear constraints discovered with the interface between OpenMDAO and pyoptsparse, this makes it far simpler and less error-prone to simply make all boundary and path constraints nonlinear unless overridden by the user. TBD: Add some documentation about constraint aliases and the linearity of constraints before merging this one in.

…nstraints.

review-notebook-app · 2022-04-25T14:59:57Z

Check out this pull request on

See visual diffs & provide feedback on Jupyter Notebooks.

Powered by ReviewNB

coveralls · 2022-04-26T23:15:55Z

Coverage decreased (-0.1%) to 95.562% when pulling c6f87d8 on robfalck:constraint_aliasing into f5aa3cb on OpenMDAO:master.

Internally dymos now converts all constraint indices to their equivalent positive values to make this detection more consistent.

…int involves the same parameter

naylor-b · 2022-04-29T17:35:12Z

dymos/phase/phase.py

+        bc_list = self._initial_boundary_constraints if loc == 'initial' else self._final_boundary_constraints
+
+        existing_bc = [bc for bc in bc_list
+                       if bc['name'] == name and bc['indices'] == indices and bc['flat_indices'] == flat_indices]


Does bc['indices'] == indices work if they're both ndarrays of equal size? I think that will give ValueError: The truth value of an array with more than one element is ambiguous. Use a.any() or a.all()

Good catch. Changed this to catch the common case where the names are the same and indices are both None.
More complicated cases involving overlap of specific indices are worked out during configure.

naylor-b · 2022-04-29T17:40:53Z

dymos/phase/phase.py

+
+        all_flat_idxs = set()
+
+        for con in cons:


Could avoid creating lists you don't need by just setting cons equal to self._initial_boundary_constraints, self._final_boundary_constraints, or self._path_constraints and then in your main loop here just continue if con['name'] == name.

naylor-b · 2022-04-29T17:46:27Z

dymos/phase/phase.py

+        all_flat_idxs = set()
+
+        for con in cons:
+            flat_idxs = set(get_constraint_flat_idxs(con).tolist())


Since you're passing flat_idxs to intersection it doesn't have to be set itself, just an iterator. So if get_constraint_flat_idxs(con) returns a flat array, you could probably pass that directly to intersection.

naylor-b · 2022-04-29T18:02:58Z

dymos/transcriptions/transcription_base.py

-                src_idxs = np.arange(-size, 0, dtype=int).reshape(shape)
+                # This is a path constraint.
+                # Remove any flat indices involved in an initial constraint from the path constraint
+                idxs_not_in_initial = list(set(flat_idxs.tolist()) - idxs_in_initial)


Not sure how big flat_idxs can get, but you compute list(set(flat_idxs.tolist()) twice. Also if flat_idxs is already a flat array you probably don't need the .tolist() either.

naylor-b · 2022-04-29T18:06:13Z

dymos/utils/indexing.py

+        The flat indices of a constraint at a single point in time.
+    """
+    if con['indices'] is None:
+        flat_idxs = np.arange(np.prod(con['shape'], dtype=int), dtype=int).ravel()


Doesn't seem like we need a call to .ravel() here since the array should already be flat.

naylor-b · 2022-04-29T18:09:12Z

dymos/utils/indexing.py

+    else:
+        flat_idxs = indexer(con['indices'], src_shape=con['shape'], flat_src=con['flat_indices']).as_array()
+
+    # Convert any negative indices to positive indices, to make detecting overlaps easier.


All of this code to get rid of negative indices is unnecessary if you call .shaped_array() instead of .as_array() above.

Kenneth-T-Moore · 2022-04-29T18:17:23Z

dymos/examples/brachistochrone/test/test_duplicate_constraints.py

+
+        phase.add_parameter('g', targets=['g'], units='m/s**2', opt=True)
+
+        # Now mistakenly add a boundary constraint at the same loc, variable, and different but equivalent_negative indices


Comment here doesn't reflect the test.

Kenneth-T-Moore · 2022-04-29T18:31:49Z

dymos/examples/brachistochrone/test/test_duplicate_constraints.py

+        expected = ("In phase traj0.phases.phase0, parameter `g` is subject to multiple boundary or path constraints.\n"
+                    "Parameters are single values that do not change in time, and may only be used in a single "
+                    "boundary or path constraint.")
+


This error message isn't part of this PR, but it doesn't seem right. Parameters aren't just single values since they can be dynamic (and I think this one is because static_target defaults to False).

Also, is there any scenario in which a parameter can have a boundary and path constraint and still be valid? I don't think it is tested.

Parameters are not dynamic, so you could argue that putting a path constraint on them makes no sense.

When we say that a parameter has static or dynamic targets, all that controls is whether or not the value of the parameter gets fanned out to all nodes in the connection. The value through the phase will not change either way. Allowing it to connect to a target shaped with num_nodes just allows us to connect a parameter or a control to the same target without having to change its shape in the ODE.

Dymos allows users to path constraint them and boundary constraint them, but doing both makes little sense and is prohibited. The last test in test_duplicate_constraints.py tests for the presence of a path and boundary constraint on the same parameter. There's no case in which they could possibly have a different value.

Kenneth-T-Moore · 2022-04-29T18:57:15Z

dymos/examples/brachistochrone/test/test_path_constraints.py

+        phase.add_objective('time', loc='final', scaler=10)
+
+        phase.add_boundary_constraint('y', loc='final', equals=5)
+        phase.add_path_constraint('y', lower=5, upper=100)


What happens if the path_constraint and boundary_constraint have different vals?

We still only get one constraint per index of a variable, since OpenMDAO will error on multiple constraints applied to different indices.

In Dymos this is resolved by removing the path constraint from the first or last index if there is an initial or final boundary constraint in place. The boundary constraint takes precedence.

Kenneth-T-Moore · 2022-04-29T19:22:27Z

dymos/phase/phase.py


-        transcription.configure_boundary_constraints('final', self)
+        transcription._configure_boundary_constraints(self)


Kind of wondering why this is the only configure-stack method that has an underscore.

I'm trying to be more cognizant of those methods and members which should be underscored. There's no reason for the user to call the configure methods so they should be underscored, I just haven't moved them all that way yet.

Kenneth-T-Moore · 2022-04-29T19:24:19Z

dymos/phase/test/test_phase.py

+                # It's a static parameter, so it shouldn't matter whether we enforce it
+                # at the start or the end of the phase, so here we'll do both.
+                # Note if we make these equality constraints, some optimizers (SLSQP) will
+                # see the problem as infeasible.


Do we do both? I only see initial on 'g'.

You're right, we tested each individually, not both simultaneously.

Kenneth-T-Moore · 2022-04-29T19:25:36Z

dymos/phase/test/test_phase.py

+
+                # We'll let g vary, but make sure it hits the desired value.
+                # It's a static parameter, so it shouldn't matter whether we enforce it
+                # at the start or the end of the phase, so here we'll do both.


Same here, maybe you meant each, and not both.

Kenneth-T-Moore · 2022-04-29T19:27:46Z

dymos/phase/test/test_phase.py

+                # We'll let g vary, but make sure it hits the desired value.
+                # It's a static parameter, so it shouldn't matter whether we enforce it
+                # at the start or the end of the phase.  In this case we'll use a path
+                # constraint to enforce it, but it will still just be a single scalar constraint.


Not sure I understand that this is a scalar constraint. Isn't static_target false?

static_target=False means that, in the ODE, g is capable of accepting either a control or a parameter. This allows us to use the ODE in different situations without needing to change the shape of g in the ODE itself.

Parameters themselves are always a single value applied throughout the entire phase ("fanned" out to each node), as opposed to a control which can adopt a different value at each node.

robfalck added 17 commits March 4, 2022 16:01

first try at radau constraint aliasing with test_state_path_constrain…

cbb858c

…t_radau

constraint aliasing works for pseudospectral phases

9d1b363

Merge branch 'master' of https://github.com/OpenMDAO/dymos into const…

3d74a73

…raint_aliasing

path fix for parameter constraints

a4e993b

Merge branch 'master' of https://github.com/OpenMDAO/dymos into const…

17cef4e

…raint_aliasing

Reworked constraint aliases for path and boundary constraints.

5732f3c

Dymos now raises an error if a parameter is subject to more than one boundary or path constraint.

refactored configuration of boundary constraints. get_boundary_constr…

961a67f

…aint_src is now _get_objective_src.

Fixed indexing for n-dimensional constrained quantities.

6da7d3c

a few more issues with how to address constraints applied to multiple…

b52310d

… indices

Still working out the way in which indices are applied to path and bo…

aabc74e

…undary constraints, but getting closer to being like numpy

Merge branch 'master' of https://github.com/OpenMDAO/dymos into const…

78ad3df

…raint_aliasing

Use openmdao indexer to resolve constraint indices.

cb1d98d

Allow user to override linear setting of a constraint. Constraints may always be nonlinear. Those that introspection finds definitely may not be linear will result in an exception if set to linear.

Merge branch 'master' of https://github.com/OpenMDAO/dymos into const…

9132f45

…raint_aliasing

Added documentation about constraint aliasing and the linearity of co…

394282f

…nstraints.

robfalck added 4 commits April 25, 2022 11:42

Fixed some missing @require_pyoptsparse tests

e529167

skip tests regarding openmdao case prefixes until that issue is resolved

7bb9184

trouble converging a partial FD example

8f702a3

updated a problematic notebook example with better IPOPT settings

33989d8

robfalck added 4 commits April 27, 2022 10:44

Added test to verify fix for Issue 733, duplicate constraint creation.

9fe109a

Internally dymos now converts all constraint indices to their equivalent positive values to make this detection more consistent.

added test for error message when more than one boundary/path constra…

5fa5ab6

…int involves the same parameter

One more fix

8083fd6

pep8

6399a73

robfalck requested review from ehariton and Kenneth-T-Moore April 27, 2022 19:52

naylor-b requested changes Apr 29, 2022

View reviewed changes

Kenneth-T-Moore requested changes Apr 29, 2022

View reviewed changes

robfalck added 3 commits May 2, 2022 07:50

Edits per Bret's review

db32bb3

Changes per Ken's review

c1164f3

set/list/array issues fixed

c6f87d8

Kenneth-T-Moore approved these changes May 2, 2022

View reviewed changes

naylor-b approved these changes May 2, 2022

View reviewed changes

robfalck merged commit 8d5a9ff into OpenMDAO:master May 3, 2022

robfalck deleted the constraint_aliasing branch August 25, 2022 16:09

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Constraint aliasing #743

Constraint aliasing #743

robfalck commented Apr 25, 2022 •

edited

Loading

review-notebook-app bot commented Apr 25, 2022

coveralls commented Apr 26, 2022 •

edited

Loading

naylor-b Apr 29, 2022

robfalck May 1, 2022

naylor-b Apr 29, 2022

naylor-b Apr 29, 2022

naylor-b Apr 29, 2022

naylor-b Apr 29, 2022

naylor-b Apr 29, 2022

Kenneth-T-Moore Apr 29, 2022

Kenneth-T-Moore Apr 29, 2022

robfalck May 2, 2022

Kenneth-T-Moore Apr 29, 2022

robfalck May 2, 2022

Kenneth-T-Moore Apr 29, 2022

robfalck May 2, 2022 •

edited

Loading

Kenneth-T-Moore Apr 29, 2022

robfalck May 2, 2022

Kenneth-T-Moore Apr 29, 2022

robfalck May 2, 2022

Kenneth-T-Moore Apr 29, 2022

robfalck May 2, 2022


		phase.add_parameter('g', targets=['g'], units='m/s**2', opt=True)

		# Now mistakenly add a boundary constraint at the same loc, variable, and different but equivalent_negative indices


		transcription.configure_boundary_constraints('final', self)
		transcription._configure_boundary_constraints(self)

Constraint aliasing #743

Constraint aliasing #743

Conversation

robfalck commented Apr 25, 2022 • edited Loading

Summary

Related Issues

Backwards incompatibilities

New Dependencies

review-notebook-app bot commented Apr 25, 2022

coveralls commented Apr 26, 2022 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

robfalck May 2, 2022 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

robfalck commented Apr 25, 2022 •

edited

Loading

coveralls commented Apr 26, 2022 •

edited

Loading

robfalck May 2, 2022 •

edited

Loading