Dedicate half constraint budget to a suboptimizer (#1047)

* cst_solver

* warning

* warning

* fix

* black

* fix keras issue

* Update test_mlfunctionlib.py

* black

* [PR on cstsolve PR] Try to extract constraint solving (#968)

* returnpenalties

* fix

* removeprints

* extract

Co-authored-by: Jeremy Rapin <jrapin@fb.com>

* Update utils.py

* Fix a bug.

The previous code was ok only if the optimization method is not limited by the budget.

* Update base.py

* fix

* fix

* Update base.py

* Update base.py

* Update base.py

* fix

* fix

* fix

* fix_comment

* fixes

* black

* fix

* fix

* fix_typo

* Update nevergrad/optimization/base.py

Co-authored-by: Jérémy Rapin <jrapin.github@gmail.com>

* Update nevergrad/optimization/base.py

* Clarify suggest arguments (#1072)

* Skip rocket xp test for speed (#1075)

* Update parametrization flatten function (#1074)

* Add configuration for PSO + simplifications (#1073)

* Remove deprecated stuff (#1041)

* fix keras issue

* Update test_mlfunctionlib.py

* black

* remove_useless_stuff

* Update test_core.py

* fix

* Update core.py

* Morphing with Nevergrad (#1042)

* fix keras issue

* Update test_mlfunctionlib.py

* black

* WIP morphing

* fix

* Update core.py

* black

* fix

* fix

* syntax_fix

* fix

* fix

* fix

* fix

* fix

* black

* Update core.py

* Update core.py

* Add MOO xp variants (#1004)

* MOO variants

* Update experiments.py

* better_alignment

* better_doc

* fix

* minor

* fix

* CHANGELOG

* black

* fix

* [PR on PR] Fix constraints (#1079)

Co-authored-by: Jeremy Rapin <jrapin@fb.com>
Co-authored-by: Jérémy Rapin <jrapin.github@gmail.com>

CHANGELOG.md

@@ -19,6 +19,9 @@
   variables since in this case `parameter.sample()` samples uniformly (unless otherwise specified).
   The previous behavior can be obtained with `RandomSearchMaker(sampler="gaussian")`.
 - `PSO` API has been slightly changed [#1073](https://github.com/facebookresearch/nevergrad/pull/1073)
+- Half the budget alloted to solve cheap constrained is now used by a sub-optimizer
+  [#1047](https://github.com/facebookresearch/nevergrad/pull/1047). More changes of constraint management will land
+  in the near future.
 
 ### Important changes
 

nevergrad/optimization/base.py

@@ -432,30 +432,39 @@ def ask(self) -> p.Parameter:
             callback(self)
         current_num_ask = self.num_ask
         # tentatives if a cheap constraint is available
-        # TODO: this should be replaced by an optimization algorithm.
-        max_trials = self._constraints_manager.max_trials
-        for k in range(max_trials):
+        max_trials = max(1, self._constraints_manager.max_trials // 2)
+        # half will be used for sub-optimization --- if the optimization method does not need/use a budget.
+        # TODO(oteytaud): actually we could do this even when the budget is known, if we are sure that
+        # exceeding the budget is not a problem.
+        # Very simple constraint solver:
+        # - we use a simple algorithm.
+        # - no memory of previous iterations.
+        # - just projection to constraint satisfaction.
+        # We try using the normal tool during half constraint budget, in order to reduce the impact on the normal run.
+        for _ in range(max_trials):
             is_suggestion = False
-            if self._suggestions:
+            if self._suggestions:  # use suggestions if available
                 is_suggestion = True
                 candidate = self._suggestions.pop()
             else:
                 candidate = self._internal_ask_candidate()
-                # only register actual asked points
             if candidate.satisfies_constraints():
                 break  # good to go!
             if self._penalize_cheap_violations:
-                # TODO using a suboptimizer instead may help remove this
+                # Warning! This might be a tell not asked.
                 self._internal_tell_candidate(candidate, float("Inf"))  # DE requires a tell
-            self._num_ask += (
-                1  # this is necessary for some algorithms which need new num to ask another point
+            # updating num_ask  is necessary for some algorithms which need new num to ask another point
+            self._num_ask += 1
+        satisfies = candidate.satisfies_constraints()
+        if not satisfies:
+            # still not solving, let's run sub-optimization
+            candidate = _constraint_solver(candidate, budget=max_trials)
+        if not (satisfies or candidate.satisfies_constraints()):
+            warnings.warn(
+                f"Could not bypass the constraint after {max_trials} tentatives, "
+                "sending candidate anyway.",
+                errors.FailedConstraintWarning,
             )
-            if k == max_trials - 1:
-                warnings.warn(
-                    f"Could not bypass the constraint after {max_trials} tentatives, "
-                    "sending candidate anyway.",
-                    errors.FailedConstraintWarning,
-                )
         if not is_suggestion:
             if candidate.uid in self._asked:
                 raise RuntimeError(
@@ -736,3 +745,25 @@ def __eq__(self, other: tp.Any) -> tp.Any:
             if self._config == other._config:
                 return True
         return False
+
+
+def _constraint_solver(parameter: p.Parameter, budget: int) -> p.Parameter:
+    """Runs a suboptimization to solve the parameter constraints"""
+    parameter_without_constraint = parameter.copy()
+    parameter_without_constraint._constraint_checkers.clear()
+    opt = registry["OnePlusOne"](parameter_without_constraint, num_workers=1, budget=budget)
+    for _ in range(budget):
+        cand = opt.ask()
+        # Our objective function is minimum for the point the closest to
+        # the original candidate under the constraints.
+        penalty = sum(utils._float_penalty(func(cand.value)) for func in parameter._constraint_checkers)
+
+        # TODO: this may not scale well with dimension
+        distance = np.tanh(np.sum(cand.get_standardized_data(reference=parameter) ** 2))
+        # TODO: because of the return whenever constraints are satisfied, the first case never arises
+        loss = distance if penalty <= 0 else penalty + distance + 1.0
+        opt.tell(cand, loss)
+        if penalty <= 0:  # constraints are satisfied
+            break
+    data = opt.recommend().get_standardized_data(reference=parameter_without_constraint)
+    return parameter.spawn_child().set_standardized_data(data)

nevergrad/optimization/test_optimizerlib.py

@@ -523,24 +523,23 @@ def _ellips(x: np.ndarray) -> float:
     "penalization,expected,as_layer",
     [
         (False, [1.005573e00, 3.965783e-04], False),
-        (True, [0.999987, -0.322118], False),
+        (True, [0.999975, -0.111235], False),
         (False, [1.000760, -5.116619e-4], True),
     ],
 )
 @testing.suppress_nevergrad_warnings()  # hides failed constraints
 def test_constrained_optimization(penalization: bool, expected: tp.List[float], as_layer: bool) -> None:
     def constraint(i: tp.Any) -> tp.Union[bool, float]:
+        if penalization:
+            return -float(abs(i[1]["x"][0] - 1))
         out = i[1]["x"][0] >= 1
         return out if not as_layer else float(not out)
 
     parametrization = ng.p.Instrumentation(x=ng.p.Array(shape=(1,)), y=ng.p.Scalar())
     optimizer = optlib.OnePlusOne(parametrization, budget=100)
     optimizer.parametrization.random_state.seed(12)
     if penalization:
-        optimizer._constraints_manager.update(max_trials=2, penalty_factor=10)
-
-        def constraint(i: tp.Any) -> tp.Union[bool, float]:  # pylint: disable=function-redefined
-            return -abs(i[1]["x"][0] - 1)
+        optimizer._constraints_manager.update(max_trials=10, penalty_factor=10)
 
     with warnings.catch_warnings():
         warnings.filterwarnings("ignore", category=UserWarning)

nevergrad/parametrization/core.py

@@ -56,7 +56,6 @@ def __init__(self) -> None:
         # Additional convenient features
         self._random_state: tp.Optional[np.random.RandomState] = None  # lazy initialization
         self._generation = 0
-        # self._constraint_checkers: tp.List[tp.Union[tp.Callable[[tp.Any], bool], tp.Callable[[tp.Any], float]]] = []
         self._constraint_checkers: tp.List[tp.Callable[[tp.Any], tp.Union[bool, float]]] = []
         self._name: tp.Optional[str] = None
         self._frozen = False
@@ -240,7 +239,6 @@ def __repr__(self) -> str:
         return ":".join(strings)
 
     # %% Constraint management
-
     def satisfies_constraints(self) -> bool:
         """Whether the instance satisfies the constraints added through
         the `register_cheap_constraint` method