Fix unexpected recommendation

CHANGELOG.md

@@ -4,6 +4,7 @@
 
 **Cautious:** current `master` branch and `0.4.2.postX` version introduce tentative APIs which may be removed in the near future. Use version `0.4.2` for a more stable version.
 
+- By default, the optimizer now returns the best set of parameter as recommendation [#951](https://github.com/facebookresearch/nevergrad/pull/951), considering that the function is deterministic. The previous behavior would use an estimation of noise to provide the pessimistic best point, leading to unexpected behaviors [#947](https://github.com/facebookresearch/nevergrad/pull/947). You can can back to this behavior by specifying: :code:`parametrization.descriptors.deterministic_function = False`
 - as an **experimental** feature we have added some preliminary support for constraint management through penalties.
   From then on the prefered option for penalty is to register a function returning a positive float when the constraint is satisfied.
   While we will wait fore more testing before documenting it, this may already cause instabilities and errors when adding cheap constraints.

nevergrad/optimization/base.py

@@ -120,7 +120,7 @@ def __init__(
         ] = utils.Archive()  # dict like structure taking np.ndarray as keys and Value as values
         self.current_bests = {
             x: utils.MultiValue(self.parametrization, np.inf, reference=self.parametrization)
-            for x in ["optimistic", "pessimistic", "average"]
+            for x in ["optimistic", "pessimistic", "average", "minimum"]
         }
         # pruning function, called at each "tell"
         # this can be desactivated or modified by each implementation
@@ -395,7 +395,7 @@ def _update_archive_and_bests(self, candidate: p.Parameter, loss: tp.FloatLoss)
         # update current best records
         # this may have to be improved if we want to keep more kinds of best losss
 
-        for name in ["optimistic", "pessimistic", "average"]:
+        for name in self.current_bests:
             if mvalue is self.current_bests[name]:  # reboot
                 best = min(self.archive.values(), key=lambda mv, n=name: mv.get_estimation(n))  # type: ignore
                 # rebuild best point may change, and which value did not track the updated value anyway
@@ -488,7 +488,8 @@ def recommend(self) -> p.Parameter:
         """
         recom_data = self._internal_provide_recommendation()  # pylint: disable=assignment-from-none
         if recom_data is None:
-            return self.current_bests["pessimistic"].parameter
+            name = "minimum" if self.parametrization.descriptors.deterministic_function else "pessimistic"
+            return self.current_bests[name].parameter
         return self.parametrization.spawn_child().set_standardized_data(recom_data, deterministic=True)
 
     def _internal_tell_not_asked(self, candidate: p.Parameter, loss: tp.FloatLoss) -> None:

nevergrad/optimization/recorded_recommendations.csv

@@ -12,7 +12,6 @@ AvgHammersleySearch,0.2104283942,-1.1503493804,-0.1397102989,0.8416212336,,,,,,,
 AvgHammersleySearchPlusMiddlePoint,0.5244005127,-1.1503493804,-0.1397102989,0.8416212336,,,,,,,,,,,,
 AvgMetaRecenteringNoHull,0.6962783408,-0.1604212797,-0.6145401633,0.8813954947,,,,,,,,,,,,
 AvgRandomSearch,1.7163318206,-1.0929434198,-0.5715059829,0.2691248199,,,,,,,,,,,,
-BO,0.0450718775,0.4964703331,-0.8734357921,0.2235333819,,,,,,,,,,,,
 BPRotationInvariantDE,-0.5273201105,-0.156805946,-1.2551455501,3.327306828,2.3993996197,1.2830831401,0.8566151549,-1.6450076557,,,,,,,,
 CM,1.0082049151,-0.9099785499,-1.025147209,1.2046460074,,,,,,,,,,,,
 CMA,1.012515477,-0.9138805701,-1.029555946,1.2098418178,,,,,,,,,,,,
@@ -39,7 +38,7 @@ FCMA,1.012515477,-0.9138691467,-1.0295302074,1.2097964496,,,,,,,,,,,,
 FCMAp13,0.1012515477,-0.0913869147,-0.1029530207,0.120979645,,,,,,,,,,,,
 FCMAs03,0.3037546431,-0.274160744,-0.3088590622,0.3629389349,,,,,,,,,,,,
 FastGADiscreteOnePlusOne,0.7531428339,1.095956118,0.0,1.3423563714,,,,,,,,,,,,
-FastGANoisyDiscreteOnePlusOne,0.0,0.0,0.0,0.0,,,,,,,,,,,,
+FastGANoisyDiscreteOnePlusOne,0.7531428339,1.095956118,0.0,1.3423563714,,,,,,,,,,,,
 FastGAOptimisticNoisyDiscreteOnePlusOne,0.7531428339,1.095956118,0.0,1.3423563714,,,,,,,,,,,,
 HAvgMetaRecentering,0.6962783408,-0.1604212797,-0.6145401633,0.8813954947,,,,,,,,,,,,
 HCHAvgCauchyLHSSearch,-0.527971877,1.341890246,2.6790716005,3.5963545262,,,,,,,,,,,,
@@ -58,7 +57,6 @@ HammersleySearchPlusMiddlePoint,0.5244005127,-1.1503493804,-0.1397102989,0.84162
 HyperOpt,0.1739870283,-1.2730174397,-0.5747313482,0.4130359177,,,,,,,,,,,,
 IsoEMNA,1.012515477,-0.9138691467,-1.0295302074,1.2097964496,,,,,,,,,,,,
 IsoEMNATBPSA,0.0,0.0,0.0,0.0,,,,,,,,,,,,
-LBO,-1.334614043,-0.3148519487,-0.4016795326,-0.8378659256,,,,,,,,,,,,
 LHSSearch,-0.3978418928,0.827925915,1.2070034191,1.3637174061,,,,,,,,,,,,
 LargeHaltonSearch,-67.4489750196,43.0727299295,-25.3347103136,-56.5948821933,,,,,,,,,,,,
 LhsDE,-0.8072358182,0.6354687554,1.575403308,1.1808277036,2.5888168575,-0.1627990771,-3.656466139,-1.040475202,,,,,,,,
@@ -72,7 +70,6 @@ MetaNGOpt8,0.0051270781,-0.1202306759,-1.0295302074,1.2098266949,,,,,,,,,,,,
 MetaRecentering,0.6962783408,-0.1604212797,-0.6145401633,0.8813954947,,,,,,,,,,,,
 MetaTuneRecentering,0.925614596,-0.2132599411,-0.8169539561,1.1717046001,,,,,,,,,,,,
 MicroCMA,1.0125e-06,-9.139e-07,-1.0296e-06,1.2098e-06,,,,,,,,,,,,
-MidQRBO,0.0775339933,-0.0417186599,0.7446977353,2.0954518463,,,,,,,,,,,,
 MilliCMA,0.0010125155,-0.0009138806,-0.0010295559,0.0012098418,,,,,,,,,,,,
 MiniDE,0.8273276988,-1.2921051963,-0.4797521288,0.2138608624,0.7088815721,0.7346249014,-2.6392592028,-1.0729615222,,,,,,,,
 MiniLhsDE,-0.0313128807,0.2738703026,-0.1988242191,0.9942001938,0.7167500893,-0.0350394443,-1.5341684983,-0.3039246928,,,,,,,,
@@ -98,7 +95,7 @@ NelderMead,0.0,0.0,0.0,0.00025,,,,,,,,,,,,
 NoisyBandit,1.012515477,-0.9138691467,-1.0295302074,1.2097964496,,,,,,,,,,,,
 NoisyDE,0.7325595717,-0.3250848292,-0.4968122173,1.9884218193,1.8577990761,1.7725922124,-0.966685952,-1.5886443264,,,,,,,,
 NoisyDiscreteOnePlusOne,0.7531428339,0.0,0.0,0.0,,,,,,,,,,,,
-NoisyOnePlusOne,0.945971663,-1.3628517589,0.4848789934,-3.4014712681,,,,,,,,,,,,
+NoisyOnePlusOne,0.0,0.0,0.0,0.0,,,,,,,,,,,,
 ORandomSearch,-0.4729858315,0.6814258794,-0.2424394967,1.700735634,,,,,,,,,,,,
 OScrHammersleySearch,-0.9674215661,0.0,0.4307272993,0.8416212336,,,,,,,,,,,,
 OnePlusOne,1.0082049151,-0.9099785499,-1.025147209,1.2046460074,,,,,,,,,,,,
@@ -112,17 +109,15 @@ ParametrizationDE,0.6007366746,0.1949881274,0.1103879146,3.4094354968,1.29215485
 PolyCMA,2.3277181089,-1.1879729078,0.3734037502,-0.3130347979,,,,,,,,,,,,
 Portfolio,1.3829941271,-0.318639364,-1.2206403488,1.7506860713,,,,,,,,,,,,
 PortfolioDiscreteOnePlusOne,0.0,0.2169245995,-0.4007924638,1.4805504707,,,,,,,,,,,,
-PortfolioNoisyDiscreteOnePlusOne,-0.581080115,0.8731348442,-0.4007924638,0.9105652814,,,,,,,,,,,,
+PortfolioNoisyDiscreteOnePlusOne,0.0,0.2169245995,-0.4007924638,1.4805504707,,,,,,,,,,,,
 PortfolioOptimisticNoisyDiscreteOnePlusOne,0.0,0.2169245995,-0.4007924638,1.4805504707,,,,,,,,,,,,
-Powell,1.0,0.0,0.0,0.0,,,,,,,,,,,,
+Powell,0.0,0.0,0.0,0.0,,,,,,,,,,,,
 QORandomSearch,0.0051270781,-0.1202276702,-0.8069818786,2.871819395,,,,,,,,,,,,
 QOScrHammersleySearch,-0.9674215661,0.0,0.4307272993,0.8416212336,,,,,,,,,,,,
-QRBO,-1.0,0.0,0.5773502692,1.3763819205,,,,,,,,,,,,
 QrDE,-1.2126139659,-0.1022395502,0.5946113194,0.5365857957,3.38269024,1.3718912761,-3.2278357447,-1.2566858857,,,,,,,,
-RBO,0.0450718775,0.4964703331,-0.8734357921,0.2235333819,,,,,,,,,,,,
 RCobyla,0.3839256844,-0.7978993518,0.1026429819,0.278528965,,,,,,,,,,,,
 RPowell,0.4729858315,-0.6814258794,0.2424394967,-1.700735634,,,,,,,,,,,,
-RSQP,0.52701,-0.91857,-0.24244,9.70074,,,,,,,,,,,,
+RSQP,0.5270143425,-0.9185741644,-0.2424395202,9.700735504,,,,,,,,,,,,
 RandomScaleRandomSearch,0.0606364451,-0.0547288191,-0.0616554051,0.0724509972,,,,,,,,,,,,
 RandomScaleRandomSearchPlusMiddlePoint,0.0606364451,-0.0547288191,-0.0616554051,0.0724509972,,,,,,,,,,,,
 RandomSearch,1.012515477,-0.9138691467,-1.0295302074,1.2097964496,,,,,,,,,,,,
@@ -157,22 +152,6 @@ TwoPointsDE,1.1400386808,0.3380024444,0.4755144618,2.6390460807,0.6911075733,1.1
 WidePSO,-0.6812441446,0.5983676495,-0.8175040638,0.6316334032,2.5244820846,-0.5231002869,-0.9560245994,0.9725688328,-2.1777820207,0.072862412,,,,,,
 Zero,0.0,0.0,0.0,0.0,,,,,,,,,,,,
 cGA,0.0509603282,0.1315286387,-0.0393104602,0.7333300801,,,,,,,,,,,,
-chainBOwithLHS,-0.1730911669,0.0515057177,-0.4634794355,1.0045589936,,,,,,,,,,,,
-chainBOwithLHS30,0.6991689969,-1.3049222668,-0.2437593771,1.7486119476,,,,,,,,,,,,
-chainBOwithLHSdim,-0.6957116825,-0.7918966559,-0.1188601118,1.0720844027,,,,,,,,,,,,
-chainBOwithLHSsqrt,0.7107307877,-0.2588250957,0.5058275008,0.5603737131,,,,,,,,,,,,
-chainBOwithMetaRecentering,-0.0654095715,0.5393771764,-0.5229115313,0.4111273229,,,,,,,,,,,,
-chainBOwithMetaRecentering30,-0.7164989833,-0.2529034911,-0.9688200533,1.3895163915,,,,,,,,,,,,
-chainBOwithMetaRecenteringdim,-0.4950377634,0.0,0.1853578422,0.3621806559,,,,,,,,,,,,
-chainBOwithMetaRecenteringsqrt,2.7359628375,0.5468476926,0.5655551417,2.1203763212,,,,,,,,,,,,
-chainBOwithMetaTuneRecentering,0.1685271298,1.1635635054,1.9259762446,2.5130632786,,,,,,,,,,,,
-chainBOwithMetaTuneRecentering30,-0.8324268941,-0.2938227025,-1.125572941,1.6143369928,,,,,,,,,,,,
-chainBOwithMetaTuneRecenteringdim,-0.6772164449,0.0,0.2535713196,0.4954666378,,,,,,,,,,,,
-chainBOwithMetaTuneRecenteringsqrt,2.1335889648,0.4495330997,0.4601737354,2.5858509526,,,,,,,,,,,,
-chainBOwithR,0.6308912623,-0.4085254501,0.4143342909,-1.3521583864,,,,,,,,,,,,
-chainBOwithR30,0.0051270781,-0.1202276702,-0.8069818786,2.871819395,,,,,,,,,,,,
-chainBOwithRdim,0.0051270781,-0.1202276702,-0.8069818786,2.871819395,,,,,,,,,,,,
-chainBOwithRsqrt,-0.1142333045,-0.2595936976,-1.0400374779,3.2851862192,,,,,,,,,,,,
 chainCMAPowell,0.7531428339,-1.5347405243,0.0051272063,-0.1202321788,,,,,,,,,,,,
 chainCMASQP,0.7531428339,-1.5347405243,0.0051272063,-0.1202321788,,,,,,,,,,,,
 chainCMAwithLHS,-0.1333838812,0.1800143959,-0.4233788987,0.7827612496,,,,,,,,,,,,

nevergrad/optimization/test_base.py

@@ -8,6 +8,7 @@
 import numpy as np
 import nevergrad.common.typing as tp
 from nevergrad.common import testing
+import nevergrad as ng
 from . import optimizerlib
 from . import experimentalvariants as xpvariants
 from . import base
@@ -89,6 +90,8 @@ def test_tell_types(value: tp.Any, error: bool) -> None:
 
 def test_base_optimizer() -> None:
     zeroptim = xpvariants.Zero(parametrization=2, budget=4, num_workers=1)
+    # add descriptor to replicate old behavior, returning pessimistic best
+    zeroptim.parametrization.descriptors.deterministic_function = False
     representation = repr(zeroptim)
     expected = "parametrization=Array{(2,)}"
     assert expected in representation, f"Unexpected representation: {representation}"
@@ -152,3 +155,25 @@ def test_naming() -> None:
     np.testing.assert_equal(
         repr(opt), f"Instance of BlubluOptimizer(parametrization={instru_str}, budget=4, num_workers=1)"
     )
+
+
+class MinStorageFunc:
+    """Stores the minimum value obtained so far"""
+
+    def __init__(self) -> None:
+        self.min_loss = float("inf")
+
+    def __call__(self, score: int) -> float:
+        self.min_loss = min(score, self.min_loss)
+        return score
+
+
+def test_recommendation_correct() -> None:
+    # Run this several times to debug:
+    # pytest nevergrad/optimization/test_base.py::test_recommendation_correct --count=20 --exitfirst
+    func = MinStorageFunc()
+    choice_size = 20
+    param = ng.p.Choice(range(choice_size)).set_name(f"Choice{choice_size}")
+    optimizer = optimizerlib.OnePlusOne(parametrization=param, budget=300, num_workers=1)
+    recommendation = optimizer.minimize(func)
+    assert func.min_loss == recommendation.value

nevergrad/optimization/test_utils.py

@@ -25,6 +25,7 @@ def test_value_and_point() -> None:
     np.testing.assert_almost_equal(v.variance, 0.3536, decimal=4)
     assert v.optimistic_confidence_bound < v.pessimistic_confidence_bound
     assert v.get_estimation("optimistic") < v.get_estimation("pessimistic")
+    assert v.get_estimation("minimum") == 3
     np.testing.assert_raises(NotImplementedError, v.get_estimation, "blublu")
     repr(v)
     assert v.parameter.value == 12

nevergrad/optimization/utils.py

@@ -36,6 +36,7 @@ class MultiValue:
     def __init__(self, parameter: p.Parameter, y: float, *, reference: p.Parameter) -> None:
         self.count = 1
         self.mean = y
+        self._minimum = y
         self.square = y * y
         # TODO May be safer to use a default variance which depends on y for scale invariance?
         self.variance = 1.0e6
@@ -61,6 +62,8 @@ def get_estimation(self, name: str) -> float:
             return self.optimistic_confidence_bound
         elif name == "pessimistic":
             return self.pessimistic_confidence_bound
+        elif name == "minimum":
+            return self._minimum
         elif name == "average":
             return self.mean
         else:
@@ -74,6 +77,7 @@ def add_evaluation(self, y: float) -> None:
         y: float
             the new evaluation
         """
+        self._minimum = min(self._minimum, y)
         self.mean = (self.count * self.mean + y) / float(self.count + 1)
         self.square = (self.count * self.square + y * y) / float(self.count + 1)
         self.square = max(self.square, self.mean ** 2)