Keep 1+1 sigma unchanged if equal loss (#1061)

nevergrad/functions/pyomo/test_core.py

@@ -67,7 +67,7 @@ def test_abstract_model_with_constraints() -> None:
 
 def test_pyomo_set() -> None:
     def square2(m: tp.Any) -> float:
-        return (m.x - 1) ** 2
+        return (m.x - 1) ** 2  # type: ignore
 
     model = pyomo.ConcreteModel()
     model.P = pyomo.Set(initialize=list(range(1, 11)))
@@ -78,6 +78,7 @@ def square2(m: tp.Any) -> float:
     model.constraint1 = pyomo.Constraint(rule=lambda m: m.x >= 2)
 
     func = core.Pyomo(model)
+    func.parametrization.random_state.seed(12)
     optimizer = ng.optimizers.OnePlusOne(parametrization=func.parametrization, budget=100)
     recommendation = optimizer.minimize(func.function)
 

nevergrad/optimization/optimizerlib.py

@@ -73,6 +73,7 @@ def __init__(
     ) -> None:
         super().__init__(parametrization, budget=budget, num_workers=num_workers)
         self._sigma: float = 1
+        self._previous_best_loss = float("inf")
         self.use_pareto = use_pareto
         all_params = paramhelpers.flatten_parameter(self.parametrization)
         arity = max(
@@ -241,6 +242,8 @@ def _internal_ask_candidate(self) -> p.Parameter:
 
     def _internal_tell(self, x: tp.ArrayLike, loss: tp.FloatLoss) -> None:
         # only used for cauchy and gaussian
+        if self._previous_best_loss != loss:
+            self._sigma *= 2.0 if loss < self._previous_best_loss else 0.84
         if self.mutation == "doerr" and self._doerr_current_best < float("inf") and self._doerr_index >= 0:
             improvement = max(0.0, self._doerr_current_best - loss)
             # Decay.
@@ -254,20 +257,16 @@ def _internal_tell(self, x: tp.ArrayLike, loss: tp.FloatLoss) -> None:
             self._doerr_index = -1
         if self.mutation == "doerr":
             self._doerr_current_best = min(self._doerr_current_best, loss)
-        self._sigma *= 2.0 if loss <= self.current_bests["pessimistic"].mean else 0.84
         if self.mutation == "adaptive":
-            factor = (
-                1.2 if loss <= self.current_bests["pessimistic"].mean else 0.731
-            )  # 0.731 = 1.2**(-np.exp(1)-1)
+            factor = 1.2 if loss <= self._previous_best_loss else 0.731  # 0.731 = 1.2**(-np.exp(1)-1)
             self._adaptive_mr = min(1.0, factor * self._adaptive_mr)
         if self.mutation == "coordinatewise_adaptive":
-            factor = (
-                1.2 if loss < self.current_bests["pessimistic"].mean else 0.731
-            )  # 0.731 = 1.2**(-np.exp(1)-1)
+            factor = 1.2 if loss < self._previous_best_loss else 0.731  # 0.731 = 1.2**(-np.exp(1)-1)
             inds = self._modified_variables
             self._velocity[inds] = np.clip(
                 self._velocity[inds] * factor, 1.0, self.arity_for_discrete_mutation / 4.0
             )
+        self._previous_best_loss = self.current_bests["pessimistic"].mean  # could be the current one
 
 
 class ParametrizedOnePlusOne(base.ConfiguredOptimizer):