XPs with sparsity on Open AI Gym (#1319)

* XPs with sparsity on Open AI Gym

* Update gymexperiments.py

nevergrad/benchmark/gymexperiments.py

@@ -53,6 +53,7 @@ def ng_full_gym(
     ng_gym: bool = False,  # pylint: disable=redefined-outer-name
     conformant: bool = False,
     gp: bool = False,
+    sparse: bool = False,
 ) -> tp.Iterator[Experiment]:
     """Gym simulator. Maximize reward.  Many distinct problems.
 
@@ -159,17 +160,25 @@ def ng_full_gym(
         )
         for neural_factor in neural_factors:
             for name in env_names:
-                try:
-                    func = nevergrad_gym.GymMulti(
-                        name, control=control, neural_factor=neural_factor, randomized=randomized
-                    )
-                except MemoryError:
-                    continue
-                for budget in budgets:
-                    for algo in optims:
-                        xp = Experiment(func, algo, budget, num_workers=1, seed=next(seedg))
-                        if not xp.is_incoherent:
-                            yield xp
+                sparse_limits: tp.List[tp.Optional[int]] = [None]
+                if sparse:
+                    sparse_limits += [10, 100, 1000]
+                for sparse_limit in sparse_limits:
+                    try:
+                        func = nevergrad_gym.GymMulti(
+                            name,
+                            control=control,
+                            neural_factor=neural_factor,
+                            randomized=randomized,
+                            sparse_limit=sparse_limit,
+                        )
+                    except MemoryError:
+                        continue
+                    for budget in budgets:
+                        for algo in optims:
+                            xp = Experiment(func, algo, budget, num_workers=1, seed=next(seedg))
+                            if not xp.is_incoherent:
+                                yield xp
 
 
 @registry.register
@@ -205,6 +214,15 @@ def gp(seed: tp.Optional[int] = None) -> tp.Iterator[Experiment]:
     return ng_full_gym(seed, gp=True)
 
 
+@registry.register
+def sparse_gp(seed: tp.Optional[int] = None) -> tp.Iterator[Experiment]:
+    """GP benchmark.
+
+    Counterpart of ng_full_gym with a specific, reduced list of problems for matching
+    a genetic programming benchmark."""
+    return ng_full_gym(seed, gp=True, sparse=True)
+
+
 @registry.register
 def ng_stacking_gym(seed: tp.Optional[int] = None) -> tp.Iterator[Experiment]:
     """Counterpart of ng_gym with a recurrent network."""