Bug fix: Correctly add noise to benchmark problems (facebook#3203)

Summary:
Pull Request resolved: facebook#3203

There was this piece of incorrect logic:

```
df["mean"] = df["Y_true"] + np.random.normal(len(df)) * df["sem"]
```

`len(df)` is not the size/shape parameter; it's the mean. So this generates one single normal draw with mean `len(df)` and variance 1, then multiplies it by `df["sem"]`.

This diff fixes that.

Reviewed By: saitcakmak

Differential Revision: D67521651

fbshipit-source-id: ef6d66761ac2ea023782a7347e841f0bdba7c5ce

ax/benchmark/benchmark_runner.py

@@ -110,7 +110,7 @@ def _add_noise(
         else:
             df["sem"] = noise_std_ser
 
-        df["mean"] = df["Y_true"] + np.random.normal(len(df)) * df["sem"]
+        df["mean"] = df["Y_true"] + np.random.normal(loc=0, scale=df["sem"])
 
     else:
         df["sem"] = 0.0

ax/benchmark/tests/test_benchmark_runner.py

@@ -15,7 +15,7 @@
 import numpy as np
 
 import torch
-from ax.benchmark.benchmark_runner import BenchmarkRunner
+from ax.benchmark.benchmark_runner import _add_noise, BenchmarkRunner
 from ax.benchmark.benchmark_test_functions.botorch_test import BoTorchTestFunction
 from ax.benchmark.benchmark_test_functions.surrogate import SurrogateTestFunction
 from ax.benchmark.problems.synthetic.hss.jenatton import (
@@ -39,6 +39,7 @@
 )
 from botorch.test_functions.synthetic import Ackley, ConstrainedHartmann, Hartmann
 from botorch.utils.transforms import normalize
+from pandas import DataFrame
 from pyre_extensions import none_throws
 
 
@@ -268,6 +269,38 @@ def test_runner(self) -> None:
                 ):
                     BenchmarkRunner.deserialize_init_args({})
 
+    def test__add_noise(self) -> None:
+        np.random.seed(0)
+        y_true = np.arange(6)
+        arm_name = ["0_0", "0_1", "0_0", "0_1", "0_0", "0_1"]
+        metric_name = ["foo", "foo", "bar", "bar", "baz", "baz"]
+
+        df = DataFrame(
+            {"Y_true": y_true, "metric_name": metric_name, "arm_name": arm_name}
+        )
+
+        noise_stds = {"foo": 1, "bar": 2, "baz": 3}
+        arm_weights = {"0_0": 1, "0_1": 2}
+        result = _add_noise(df=df, noise_stds=noise_stds, arm_weights=arm_weights)
+        self.assertEqual(set(result.columns), set(df.columns) | {"mean", "sem"})
+        expected_sem = df["metric_name"].map(noise_stds) / np.sqrt(
+            df["arm_name"].map(arm_weights) / 3
+        )
+        self.assertEqual(result["sem"].tolist(), expected_sem.tolist())
+        noise = df["mean"] - df["Y_true"]
+        self.assertNotEqual(noise.std(), 0)
+
+        z_scores = noise / expected_sem
+        self.assertNotEqual(z_scores.std(), 0)
+
+        chi_squared_stat = (z_scores**2).sum()
+        # None of these assertions would have failed in 10M simulations.
+        # Each has some tolerance from the most extreme value seen in 10M sims.
+        self.assertGreater(chi_squared_stat, 0.005)
+        self.assertLess(chi_squared_stat, 45)
+        self.assertLess(np.abs(z_scores).min(), 2)
+        self.assertGreater(z_scores.max(), 0.05)
+
     def test_heterogeneous_noise(self) -> None:
         outcome_names = ["objective_0", "constraint"]
         noise_dict = {"objective_0": 0.1, "constraint": 0.05}