Merge pull request #5365 from nabenabe0928/hotfix/fix-prob-zero-error…

…-in-gp-sampler

Add a unit test for convergence of acquisition function in `GPSampler`

optuna/_gp/optim_mixed.py

@@ -262,7 +262,7 @@ def local_search_mixed(
             # Parameters not changed from the beginning.
             return (best_normalized_params, best_fval)
 
-    _logger.warn("local_search_mixed: Local search did not converge.")
+    _logger.warning("local_search_mixed: Local search did not converge.")
     return (best_normalized_params, best_fval)
 
 
@@ -298,10 +298,13 @@ def optimize_acqf_mixed(
     probs = np.exp(f_vals - f_vals[max_i])
     probs[max_i] = 0.0  # We already picked the best param, so remove it from roulette.
     probs /= probs.sum()
+    n_non_zero_probs_improvement = np.count_nonzero(probs > 0.0)
     # n_additional_warmstart becomes smaller when study starts to converge.
     n_additional_warmstart = min(
-        n_local_search - len(warmstart_normalized_params_array) - 1, np.count_nonzero(probs > 0.0)
+        n_local_search - len(warmstart_normalized_params_array) - 1, n_non_zero_probs_improvement
     )
+    if n_additional_warmstart == n_non_zero_probs_improvement:
+        _logger.warning("Study already converged, so the number of local search is reduced.")
     chosen_idxs = np.array([max_i])
     if n_additional_warmstart > 0:
         additional_idxs = rng.choice(

tests/samplers_tests/test_gp.py

@@ -0,0 +1,45 @@
+from __future__ import annotations
+
+from _pytest.logging import LogCaptureFixture
+import numpy as np
+
+import optuna
+import optuna._gp.acqf as acqf
+import optuna._gp.optim_mixed as optim_mixed
+import optuna._gp.prior as prior
+import optuna._gp.search_space as gp_search_space
+
+
+def test_after_convergence(caplog: LogCaptureFixture) -> None:
+    # A large `optimal_trials` causes the instability in the kernel inversion, leading to
+    # instability in the variance calculation.
+    X_uniform = [(i + 1) / 10 for i in range(10)]
+    X_uniform_near_optimal = [(i + 1) / 1e5 for i in range(20)]
+    X_optimal = [0.0] * 10
+    X = np.array(X_uniform + X_uniform_near_optimal + X_optimal)
+    score_vals = -(X - np.mean(X)) / np.std(X)
+    search_space = gp_search_space.SearchSpace(
+        scale_types=np.array([gp_search_space.ScaleType.LINEAR]),
+        bounds=np.array([[0.0, 1.0]]),
+        steps=np.zeros(1, dtype=float),
+    )
+    kernel_params = optuna._gp.gp.fit_kernel_params(
+        X=X[:, np.newaxis],
+        Y=score_vals,
+        is_categorical=np.array([False]),
+        log_prior=prior.default_log_prior,
+        minimum_noise=prior.DEFAULT_MINIMUM_NOISE_VAR,
+        deterministic_objective=False,
+    )
+    acqf_params = acqf.create_acqf_params(
+        acqf_type=acqf.AcquisitionFunctionType.LOG_EI,
+        kernel_params=kernel_params,
+        search_space=search_space,
+        X=X[:, np.newaxis],
+        Y=score_vals,
+    )
+    caplog.clear()
+    optuna.logging.enable_propagation()
+    optim_mixed.optimize_acqf_mixed(acqf_params, rng=np.random.RandomState(42))
+    # len(caplog.text) > 0 means the optimization has already converged.
+    assert len(caplog.text) > 0, "Did you change the kernel implementation?"