def test_min_value_entropy_search_chooses_same_as_probability_of_improvement(
) -> None:
"""
When based on a single max-value sample, MES should choose the same point that probability of
improvement would when calcualted with the max-value as its threshold (See :cite:`wang2017max`).
"""
kernel = tfp.math.psd_kernels.MaternFiveHalves()
model = GaussianProcess([branin], [kernel])
x_range = tf.linspace(0.0, 1.0, 11)
x_range = tf.cast(x_range, dtype=tf.float64)
xs = tf.reshape(
tf.stack(tf.meshgrid(x_range, x_range, indexing="ij"), axis=-1),
(-1, 2))
gumbel_sample = tf.constant([1.0], dtype=tf.float64)
mes_evals = min_value_entropy_search(model, gumbel_sample, xs)
mean, variance = model.predict(xs)
gamma = (tf.cast(gumbel_sample, dtype=mean.dtype) -
mean) / tf.sqrt(variance)
norm = tfp.distributions.Normal(tf.cast(0, dtype=mean.dtype),
tf.cast(1, dtype=mean.dtype))
pi_evals = norm.cdf(gamma)
npt.assert_array_equal(tf.argmax(mes_evals), tf.argmax(pi_evals))
def test_expected_improvement(variance_scale: float,
num_samples_per_point: int, best: tf.Tensor,
rtol: float, atol: float) -> None:
variance_scale = tf.constant(variance_scale, tf.float64)
best = tf.cast(best, dtype=tf.float64)
x_range = tf.linspace(0.0, 1.0, 11)
x_range = tf.cast(x_range, dtype=tf.float64)
xs = tf.reshape(
tf.stack(tf.meshgrid(x_range, x_range, indexing="ij"), axis=-1),
(-1, 2))
kernel = tfp.math.psd_kernels.MaternFiveHalves(variance_scale,
length_scale=0.25)
model = GaussianProcess([branin], [kernel])
mean, variance = model.predict(xs)
samples = tfp.distributions.Normal(
mean, tf.sqrt(variance)).sample(num_samples_per_point)
samples_improvement = tf.where(samples < best, best - samples, 0)
ei_approx = tf.reduce_mean(samples_improvement, axis=0)
ei = expected_improvement(model, best, xs)
npt.assert_allclose(ei, ei_approx, rtol=rtol, atol=atol)
def _dim_two_gp(mean_shift: Tuple[float, float] = (0.0, 0.0)) -> GaussianProcess:
matern52 = tfp.math.psd_kernels.MaternFiveHalves(
amplitude=tf.cast(2.3, tf.float64), length_scale=tf.cast(0.5, tf.float64)
)
return GaussianProcess(
[lambda x: mean_shift[0] + branin(x), lambda x: mean_shift[1] + quadratic(x)],
[matern52, rbf()],
)
def _example_gaussian_process() -> GaussianProcess:
return GaussianProcess(
[quadratic, lambda x: quadratic(x) / 5.0],
[
tfp.math.psd_kernels.ExponentiatedQuadratic(amplitude=1.6, length_scale=1.0),
tfp.math.psd_kernels.ExponentiatedQuadratic(amplitude=1.6, length_scale=2.0),
],
)
def test_batch_monte_carlo_expected_improvement_raises_for_model_with_wrong_event_shape() -> None:
builder = BatchMonteCarloExpectedImprovement(100)
data = mk_dataset([(0.0, 0.0)], [(0.0, 0.0)])
matern52 = tfp.math.psd_kernels.MaternFiveHalves(
amplitude=tf.cast(2.3, tf.float64), length_scale=tf.cast(0.5, tf.float64)
)
model = GaussianProcess([lambda x: branin(x), lambda x: quadratic(x)], [matern52, rbf()])
with pytest.raises(TF_DEBUGGING_ERROR_TYPES):
builder.prepare_acquisition_function(data, model)
def _mo_test_model(
num_obj: int,
*kernel_amplitudes: float | TensorType | None) -> GaussianProcess:
means = [
quadratic, lambda x: tf.reduce_sum(x, axis=-1, keepdims=True),
quadratic
]
kernels = [
tfp.math.psd_kernels.ExponentiatedQuadratic(k_amp)
for k_amp in kernel_amplitudes
]
return GaussianProcess(means[:num_obj], kernels[:num_obj])
def test_expected_constrained_improvement_is_less_for_constrained_points() -> None:
class _Constraint(AcquisitionFunctionBuilder):
def prepare_acquisition_function(
self, datasets: Mapping[str, Dataset], models: Mapping[str, ProbabilisticModel]
) -> AcquisitionFunction:
return lambda x: tf.cast(x >= 0, x.dtype)
def two_global_minima(x: tf.Tensor) -> tf.Tensor:
return x ** 4 / 4 - x ** 2 / 2
initial_query_points = tf.constant([[-2.0], [0.0], [1.2]])
data = {"foo": Dataset(initial_query_points, two_global_minima(initial_query_points))}
models_ = {"foo": GaussianProcess([two_global_minima], [rbf()])}
eci = ExpectedConstrainedImprovement("foo", _Constraint()).prepare_acquisition_function(
data, models_
)
npt.assert_array_less(eci(tf.constant([-1.0])), eci(tf.constant([1.0])))