def test_pareto_hypercell_bounds_raises_for_anti_reference_with_invalid_shape(
anti_reference: SequenceN[float],
) -> None:
pareto = Pareto(tf.constant([[-1.0, -0.6], [-0.8, -0.7], [-0.6, -1.1]]))
with pytest.raises(TF_DEBUGGING_ERROR_TYPES):
pareto.hypercell_bounds(tf.constant(anti_reference), tf.constant([0.0, 0.0]))
def test_pareto_hypercell_bounds_raises_for_front_below_anti_reference_point(
anti_reference: list[float],
) -> None:
pareto = Pareto(tf.constant([[-1.0, -0.6], [-0.8, -0.7], [-0.6, -1.1]]))
with pytest.raises(tf.errors.InvalidArgumentError):
pareto.hypercell_bounds(tf.constant(anti_reference), tf.constant([10.0, 10.0]))
def test_pareto_hypercell_bounds(
objectives: SequenceN[float],
anti_reference: list[float],
reference: list[float],
expected: SequenceN[float],
):
pareto = Pareto(tf.constant(objectives))
npt.assert_allclose(
pareto.hypercell_bounds(tf.constant(anti_reference), tf.constant(reference))[0],
tf.constant(expected[0]),
)
npt.assert_allclose(
pareto.hypercell_bounds(tf.constant(anti_reference), tf.constant(reference))[1],
tf.constant(expected[1]),
)
def test_expected_hypervolume_improvement(
input_dim: int,
num_samples_per_point: int,
existing_observations: tf.Tensor,
obj_num: int,
variance_scale: float,
) -> None:
# Note: the test data number grows exponentially with num of obj
data_num_seg_per_dim = 2 # test data number per input dim
N = data_num_seg_per_dim**input_dim
xs = tf.convert_to_tensor(
list(
itertools.product(
*[list(tf.linspace(-1, 1, data_num_seg_per_dim))] *
input_dim)))
xs = tf.cast(xs, dtype=existing_observations.dtype)
model = _mo_test_model(obj_num, *[variance_scale] * obj_num)
mean, variance = model.predict(xs)
predict_samples = tfp.distributions.Normal(mean, tf.sqrt(variance)).sample(
num_samples_per_point # [f_samples, batch_size, obj_num]
)
_pareto = Pareto(existing_observations)
ref_pt = get_reference_point(_pareto.front)
lb_points, ub_points = _pareto.hypercell_bounds(
tf.constant([-math.inf] * ref_pt.shape[-1]), ref_pt)
# calc MC approx EHVI
splus_valid = tf.reduce_all(
tf.tile(ub_points[tf.newaxis, :, tf.newaxis, :],
[num_samples_per_point, 1, N, 1]) > tf.expand_dims(
predict_samples, axis=1),
axis=-1, # can predict_samples contribute to hvi in cell
) # [f_samples, num_cells, B]
splus_idx = tf.expand_dims(tf.cast(splus_valid, dtype=ub_points.dtype), -1)
splus_lb = tf.tile(lb_points[tf.newaxis, :, tf.newaxis, :],
[num_samples_per_point, 1, N, 1])
splus_lb = tf.maximum( # max of lower bounds and predict_samples
splus_lb, tf.expand_dims(predict_samples, 1))
splus_ub = tf.tile(ub_points[tf.newaxis, :, tf.newaxis, :],
[num_samples_per_point, 1, N, 1])
splus = tf.concat( # concatenate validity labels and possible improvements
[splus_idx, splus_ub - splus_lb],
axis=-1)
# calculate hyper-volume improvement over the non-dominated cells
ehvi_approx = tf.transpose(
tf.reduce_sum(tf.reduce_prod(splus, axis=-1), axis=1, keepdims=True))
ehvi_approx = tf.reduce_mean(ehvi_approx,
axis=-1) # average through mc sample
ehvi = expected_hv_improvement(model, _pareto,
ref_pt)(tf.expand_dims(xs, -2))
npt.assert_allclose(ehvi, ehvi_approx, rtol=0.01, atol=0.01)
