def test_custom_centroids(use_kd_tree):
centroids = np.array([[-0.25, -0.25], [0.25, 0.25]])
archive = CVTArchive(centroids.shape[0], [(-1, 1), (-1, 1)],
custom_centroids=centroids,
use_kd_tree=use_kd_tree)
archive.initialize(solution_dim=3)
assert archive.samples is None
assert (archive.centroids == centroids).all()
def _cvt_archive():
"""Deterministically created CVTArchive."""
archive = CVTArchive(100, [(-1, 1), (-1, 1)],
samples=1000,
use_kd_tree=True,
seed=42)
archive.initialize(solution_dim=2)
_add_uniform_sphere(archive, (-1, 1), (-1, 1))
return archive
def _long_cvt_archive():
"""Same as above, but the behavior space is longer in one direction."""
archive = CVTArchive(100, [(-2, 2), (-1, 1)],
samples=1000,
use_kd_tree=True,
seed=42)
archive.initialize(solution_dim=2)
_add_uniform_sphere(archive, (-2, 2), (-1, 1))
return archive
def test_custom_centroids_bad_shape(use_kd_tree):
with pytest.raises(ValueError):
# The centroids array should be of shape (10, 2) instead of just (1, 2),
# hence a ValueError will be raised.
CVTArchive(10, [(-1, 1), (-1, 1)],
custom_centroids=[[0.0, 0.0]],
use_kd_tree=use_kd_tree)
def test_heatmap_fails_on_non_2d(archive_type):
archive = {
"grid": lambda: GridArchive([20, 20, 20], [(-1, 1)] * 3),
"cvt": lambda: CVTArchive(100, [(-1, 1)] * 3, samples=100),
"sliding":
lambda: SlidingBoundariesArchive([20, 20, 20], [(-1, 1)] * 3),
}[archive_type]()
archive.initialize(solution_dim=2) # Arbitrary.
with pytest.raises(ValueError):
{
"grid": grid_archive_heatmap,
"cvt": cvt_archive_heatmap,
"sliding": sliding_boundaries_archive_heatmap,
}[archive_type](archive)
def setup():
archive = CVTArchive(1000, [(-1, 1), (-1, 1)],
samples=20_000,
use_kd_tree=use_kd_tree)
archive.initialize(solutions.shape[1])
# Let numba compile.
archive.add(solutions[0], objective_values[0], behavior_values[0])
return (archive,), {}
def benchmark_get_10k_random_elites(use_kd_tree, benchmark, benchmark_data_10k):
n, solutions, objective_values, behavior_values = benchmark_data_10k
archive = CVTArchive(1000, [(-1, 1), (-1, 1)],
samples=20_000,
use_kd_tree=use_kd_tree)
archive.initialize(solutions.shape[1])
for i in range(n):
archive.add(solutions[i], objective_values[i], behavior_values[i])
@benchmark
def get_elites():
for i in range(n):
sol, obj, beh = archive.get_random_elite()
def benchmark_as_pandas_2000_items(benchmark):
bins = 2000
archive = CVTArchive(bins, [(-1, 1), (-1, 1)],
use_kd_tree=True,
samples=50_000)
archive.initialize(10)
for x, y in archive.centroids:
sol = np.random.random(10)
sol[0] = x
sol[1] = y
archive.add(sol, 1.0, np.array([x, y]))
# Archive should be full.
assert len(archive.as_pandas()) == bins
benchmark(archive.as_pandas)
def main():
"""Creates archives, times insertion into them, and plots results."""
archive = None
n_bins = [10, 50, 100, 500, 1_000, 5_000, 10_000, 100_000]
# Pre-made solutions to insert.
n_vals = 100_000
solutions = np.random.uniform(-1, 1, (n_vals, 10))
objective_values = np.random.randn(n_vals)
behavior_values = np.random.uniform(-1, 1, (n_vals, 2))
# Set up these archives so we can use the same centroids across all
# experiments for a certain number of bins (and also save time).
ref_archives = {
bins: CVTArchive(
bins,
[(-1, 1), (-1, 1)],
# Use 200k bins to avoid dropping clusters.
samples=n_vals if bins != 10_000 else 200_000,
use_kd_tree=False)
for bins in n_bins
}
def get_archive_data(name, dtype=np.float64):
"""Returns ArchiveFixtureData to use for testing each archive.
The archives vary, but there will always be an empty 2D archive, as well as
a 2D archive with a single solution added to it. This solution will have a
value of [1, 2, 3], its objective value will be 1.0, and its behavior values
will be [0.25, 0.25].
The name is the name of an archive to create. It should come from
ARCHIVE_NAMES.
"""
# Characteristics of a single solution to insert into archive_with_entry.
solution = np.array([1, 2, 3])
objective_value = 1.0
behavior_values = np.array([0.25, 0.25])
grid_indices = None
centroid = None
if name == "GridArchive":
# Grid archive with 10 bins and range (-1, 1) in first dim, and 20 bins
# and range (-2, 2) in second dim.
bins = 10 * 20
archive = GridArchive([10, 20], [(-1, 1), (-2, 2)], dtype=dtype)
archive.initialize(len(solution))
archive_with_entry = GridArchive([10, 20], [(-1, 1), (-2, 2)],
dtype=dtype)
archive_with_entry.initialize(len(solution))
grid_indices = (6, 11)
elif name.startswith("CVTArchive-"):
# CVT archive with bounds (-1,1) and (-1,1), and 4 centroids at (0.5,
# 0.5), (-0.5, 0.5), (-0.5, -0.5), and (0.5, -0.5). The entry in
# archive_with_entry should match with centroid (0.5, 0.5).
bins = 4
kd_tree = name == "CVTArchive-kd_tree"
samples = [[0.5, 0.5], [-0.5, 0.5], [-0.5, -0.5], [0.5, -0.5]]
centroid = [0.5, 0.5]
archive = CVTArchive(4, [(-1, 1), (-1, 1)],
samples=samples,
use_kd_tree=kd_tree,
dtype=dtype)
archive.initialize(len(solution))
archive_with_entry = CVTArchive(4, [(-1, 1), (-1, 1)],
samples=samples,
use_kd_tree=kd_tree,
dtype=dtype)
archive_with_entry.initialize(len(solution))
elif name == "SlidingBoundariesArchive":
# Sliding boundary archive with 10 bins and range (-1, 1) in first dim,
# and 20 bins and range (-2, 2) in second dim.
bins = 10 * 20
archive = SlidingBoundariesArchive([10, 20], [(-1, 1), (-2, 2)],
remap_frequency=100,
buffer_capacity=1000,
dtype=dtype)
archive.initialize(len(solution))
archive_with_entry = SlidingBoundariesArchive([10, 20], [(-1, 1),
(-2, 2)],
remap_frequency=100,
buffer_capacity=1000,
dtype=dtype)
archive_with_entry.initialize(len(solution))
grid_indices = (6, 11)
archive_with_entry.add(solution, objective_value, behavior_values)
return ArchiveFixtureData(
archive,
archive_with_entry,
solution,
objective_value,
behavior_values,
grid_indices,
centroid,
bins,
)
def init():
archive = CVTArchive(1000, [(-1, 1), (-1, 1)],
samples=20_000,
use_kd_tree=use_kd_tree)
archive.initialize(solution_dim=2)
def setup(bins, use_kd_tree):
nonlocal archive
archive = CVTArchive(bins, [(-1, 1), (-1, 1)],
custom_centroids=ref_archives[bins].centroids,
use_kd_tree=use_kd_tree)
archive.initialize(solutions.shape[1])
def test_samples_bad_shape(use_kd_tree):
# The behavior space is 2D but samples are 3D.
with pytest.raises(ValueError):
CVTArchive(10, [(-1, 1), (-1, 1)],
samples=[[-1, -1, -1], [1, 1, 1]],
use_kd_tree=use_kd_tree)
