def test_dominates_complete_vs_incomplete(t1_state: TrialState) -> None:
directions = [StudyDirection.MINIMIZE, StudyDirection.MAXIMIZE]
t1 = create_trial(values=[0, 2], state=t1_state)
t2 = create_trial(values=[1, 1], state=TrialState.COMPLETE)
assert _dominates(t2, t1, list(directions))
assert not _dominates(t1, t2, list(directions))
def test_dominates_1d_not_equal(v1: float, v2: float) -> None:
t1 = create_trial(values=[v1])
t2 = create_trial(values=[v2])
assert _dominates(t1, t2, [StudyDirection.MINIMIZE])
assert not _dominates(t2, t1, [StudyDirection.MINIMIZE])
assert _dominates(t2, t1, [StudyDirection.MAXIMIZE])
assert not _dominates(t1, t2, [StudyDirection.MAXIMIZE])
def test_dominates_invalid() -> None:
directions = [StudyDirection.MINIMIZE, StudyDirection.MAXIMIZE]
# The numbers of objectives for `t1` and `t2` don't match.
t1 = create_trial(values=[1]) # One objective.
t2 = create_trial(values=[1, 2]) # Two objectives.
with pytest.raises(ValueError):
_dominates(t1, t2, directions)
# The numbers of objectives and directions don't match.
t1 = create_trial(values=[1]) # One objective.
t2 = create_trial(values=[1]) # One objective.
with pytest.raises(ValueError):
_dominates(t1, t2, directions)
def test_dominates_2d() -> None:
directions = [StudyDirection.MINIMIZE, StudyDirection.MAXIMIZE]
# Check all pairs of trials consisting of these values, i.e.,
# [-inf, -inf], [-inf, -1], [-inf, 1], [-inf, inf], [-1, -inf], ...
# These values should be specified in ascending order.
vals = [-float("inf"), -1, 1, float("inf")]
# The following table illustrates an example of dominance relations.
# "d" cells in the table dominates the "t" cell in (MINIMIZE, MAXIMIZE) setting.
#
# value1
# ╔═════╤═════╤═════╤═════╤═════╗
# ║ │ -∞ │ -1 │ 1 │ ∞ ║
# ╟─────┼─────┼─────┼─────┼─────╢
# ║ -∞ │ │ │ d │ d ║
# ╟─────┼─────┼─────┼─────┼─────╢
# ║ -1 │ │ │ d │ d ║
# value0 ╟─────┼─────┼─────┼─────┼─────╢
# ║ 1 │ │ │ t │ d ║
# ╟─────┼─────┼─────┼─────┼─────╢
# ║ ∞ │ │ │ │ ║
# ╚═════╧═════╧═════╧═════╧═════╝
#
# In the following code, we check that for each position of "t" cell, the relation
# above holds.
# Generate the set of all possible indices.
all_indices = set(
(i, j) for i in range(len(vals)) for j in range(len(vals)))
for (t_i, t_j) in all_indices:
# Generate the set of all indices that dominates the current index.
dominating_indices = set((d_i, d_j) for d_i in range(t_i + 1)
for d_j in range(t_j, len(vals)))
dominating_indices -= {(t_i, t_j)}
for (d_i, d_j) in dominating_indices:
trial1 = create_trial(values=[vals[t_i], vals[t_j]])
trial2 = create_trial(values=[vals[d_i], vals[d_j]])
assert _dominates(trial2, trial1, directions)
for (d_i, d_j) in all_indices - dominating_indices:
trial1 = create_trial(values=[vals[t_i], vals[t_j]])
trial2 = create_trial(values=[vals[d_i], vals[d_j]])
assert not _dominates(trial2, trial1, directions)
def test_constrained_dominates_feasible_vs_feasible(
direction1: StudyDirection, direction2: StudyDirection,
constraints_list: List[List[float]]) -> None:
directions = [direction1, direction2]
# Check all pairs of trials consisting of these values, i.e.,
# [-inf, -inf], [-inf, -1], [-inf, 1], [-inf, inf], [-1, -inf], ...
values_list = [
[x, y]
for x in [-float("inf"), -1, 1, float("inf")]
for y in [-float("inf"), -1, 1, float("inf")]
]
values_constraints_list = [(vs, cs) for vs in values_list
for cs in constraints_list]
# The results of _constrained_dominates match _dominates in all feasible cases.
for (values1, constraints1) in values_constraints_list:
for (values2, constraints2) in values_constraints_list:
t1 = _create_frozen_trial(0, values1, constraints1)
t2 = _create_frozen_trial(1, values2, constraints2)
assert _constrained_dominates(t1, t2, directions) == _dominates(
t1, t2, directions)
def _constrained_dominates(trial0: FrozenTrial, trial1: FrozenTrial,
directions: Sequence[StudyDirection]) -> bool:
"""Checks constrained-domination.
A trial x is said to constrained-dominate a trial y, if any of the following conditions is
true:
1) Trial x is feasible and trial y is not.
2) Trial x and y are both infeasible, but solution x has a smaller overall constraint
violation.
3) Trial x and y are feasible and trial x dominates trial y.
"""
constraints0 = trial0.system_attrs.get(_CONSTRAINTS_KEY)
constraints1 = trial1.system_attrs.get(_CONSTRAINTS_KEY)
if constraints0 is None:
warnings.warn(f"Trial {trial0.number} does not have constraint values."
" It will be dominated by the other trials.")
if constraints1 is None:
warnings.warn(f"Trial {trial1.number} does not have constraint values."
" It will be dominated by the other trials.")
if constraints0 is None and constraints1 is None:
# Neither Trial x nor y has constraints values
return _dominates(trial0, trial1, directions)
if constraints0 is not None and constraints1 is None:
# Trial x has constraint values, but y doesn't.
return True
if constraints0 is None and constraints1 is not None:
# If Trial y has constraint values, but x doesn't.
return False
assert isinstance(constraints0, (list, tuple))
assert isinstance(constraints1, (list, tuple))
if len(constraints0) != len(constraints1):
raise ValueError(
"Trials with different numbers of constraints cannot be compared.")
if trial0.state != TrialState.COMPLETE:
return False
if trial1.state != TrialState.COMPLETE:
return True
satisfy_constraints0 = all(v <= 0 for v in constraints0)
satisfy_constraints1 = all(v <= 0 for v in constraints1)
if satisfy_constraints0 and satisfy_constraints1:
# Both trials satisfy the constraints.
return _dominates(trial0, trial1, directions)
if satisfy_constraints0:
# trial0 satisfies the constraints, but trial1 violates them.
return True
if satisfy_constraints1:
# trial1 satisfies the constraints, but trial0 violates them.
return False
# Both trials violate the constraints.
violation0 = sum(v for v in constraints0 if v > 0)
violation1 = sum(v for v in constraints1 if v > 0)
return violation0 < violation1
def test_dominates_1d_equal(v: float, direction: StudyDirection) -> None:
assert not _dominates(create_trial(values=[v]), create_trial(values=[v]),
[direction])
