def test_simple_run(self):
arm_num = 5
horizon = 10
learner = UCB(arm_num=arm_num)
learner.reset()
mock_ucb = np.array([1.2, 1, 1, 1, 1])
# pylint: disable=protected-access
learner._UCB__UCB = MagicMock(return_value=mock_ucb)
# During the initial time steps, each arm is pulled once
for time in range(1, arm_num + 1):
assert learner.actions(
Context()).SerializeToString() == text_format.Parse(
"""
arm_pulls <
arm <
id: {arm_id}
>
times: 1
>
""".format(arm_id=time - 1), Actions()).SerializeToString()
learner.update(
text_format.Parse(
"""
arm_feedbacks <
arm <
id: {arm_id}
>
rewards: 0
>
""".format(arm_id=time - 1), Feedback()))
# For the left time steps, arm 0 is always the choice
for _ in range(arm_num + 1, horizon + 1):
assert learner.actions(
Context()).SerializeToString() == text_format.Parse(
"""
arm_pulls <
arm <
id: 0
>
times: 1
>
""", Actions()).SerializeToString()
learner.update(
text_format.Parse(
"""
arm_feedbacks <
arm <
id: 0
>
rewards: 0
>
""", Feedback()))
def test_simple_run(self):
arm_num = 5
horizon = 10
learner = Uniform(arm_num=arm_num)
learner.reset()
for time in range(1, horizon + 1):
assert learner.actions(
Context()).SerializeToString() == text_format.Parse(
"""
arm_pulls <
arm <
id: {arm_id}
>
times: 1
>
""".format(arm_id=(time - 1) % arm_num),
Actions()).SerializeToString()
learner.update(
text_format.Parse(
"""
arm_feedbacks <
arm <
id: 0
>
rewards: 0
>
""", Feedback()))
def feed(self, actions: Actions) -> Feedback:
feedback = Feedback()
for arm_pull in actions.arm_pulls:
arm_feedback = self._take_action(arm_pull=arm_pull)
if arm_feedback.rewards:
feedback.arm_feedbacks.append(arm_feedback)
return feedback
def feed(self, actions: Actions) -> Feedback:
feedback = Feedback()
for arm_pull in actions.arm_pulls:
if arm_pull.times > 0:
arm_feedback = self._take_action(arm_pull=arm_pull)
feedback.arm_feedbacks.append(arm_feedback)
return feedback
def test_simple_run(self):
means = [0, 0.5, 0.7, 1]
arms = [BernoulliArm(mean) for mean in means]
learner = EpsGreedy(arm_num=len(arms))
learner.reset()
# Pull each arm once during the initial steps
for time in range(1, len(arms) + 1):
assert learner.actions(
Context()).SerializeToString() == text_format.Parse(
"""
arm_pulls <
arm <
id: {arm_id}
>
times: 1
>
""".format(arm_id=time - 1), Actions()).SerializeToString()
learner.update(
text_format.Parse(
"""
arm_feedbacks <
arm <
id: {arm_id}
>
rewards: 0
>
""".format(arm_id=time - 1), Feedback()))
def test_simple_run(self):
arm_num = 5
budget = 20
learner = SH(arm_num=arm_num, budget=budget)
learner.reset()
while True:
actions = learner.actions(Context())
if not actions.arm_pulls:
break
feedback = Feedback()
for arm_pull in actions.arm_pulls:
arm_feedback = feedback.arm_feedbacks.add()
arm_feedback.arm.id = arm_pull.arm.id
arm_feedback.rewards.extend(list(np.zeros(arm_pull.times)))
learner.update(feedback)
assert learner.best_arm in list(range(arm_num))
def test_simple_run(self):
arm_num = 3
confidence = 0.95
learner = ExpGap(arm_num=arm_num, confidence=confidence)
learner.reset()
while True:
actions = learner.actions(Context())
if not actions.arm_pulls:
break
feedback = Feedback()
for arm_pull in actions.arm_pulls:
arm_feedback = feedback.arm_feedbacks.add()
arm_feedback.arm.id = arm_pull.arm.id
arm_feedback.rewards.extend(
list(
np.random.normal(arm_pull.arm.id / arm_num, 1,
arm_pull.times)))
learner.update(feedback)
assert learner.best_arm in list(range(arm_num))
def test_simple_run(self):
arm_num = 5
horizon = 10
learner = ExploreThenCommit(arm_num=arm_num, T_prime=6)
learner.reset()
for _ in range(1, horizon + 1):
actions = learner.actions(Context())
assert len(actions.arm_pulls) == 1
arm_pull = actions.arm_pulls[0]
arm_id = arm_pull.arm.id
assert arm_pull.times == 1
learner.update(
text_format.Parse(
"""
arm_feedbacks <
arm <
id: {arm_id}
>
rewards: 0
>
""".format(arm_id=arm_id), Feedback()))
def test_simple_run(self):
horizon = 10
features = [
np.array([1, 0]),
np.array([1, 0]),
np.array([1, 0]),
np.array([1, 0]),
np.array([0, 1])
]
learner = LinUCB(features, 0.1, 1e-3)
learner.reset()
mock_ucb = np.array([1.2, 1, 1, 1, 1])
# pylint: disable=protected-access
learner._LinUCB__LinUCB = MagicMock(return_value=mock_ucb)
# Always 0th arm is picked
# not the most efficient test
for _ in range(1, horizon + 1):
assert learner.actions(
Context()).SerializeToString() == text_format.Parse(
"""
arm_pulls <
arm <
id: 0
>
times: 1
>
""", Actions()).SerializeToString()
learner.update(
text_format.Parse(
"""
arm_feedbacks <
arm <
id: 0
>
rewards: 0
>
""", Feedback()))
def test_actions(self):
# Test actions are in the right range
arm_num = 10
budget = 15
apt = APT(arm_num=arm_num, theta=0.5, eps=0)
apt.reset()
for _ in range(budget):
actions = apt.actions(Context())
assert len(actions.arm_pulls) == 1
arm_id = actions.arm_pulls[0].arm.id
assert 0 <= arm_id < arm_num
apt.update(
text_format.Parse(
"""
arm_feedbacks <
arm <
id: {arm_id}
>
rewards: 0
>
""".format(arm_id=arm_id), Feedback()))
def test_simple_run(self):
arm_num = 3
confidence = 0.95
learner = CentralizedLilUCBHeuristic(arm_num=arm_num,
confidence=confidence,
assigned_arms=np.arange(arm_num))
learner.reset()
while True:
actions = learner.actions()
if not actions.arm_pulls:
break
feedback = Feedback()
for arm_pull in actions.arm_pulls:
arm_feedback = feedback.arm_feedbacks.add()
arm_feedback.arm.id = arm_pull.arm.id
arm_feedback.rewards.extend(
list(
np.random.normal(arm_pull.arm.id / arm_num, 1,
arm_pull.times)))
learner.update(feedback)
assert learner.best_arm in list(range(arm_num))