def test_should_always_select_winner_with_eps_1(self):
loosing_arm = EpsilonGreedyArm('loosing_arm', n=1, reward_sum=1)
winning_arm = EpsilonGreedyArm('winning_arm', n=1, reward_sum=1000.0)
egreedy = EGreedy('my_bandit', [loosing_arm, winning_arm], epsilon=1.0)
for _ in range(100):
assert egreedy.choose().arm_id == 'winning_arm'
def test_should_be_constructed_from_config_with_default_fallbacks(self):
config = BanditConfig(bandit_id="test_bandit",
arm_ids=["arm1", "arm2", "arm3"])
egreedy: EGreedy = EGreedy.make_from_bandit_config(config)
assert egreedy.bandit_id == "test_bandit"
assert egreedy.epsilon == 0.90
assert egreedy.arms_dict == {
"arm1": EpsilonGreedyArm("arm1", 1, 0),
"arm2": EpsilonGreedyArm("arm2", 1, 0),
"arm3": EpsilonGreedyArm("arm3", 1, 0)
}
def test_should_choose_an_arm(self):
arm_names = ['arm1', 'arm2']
arm = [EpsilonGreedyArm(name, n=1, reward_sum=1) for name in arm_names]
egreedy = EGreedy('my_bandit', arm, epsilon=0.1)
chosen_arm = egreedy.choose()
assert chosen_arm.arm_id == 'arm1' or chosen_arm.arm_id == 'arm2'
def test_should_be_constructed_from_config(self):
config = BanditConfig(bandit_id="test_bandit",
arm_ids=["arm1", "arm2", "arm3"],
priors={
'n': 2,
'reward_sum': 2
},
parameters={'epsilon': 0.80})
egreedy: EGreedy = EGreedy.make_from_bandit_config(config)
assert egreedy.bandit_id == "test_bandit"
assert egreedy.epsilon == 0.80
assert egreedy.arms_dict == {
"arm1": EpsilonGreedyArm("arm1", 2, 2),
"arm2": EpsilonGreedyArm("arm2", 2, 2),
"arm3": EpsilonGreedyArm("arm3", 2, 2)
}
def test_should_be_instantiated_with_internal_map(self):
arm_names = ["arm1", "arm2", "amr3"]
arms = [EpsilonGreedyArm(name, 1, 0) for name in arm_names]
bandit1 = EGreedy("bandit1", arms, epsilon=0.9)
bandit2 = EGreedy("bandit2", arms, epsilon=0.9)
pulpo = Pulpo([bandit1, bandit2])
assert list(pulpo.bandits.values()) == [bandit1, bandit2]
def test_should_choose_an_arm_per_bandit(self):
arm_names = ["arm1", "arm2", "amr3"]
arms = [EpsilonGreedyArm(name, 1, 0) for name in arm_names]
bandit1 = EGreedy("bandit1", arms, epsilon=0.9)
bandit2 = EGreedy("bandit2", arms, epsilon=0.9)
pulpo = Pulpo([bandit1, bandit2])
assert pulpo.choose('bandit1') in ["arm1", "arm2", "amr3"]
assert pulpo.choose('bandit2') in ["arm1", "arm2", "amr3"]
def test_should_update_an_arm_with_feedback(self):
arm_names = ['arm1', 'arm2']
arm = [EpsilonGreedyArm(name, n=1, reward_sum=1) for name in arm_names]
egreedy = EGreedy('my_bandit', arm, epsilon=0.1)
feedback = Feedback('arm1', reward=1)
egreedy.update(feedback)
assert egreedy.arms_dict['arm1'].reward_sum == 2 and egreedy.arms_dict[
'arm1'].n == 2
def test_should_be_updated_with_feedback(self):
arm_names = ["arm1", "arm2", "amr3"]
arms = [EpsilonGreedyArm(name, 1, 0) for name in arm_names]
bandit1 = EGreedy("bandit1", arms, epsilon=0.9)
bandit2 = EGreedy("bandit2", arms, epsilon=0.9)
pulpo = Pulpo([bandit1, bandit2])
pulpo.update("bandit2", "arm2", 100)
bandit2_arm2 = pulpo.bandits['bandit2'].arms_dict['arm2']
assert bandit2_arm2.reward_sum == 100 and bandit2_arm2.n == 2
def test_should_build_from_json_with_optional_parameters_omited(self):
json_string = """
[
{
"bandit_id" : "test_bandit",
"bandit_type" : "epsilon_greedy",
"arm_ids" : ["arm1", "arm2", "arm3"]
}
]
"""
factory = BanditFactory()
bandit_list: List[OnlineBandit] = factory.make_bandits_list(json_string)
new_bandit: EGreedy = bandit_list[0]
assert new_bandit.bandit_id == "test_bandit"
assert new_bandit.epsilon == 0.90
assert new_bandit.arms_dict == {
"arm1": EpsilonGreedyArm("arm1", 1, 0),
"arm2": EpsilonGreedyArm("arm2", 1, 0),
"arm3": EpsilonGreedyArm("arm3", 1, 0)
}
def test_should_run_for_several_iterations_and_sample_all_arms(self):
arm_names = ['arm1', 'arm2', 'arm3']
arm = [EpsilonGreedyArm(name, n=1, reward_sum=1) for name in arm_names]
egreedy = EGreedy('my_bandit', arm, epsilon=0.1)
num_steps = 100
sampled_arm_ids = set()
for _ in range(num_steps):
arm = egreedy.choose()
sampled_arm_ids.add(arm.arm_id)
reward = 1.0 if random.random() < 0.2 else 0.0
egreedy.update(Feedback(arm.arm_id, reward))
assert sampled_arm_ids == set(arm_names)
def test_should_build_from_json(self):
json_string = """
[
{
"bandit_id" : "test_bandit",
"bandit_type" : "epsilon_greedy",
"arm_ids" : ["arm1", "arm2", "arm3"],
"priors" : {"n": 2, "reward_sum": 2},
"parameters" : {"epsilon": 0.80}
}
]
"""
factory = BanditFactory()
bandit_list: List[OnlineBandit] = factory.make_bandits_list(json_string)
new_bandit: EGreedy = bandit_list[0]
assert new_bandit.bandit_id == "test_bandit"
assert new_bandit.epsilon == 0.80
assert new_bandit.arms_dict == {
"arm1": EpsilonGreedyArm("arm1", 2, 2),
"arm2": EpsilonGreedyArm("arm2", 2, 2),
"arm3": EpsilonGreedyArm("arm3", 2, 2)
}
def test_should_reset_value_to_default(self):
arm_names = ['arm1', 'arm2', 'arm3']
arm = [EpsilonGreedyArm(name, n=1, reward_sum=1) for name in arm_names]
egreedy = EGreedy('my_bandit', arm, epsilon=0.1)
num_steps = 20
for _ in range(num_steps):
arm = egreedy.choose()
reward = 1.0 if random.random() < 0.2 else 0.0
egreedy.update(Feedback(arm_id=arm.arm_id, reward=reward))
egreedy.reset()
assert all([
arm.reward_sum == 0 and arm.n == 0.001
for arm in egreedy.arms_dict.values()
])
def make_from_bandit_config(cls, config: BanditConfig):
if config.priors:
n = config.priors[fields.N]
reward_sum = config.priors[fields.REWARD_SUM]
else:
n = EGreedy._DEFAULT_N
reward_sum = EGreedy._DEFAULT_REWARD_SUM
if config.parameters:
epsilon = config.parameters[fields.EPSILON]
else:
epsilon = EGreedy._DEFAULT_EPSILON
arms = [
EpsilonGreedyArm(arm_id, n, reward_sum)
for arm_id in config.arm_ids
]
return cls(config.bandit_id, arms, epsilon)