def test_invalid_add_arm_scaler(self):
scaler = StandardScaler()
arm_to_scaler = {0: deepcopy(scaler), 1: deepcopy(scaler)}
mab = MAB([0, 1], LearningPolicy.LinUCB(arm_to_scaler=arm_to_scaler))
with self.assertRaises(TypeError):
mab.add_arm(2, scaler=deepcopy(scaler))
def test_add_arm_scaler(self):
scaler = StandardScaler()
scaler.fit(
np.array([[1, 2, 3, 4, 5], [5, 4, 3, 2, 1]]).astype('float64'))
arm_to_scaler = {0: deepcopy(scaler), 1: deepcopy(scaler)}
mab = MAB([0, 1], LearningPolicy.LinUCB(arm_to_scaler=arm_to_scaler))
mab.add_arm(2, scaler=deepcopy(scaler))
def test_invalid_add_arm(self):
mab = MAB([1, 2, 3], LearningPolicy.EpsilonGreedy(epsilon=0))
with self.assertRaises(ValueError):
mab.add_arm(None)
with self.assertRaises(ValueError):
mab.add_arm(np.nan)
with self.assertRaises(ValueError):
mab.add_arm(np.inf)
with self.assertRaises(ValueError):
mab.add_arm(3)
prediction = radius.predict(test)
# Expectation of each ad based on learning from past ad revenues
expectations = radius.predict_expectations(test)
# Results
print("Radius: ", prediction, " ", expectations)
assert (prediction == [4, 4])
# Online update of model
radius.partial_fit(decisions=prediction,
rewards=test_df_revenue,
contexts=test)
# Updating of the model with new arm
radius.add_arm(6)
########################################################
# KNearest Neighborhood Policy with UCB1 Learning Policy
########################################################
# KNearest context policy with k equals to 5 and ucb1 learning with alpha of 1.25
knearest = MAB(arms=ads,
learning_policy=LearningPolicy.UCB1(alpha=1.25),
neighborhood_policy=NeighborhoodPolicy.KNearest(k=5))
# Learn from previous ads shown and revenues generated
knearest.fit(decisions=train_df['ad'],
rewards=train_df['revenues'],
contexts=train)
# Expected revenues of each layouts learnt from historical data based on epsilon greedy policy
expectations = greedy.predict_expectations()
# Results
print("Epsilon Greedy: ", prediction, " ", expectations)
assert (prediction == 1)
# Additional historical data becomes available which allows _online learning
additional_layouts = [1, 2, 1, 2]
additional_revenues = [0, 12, 7, 19]
# Online updating of the model
greedy.partial_fit(additional_layouts, additional_revenues)
# Adding a new layout option
greedy.add_arm(3)
#################################################
# Randomized Popularity Learning Policy
#################################################
# Randomized Popularity learning policy that select arms
# with weighted probability based on the mean reward for each arm
popularity = MAB(arms=options,
learning_policy=LearningPolicy.Popularity(),
seed=123456)
# Learn from previous layouts decisions and revenues generated
popularity.fit(decisions=layouts, rewards=revenues)
# Predict the next best layouts decision
prediction = linucb.predict(test)
# Expectation of each ad based on learning from past ad revenues
expectations = linucb.predict_expectations(test)
# Results
print("LinUCB: ", prediction, " ", expectations)
assert (prediction == [5, 2])
# Online update of model
linucb.partial_fit(decisions=prediction,
rewards=test_df_revenue,
contexts=test)
# Update the model with new arm
linucb.add_arm(6)
###################################################################
# LinUCB Learning Policy combined with Radius Neighborhood Policy
###################################################################
# Radius context policy with radius equals to 1 and LinUCB learning with alpha of 1
radius = MAB(arms=ads,
learning_policy=LearningPolicy.LinUCB(alpha=1),
neighborhood_policy=NeighborhoodPolicy.Radius(radius=1))
# Learn from previous ads shown and revenues generated
radius.fit(decisions=train_df['ad'],
rewards=train_df['revenues'],
contexts=train)
