def _predict_contexts(self, contexts: np.ndarray, is_predict: bool,
seeds: Optional[np.ndarray] = None, start_index: Optional[int] = None) -> List:
# Get local copy of model, arm_to_expectation and arms to minimize
# communication overhead between arms (processes) using shared objects
arm_to_model = deepcopy(self.arm_to_model)
arm_to_expectation = deepcopy(self.arm_to_expectation)
arms = deepcopy(self.arms)
# Create an empty list of predictions
predictions = [None] * len(contexts)
for index, row in enumerate(contexts):
# Each row needs a separately seeded rng for reproducibility in parallel
rng = np.random.RandomState(seed=seeds[index])
for arm in arms:
# Copy the row rng to the deep copied model in arm_to_model
arm_to_model[arm].rng = rng
# Get the expectation of each arm from its trained model
arm_to_expectation[arm] = arm_to_model[arm].predict(row)
if is_predict:
predictions[index] = argmax(arm_to_expectation)
else:
predictions[index] = arm_to_expectation.copy()
# Return list of predictions
return predictions
def predict(self, contexts: np.ndarray = None) -> Arm:
# Return a random arm with less than epsilon probability
if self.rng.rand() < self.epsilon:
return self.arms[self.rng.randint(0, len(self.arms))]
# Return the first arm with maximum expectation.
return argmax(self.arm_to_expectation)
def predict(self, contexts: np.ndarray = None):
# Generate expectations based on epsilon greedy policy
arm_to_exp = super().predict_expectations()
# Arm with the highest expectation
max_arm = argmax(arm_to_exp)
max_exp = arm_to_exp[max_arm]
print("Max arm: ", max_arm, " with expectation: ", max_exp)
# Arm with second highest expectation
del arm_to_exp[max_arm]
next_max_arm = argmax(arm_to_exp)
next_max_exp = arm_to_exp[next_max_arm]
print("Next max arm: ", next_max_arm, " with expectation: ",
next_max_exp)
# Regret between best and the second best decision
regret = max_exp - next_max_exp
print("Regret: ", regret, " margin: ", self.margin)
# Return the arm with maximum expectation
# if and only if the regret beats the given operational margin
return max_arm if regret >= self.margin else next_max_arm
def _predict_contexts(self, contexts: np.ndarray, is_predict: bool,
seeds: Optional[np.ndarray] = None, start_index: Optional[int] = None) -> List:
# Get local copy of model, arm_to_expectation and arms to minimize
# communication overhead between arms (processes) using shared objects
arm_to_model = deepcopy(self.arm_to_model)
arm_to_expectation = deepcopy(self.arm_to_expectation)
arms = deepcopy(self.arms)
# Create an empty list of predictions
predictions = [None] * len(contexts)
for index, row in enumerate(contexts):
for arm in arms:
# Get the expectation of each arm from its trained model
arm_to_expectation[arm] = arm_to_model[arm].predict(row)
if is_predict:
predictions[index] = argmax(arm_to_expectation)
else:
predictions[index] = arm_to_expectation.copy()
# Return list of predictions
return predictions
def predict(self, contexts: np.ndarray = None) -> Arm:
# Return the arm with maximum expectation. If multiple max value exists, return the first one
return argmax(self.predict_expectations())
def predict(self, contexts: np.ndarray = None) -> Arm:
# Return the first arm with maximum expectation
return argmax(self.arm_to_expectation)
