def _act_impl(self, observation, reward, done):
"""Returns an action from `self.action_space`.
Args:
observation: An observation in self.observation_space.
reward: A scalar value that can be used as a supervising signal.
done: A boolean indicating whether the episode is over.
Raises:
core.EpisodeDoneError if `done` is True.
core.InvalidObservationError if observation is not in
`self.observation_space`.
core.InvalidRewardError if reward is not a scalar or None.
"""
if done:
raise core.EpisodeDoneError("Called act on a done episode.")
if not self.observation_space.contains(observation):
raise core.InvalidObservationError("Invalid observation: %s" %
observation)
core.validate_reward(reward)
# Use `_sample_from` so that the randomness comes from the agent's random
# state rather than the action_space's random_state may be changed by other
# parties.
return self.sample_from(self.action_space)
def _act_impl(self, observation, reward, done):
"""Returns a treatment action.
Args:
observation: An observation in Dict Space with 'population' and
'population_graph' keys.
reward: A scalar float reward value.
done: A boolean indicating whether the simulation has finished.
Returns:
A numpy ndarray containing population indices that represents a treatment
action.
"""
if done:
raise core.EpisodeDoneError('Called act on a done episode.')
if not self.observation_space.contains(observation):
raise core.InvalidObservationError('Invalid observation: %s.' %
observation)
return self._triage(observation)
def _act_impl(self, observation, reward,
done):
"""Returns an action from `self.action_space`.
Args:
observation: An observation in self.observation_space.
reward: A scalar value that can be used as a supervising signal.
done: A boolean indicating whether the episode is over.
Raises:
core.EpisodeDoneError if `done` is True.
core.InvalidObservationError if observation is not in
`self.observation_space`.
"""
if done:
raise core.EpisodeDoneError('Called act on a done episode.')
if not self.observation_space.contains(observation):
raise core.InvalidObservationError('Invalid observation: %s' %
observation)
features = self._feature_selection_fn(observation)
labels = self._label_fn(observation)
# No need to build up a dataset of observations when the model is frozen.
if not self._freeze_classifier():
# Update dataset and fit new model
self._update_dataset(features, labels)
self._train_model()
# Since this method is called afetr initial action in the first step.
self._steps += 1
if self._steps < self._burnin:
return {
'threshold': np.asarray(self._initial_threshold),
'epsilon_prob': np.asarray(self._get_epsilon_prob())
}
return {
'threshold': np.asarray(self._threshold),
'epsilon_prob': np.asarray(self._get_epsilon_prob())
}
def _act_impl(self, observation, reward, done):
"""Returns a fixed threshold.
Args:
observation: An observation in self.observation_space.
reward: A scalar value that can be used as a supervising signal.
done: A boolean indicating whether the episode is over.
Raises:
core.EpisodeDoneError if `done` is True.
core.InvalidObservationError if observation is not in
`self.observation_space`.
"""
if done:
raise core.EpisodeDoneError("Called act on a done episode.")
if not self.observation_space.contains(observation):
raise core.InvalidObservationError("Invalid observation: %s" %
observation)
self._steps += 1
return self.initial_action()
def _act_impl(self, observation, reward, done):
"""Returns an action from 'self.action_space'.
Args:
observation: An observation in self.observation_space.
reward: A scalar value that can be used as a supervising signal.
done: A boolean indicating whether the episode is over.
Returns:
An action from self.action space.
Raises:
core.EpisodeDoneError if `done` is True.
core.InvalidObservationError if observation is not contained in
'self.observation_space'.
gym.error.InvalidAction if the generated action to return is not contained
in 'self.action_space'.
"""
if done:
raise core.EpisodeDoneError("Called act on a done episode.")
if not self.observation_space.contains(observation):
raise core.InvalidObservationError("Invalid ovservation: %s" %
observation)
if self.params.observation_adjustment_fn:
observation = self.params.observation_adjustment_fn(
self.rng, self.beliefs, observation)
features = self.feature_selection_fn(observation)
self.beliefs = self._update_beliefs(features, self.beliefs)
action = self._allocate(self._n_resource, self.beliefs)
if not self.action_space.contains(action):
raise gym.error.InvalidAction("Invalid action: %s" % action)
return action
