def testDistributionRaisesError(self):
# Set up the smoothing policy.
policy = StateIncrementPolicy(self._time_step_spec, self._action_spec)
smoothed_policy = temporal_action_smoothing.TemporalActionSmoothing(
policy, smoothing_coefficient=0.5)
# Create actions sampled in time order.
policy_state = smoothed_policy.get_initial_state(batch_size=1)
with self.assertRaises(NotImplementedError):
smoothed_policy.distribution(self._time_step, policy_state)
def testSmoothedActions(self, smoothing_coefficient, expected_actions):
# Set up the smoothing policy.
policy = StateIncrementPolicy(self._time_step_spec, self._action_spec)
smoothed_policy = temporal_action_smoothing.TemporalActionSmoothing(
policy, smoothing_coefficient)
# Create actions sampled in time order.
policy_state = smoothed_policy.get_initial_state(batch_size=1)
smoothed_actions = []
for _ in range(5):
action, policy_state, unused_policy_info = smoothed_policy.action(
self._time_step, policy_state=policy_state)
smoothed_actions.append(action)
# Make sure smoothed actions are as expected.
smoothed_actions_ = self.evaluate(smoothed_actions)
self.assertAllClose(np.squeeze(smoothed_actions_), expected_actions)