def reset(self):
"""Start a new episode."""
# Build a new game and retrieve its first set of state/reward/discount.
self._current_game = self._game_factory()
self._state = environment.StepType.FIRST
# Collect environment returns from starting the game and update state.
observations, reward, discount = self._current_game.its_showtime()
self._update_for_game_step(observations, reward, discount)
return environment.TimeStep(step_type=self._state,
reward=None,
discount=None,
observation=self.last_observations)
def step(self, action):
"""Apply action, step the world forward, and return observations."""
if self._action_size == 1:
# Handle a float or single-element arrays of any dimensionality. Strictly
# speaking, a single-element list will also work, but it's best not to
# confuse matters in the docstring with this option.
all_actions = [np.asarray(action).item()]
else:
all_actions = [np.asarray(a).item() for a in action]
if len(all_actions) != self._action_size:
raise RuntimeError("A pycolab Environment adapter's step method "
'was called with actions that were not compatible '
'with what the pycolab game expects.')
# Clear episode internals and start a new episode, if episode ended or if
# the game was not already underway.
if self._state == environment.StepType.LAST:
self._drop_last_episode()
if self._current_game is None:
return self.reset()
# Execute the action in pycolab.
action = all_actions[0] if self._action_size == 1 else all_actions
observations, reward, discount = self._current_game.play(action)
self._update_for_game_step(observations, reward, discount)
# Check the current status of the game.
if self._game_over:
self._state = environment.StepType.LAST
else:
self._state = environment.StepType.MID
return environment.TimeStep(
step_type=self._state,
reward=self._last_reward,
discount=self._last_discount,
observation=self.last_observations)