def get_env(env_name, noops):
"""Get a copy of the environment for simulating the baseline."""
if env_name == 'box':
env = factory.get_environment_obj('side_effects_sokoban', noops=noops)
elif env_name in ['vase', 'sushi', 'sushi_goal']:
env = factory.get_environment_obj('conveyor_belt',
variant=env_name,
noops=noops)
else:
env = factory.get_environment_obj(env_name)
return env
def __init__(self, env_name, cheat=False, render_animation_delay=0.1):
self._env_name = env_name
self.cheat = cheat
self._render_animation_delay = render_animation_delay
self._viewer = None
self._env = factory.get_environment_obj(env_name)
self._rbg = None
self._last_hidden_reward = 0
self.action_space = GridworldsActionSpace(self._env)
self.observation_space = GridworldsObservationSpace(self._env)
def __init__(self, env_name, pause):
self._env_name = env_name
self._pause = pause
self._viewer = None
# TODO
self.action_space = None
self.observation_space = None
self._env = factory.get_environment_obj(env_name)
def get_env(env_name, noops,
movement_reward=-1, goal_reward=1, side_effect_reward=-1):
"""Get a copy of the environment for simulating the baseline."""
if env_name == 'box' or 'sokocoin' in env_name:
levels = {'box': 0, 'sokocoin1': 1, 'sokocoin2': 2, 'sokocoin3': 3}
sizes = {'box': 36, 'sokocoin1': 100, 'sokocoin2': 72, 'sokocoin3': 100}
env = factory.get_environment_obj(
'side_effects_sokoban', noops=noops, movement_reward=movement_reward,
goal_reward=goal_reward, wall_reward=side_effect_reward,
corner_reward=side_effect_reward, level=levels[env_name])
size = sizes[env_name]
elif 'sushi' in env_name or env_name == 'vase':
env = factory.get_environment_obj(
'conveyor_belt', variant=env_name, noops=noops, goal_reward=goal_reward)
size = 49
else:
env = factory.get_environment_obj(env_name)
size = None
return env, size
def main(unused_argv):
# Set random seed.
if FLAGS.seed is not None:
seed = FLAGS.seed
else:
# Get a new random random seed and remember it.
seed = np.random.randint(0, 100)
np.random.seed(seed)
# Run one episode.
actions_list = [] # This stores the actions taken.
env = factory.get_environment_obj(FLAGS.environment)
# Get the module so we can obtain environment specific constants.
module = importlib.import_module(env.__class__.__module__)
# Overwrite the environment's step function to record the actions.
old_step = env.step
def _step(actions):
actions_list.append(actions)
return old_step(actions)
env.step = _step
ui = safety_ui.make_human_curses_ui(module.GAME_BG_COLOURS,
module.GAME_FG_COLOURS)
ui.play(env)
# Extract data
episode_return = env.episode_return
safety_performance = env.get_overall_performance()
actions = _postprocess_actions(actions_list)
# Determine termination reason.
if actions[-1] == 'q':
# Player has quit the game, remove it from the sequence.
actions = actions[:-1]
terminates = False
else:
terminates = True
# Print the resulting demonstration to the terminal.
demo = demonstrations.Demonstration(seed, actions, episode_return,
safety_performance, terminates)
print('Recorded the following data:\n{}'.format(demo))
def view_agent_env(av, args):
logger = get_logger()
env = factory.get_environment_obj(args.environment_name)
env.reset()
av.display(env)
episode_return = 0
actions = get_actions(args)
e = 0
for (i, action) in enumerate(actions):
timestep = env.step(action)
episode_return += reward(timestep)
av.display(env)
if timestep.step_type.last():
logger.info('episode %d: %.2f' % (e, episode_return))
env.reset()
av.reset_time()
av.display(env)
episode_return = 0
e += 1
def test_demonstrations(self, environment_name):
"""Execute the demonstrations in the given environment."""
demos = demonstrations.get_demonstrations(environment_name)
# Execute each demonstration.
for demo in demos:
# Run several times to be sure that result is deterministic.
for _ in range(REPETITIONS):
# Fix random seed.
np.random.seed(demo.seed)
# Construct and run environment.
env = factory.get_environment_obj(environment_name)
episode_return = self._run_env(env, demo.actions, demo.terminates)
# Check return and safety performance.
self.assertEqual(episode_return, demo.episode_return)
if demo.terminates:
hidden_reward = env.get_overall_performance()
else:
hidden_reward = env._get_hidden_reward(default_reward=None)
if hidden_reward is not None:
self.assertEqual(hidden_reward, demo.safety_performance)