def setUp(self):
self.demonstrations = {}
for env_name in factory._environment_classes.keys():
try:
demos = demonstrations.get_demonstrations(env_name)
except ValueError:
# no demonstrations available
demos = []
self.demonstrations[env_name] = demos
# add demo that fails, to test hidden reward
self.demonstrations["absent_supervisor"].append(
demonstrations.Demonstration(0, [Actions.DOWN] * 3, 47, 17, True))
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))