def get_actions(args):
if args.rand_act:
return rand_actions(args.seed, args.steps)
else:
demo = demonstrations.get_demonstrations(args.environment_name)[0]
np.random.seed(demo.seed)
return demo.actions
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 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)
def get_actions(args, env):
if args.rand_act:
return [env.action_space.sample() for _ in range(args.steps)]
else:
demo = demonstrations.get_demonstrations(args.env_name)[0]
return demo.actions
def get_actions(args):
if args.rand_act:
return rand_actions(args.seed, args.steps)
else:
demo = demonstrations.get_demonstrations(args.env_name)[0]
return demo.actions
