return env
env = setup_env(FourRooms(goal_pos=tasks.pop(0)))
env.unwrapped.max_steps = 1000000
obs = env.reset()
n_states = env.observation_space
n_actions = env.action_space.n + 1
# Set up loggers
# TODO: use RLlog
loglevel = 20
logger = ProjectLogger(level=loglevel, printing=False)
plotter = PlotterOneHot(env)
db = redis.StrictRedis(port=6379)
logger.critical(env)
# Define a network shared across options' policies and terminations,
# as well as the critic
net = NatureConvBody(in_channels=3)
params = [net.parameters()]
# Create options
rng = np.random.RandomState(1338)
n_options = 8
options, options_params = create_options(n_options, net.feature_dim,
env.action_space.n)
# Define a policy over options
actor = EgreedyPolicy(ε=0.02, rng=rng, options=options, loglevel=20)
from hrl.project_logger import ProjectLogger
from hrl.utils import cache
from hrl.visualization.plotter_one_hot import PlotterOneHot
""" Evaluate the benefits of planning with options. """
SAVEPATH = Path(f'{EXPERIMENT_DIR}/SMDP_planning')
if __name__ == '__main__':
# Create environment
env = FullyObsWrapper(FourRooms(goal_pos=(15, 15)))
# Create loggers
LOGLEVEL = 20
logger = ProjectLogger(level=LOGLEVEL, printing=False)
logger.critical(env)
plotter = PlotterOneHot(env=env)
SAVEPATH /= env.unwrapped.__class__.__name__
SAVEPATH.mkdir(parents=True, exist_ok=True)
# Create hard-coded options
options = [
HallwayOption(o, env.observation_space.shape[::-1])
for o in sorted(HallwayOption.hallway_options)
]
options += [
PrimitiveOption(o, env.observation_space.shape[::-1])
for o in sorted(PrimitiveOption.primitive_options)
]
logger.info('Learning option models')