def test_continuous_state_env_3():
env = NRoom(nrooms=3, array_observation=True)
counter = OnlineDiscretizationCounter(env.observation_space,
env.action_space,
rate_power=0.5,
min_dist=0.0)
for N in range(10, 20):
assert counter.get_n_visited_states() == 0
assert counter.get_entropy() == 0.0
for ss in range(env.discrete_observation_space.n):
for aa in range(env.action_space.n):
for _ in range(N):
ns, rr, _, _ = env.sample(ss, aa)
continuous_ss = env._convert_index_to_float_coord(ss)
counter.update(continuous_ss, aa, None, rr)
assert counter.N_sa[ss, aa] == N
assert counter.count(continuous_ss, aa) == N
assert np.allclose(counter.measure(continuous_ss, aa),
np.sqrt(1.0 / N))
assert counter.get_n_visited_states(
) == env.discrete_observation_space.n
assert np.allclose(counter.get_entropy(),
np.log2(env.discrete_observation_space.n))
counter.reset()
def test_n_room(reward_free, array_observation, initial_state_distribution):
env = NRoom(reward_free=reward_free,
array_observation=array_observation,
initial_state_distribution=initial_state_distribution)
initial_state = env.reset()
next_state, reward, _, _ = env.step(1)
if initial_state_distribution == 'uniform':
assert env.initial_state_distribution[
0] == 1.0 / env.observation_space.n
assert env.observation_space.contains(initial_state)
assert env.observation_space.contains(next_state)
if reward_free:
assert env.reward_at == {}
if array_observation:
assert isinstance(initial_state, np.ndarray)
assert isinstance(next_state, np.ndarray)
"""
=====================
Demo: demo_vis2d
=====================
"""
from rlberry.envs.benchmarks.grid_exploration.nroom import NRoom, get_nroom_state_coord
from rlberry.envs.classic_control import MountainCar
from rlberry.wrappers.vis2d import Vis2dWrapper
from rlberry.agents import RSUCBVIAgent
from rlberry.agents.dynprog import ValueIterationAgent
CHOICE = 1
if CHOICE == 0:
env = NRoom(nrooms=5, array_observation=False, reward_free=True)
env = Vis2dWrapper(env,
n_bins_obs=20,
memory_size=100,
state_preprocess_fn=get_nroom_state_coord)
agent = ValueIterationAgent(env.unwrapped,
gamma=0.99,
horizon=200,
copy_env=False)
else:
env = MountainCar()
env = Vis2dWrapper(env, n_bins_obs=20, memory_size=200)
agent = RSUCBVIAgent(
env,
gamma=0.99,
===============================
Illustration of NRooms environment
.. video:: ../../video_plot_rooms.mp4
:width: 600
"""
# sphinx_gallery_thumbnail_path = 'thumbnails/video_plot_rooms.jpg'
from rlberry.envs.benchmarks.grid_exploration.nroom import NRoom
from rlberry.agents.dynprog import ValueIterationAgent
env = NRoom(
nrooms=9,
remove_walls=False,
room_size=9,
initial_state_distribution="center",
include_traps=True,
)
horizon = env.observation_space.n
agent = ValueIterationAgent(env, gamma=0.999, horizon=horizon)
print("fitting...")
info = agent.fit()
print(info)
env.enable_rendering()
for _ in range(10):
state = env.reset()
for tt in range(horizon):
from rlberry.envs.benchmarks.grid_exploration.four_room import FourRoom
from rlberry.envs.benchmarks.grid_exploration.six_room import SixRoom
from rlberry.envs.benchmarks.grid_exploration.nroom import NRoom
from rlberry.agents.dynprog import ValueIterationAgent
# env = FourRoom(reward_free=False, difficulty=0, array_observation=False)
# env = SixRoom(reward_free=False, array_observation=False)
env = NRoom(nrooms=9,
remove_walls=False,
room_size=9,
initial_state_distribution='center',
include_traps=True)
horizon = env.observation_space.n
agent = ValueIterationAgent(env, gamma=0.999, horizon=horizon)
print("fitting...")
info = agent.fit()
print(info)
env.enable_rendering()
for _ in range(10):
state = env.reset()
for tt in range(horizon):
# action = agent.policy(state)
action = env.action_space.sample()
next_s, _, done, _ = env.step(action)
if done:
break
state = next_s