def __init__(self,
reward_free=False,
difficulty=0,
array_observation=False):
self.reward_free = reward_free
self.difficulty = difficulty
self.array_observation = array_observation
if difficulty not in [0, 1, 2]:
raise ValueError("FourRoom difficulty must be in [0, 1, 2]")
# Common parameters
nrows = 9
ncols = 9
start_coord = (0, 0)
terminal_states = ((8, 0), )
success_probability = 0.95
#
walls = ()
for ii in range(9):
if ii not in [2, 6]:
walls += ((ii, 4), )
for jj in range(9):
if jj != 7:
walls += ((4, jj), )
# Default reward according to the difficulty
if difficulty in [0, 1]:
default_reward = 0.0
elif difficulty == 2:
default_reward = -0.005
# Rewards according to the difficulty
if self.reward_free:
reward_at = {}
else:
if difficulty == 0:
reward_at = {(8, 0): 1.0}
elif difficulty in [1, 2]:
reward_at = {
(8, 0): 1.0,
(3, 3): 0.1,
}
# Init base class
GridWorld.__init__(
self,
nrows=nrows,
ncols=ncols,
start_coord=start_coord,
terminal_states=terminal_states,
success_probability=success_probability,
reward_at=reward_at,
walls=walls,
default_reward=default_reward,
)
# spaces
if self.array_observation:
self.observation_space = spaces.Box(0.0, 1.0, shape=(2, ))
def test_lsvi_random_exploration():
env = GridWorld(nrows=2, ncols=2, walls=(), success_probability=0.95)
env.reseed(123)
def feature_map_fn(_env):
return OneHotFeatureMap(_env.observation_space.n, _env.action_space.n)
agent = LSVIUCBAgent(
env,
feature_map_fn=feature_map_fn,
horizon=20,
gamma=0.99,
reg_factor=1e-5,
bonus_scale_factor=0.0,
)
agent.reseed(123)
agent.fit(budget=250)
# estimated Q
S = env.observation_space.n
Q_est = agent._run_lsvi(bonus_factor=0.0)[0, :].reshape((S, -1))
# near optimal Q
agent_opt = ValueIterationAgent(env, gamma=0.99, horizon=20)
agent_opt.fit()
Q = agent_opt.Q[0, :, :]
print(Q)
print("---")
print(Q_est)
print("-------")
print(np.abs(Q - Q_est))
# Check error
assert np.abs(Q - Q_est).mean() < 0.1
def __init__(self, reward_free=False, array_observation=False):
self.reward_free = reward_free
self.array_observation = array_observation
# Common parameters
nrows = 13
ncols = 17
start_coord = (5, 1)
terminal_states = ((7, 7),)
success_probability = 0.95
#
walls = ()
for ii in range(13):
walls += ((ii, 0),)
walls += ((ii, 16),)
for jj in range(17):
walls += ((0, jj),)
walls += ((12, jj),)
for ii in range(13):
if ii not in [1, 11]:
walls += ((ii, 6),)
walls += ((ii, 10),)
walls += ((11, 6),)
for jj in range(17):
if jj not in [1, 15]:
walls += ((6, jj),)
# Default reward according to the difficulty
default_reward = 0
# Rewards according to the difficulty
if self.reward_free:
reward_at = {}
else:
reward_at = {
(7, 7): 10.0,
(8, 2): 1.0,
(10, 3): 1.0
}
for jj in range(7, 16):
for ii in range(1, 12):
if (ii, jj) not in walls and (ii, jj) != (7, 7):
reward_at[(ii, jj)] = -0.05
# Init base class
GridWorld.__init__(self,
nrows=nrows,
ncols=ncols,
start_coord=start_coord,
terminal_states=terminal_states,
success_probability=success_probability,
reward_at=reward_at,
walls=walls,
default_reward=default_reward)
# spaces
if self.array_observation:
self.observation_space = spaces.Box(0.0, 1.0, shape=(2,))
def test_gridworld_aux_functions():
env = GridWorld(nrows=5,
ncols=5,
walls=((1, 1), ),
reward_at={
(4, 4): 1,
(4, 3): -1
})
env.log() # from FiniteMDP
env.render_ascii() # from GridWorld
vals = np.ones(env.observation_space.n)
env.display_values(vals)
env.print_transition_at(0, 0, 'up')
def __init__(self, reward_free=False, array_observation=False):
self.reward_free = reward_free
self.array_observation = array_observation
# Common parameters
nrows = 11
ncols = 17
start_coord = (0, 0)
terminal_states = ((10, 0), )
success_probability = 0.95
#
walls = ()
for ii in range(11):
if ii not in [2, 8]:
walls += ((ii, 5), )
walls += ((ii, 11), )
for jj in range(17):
if jj != 15:
walls += ((5, jj), )
# Default reward according to the difficulty
default_reward = -0.001
# Rewards according to the difficulty
if self.reward_free:
reward_at = {}
else:
reward_at = {
(10, 0): 10.0,
(4, 4): 0.1,
}
# Init base class
GridWorld.__init__(
self,
nrows=nrows,
ncols=ncols,
start_coord=start_coord,
terminal_states=terminal_states,
success_probability=success_probability,
reward_at=reward_at,
walls=walls,
default_reward=default_reward,
)
# spaces
if self.array_observation:
self.observation_space = spaces.Box(0.0, 1.0, shape=(2, ))
def test_lsvi_optimism():
env = GridWorld(nrows=2, ncols=2, walls=())
def feature_map_fn(_env):
return OneHotFeatureMap(_env.observation_space.n, _env.action_space.n)
agent = LSVIUCBAgent(env, n_episodes=250, gamma=0.99,
feature_map_fn=feature_map_fn,
horizon=3,
bonus_scale_factor=3,
reg_factor=0.000001)
agent.fit()
# near optimal Q
agent_opt = ValueIterationAgent(env, gamma=0.99, horizon=3)
agent_opt.fit()
Q = agent_opt.Q[0, :, :]
# optimistic Q
S = env.observation_space.n
A = env.action_space.n
Q_optimistic = np.zeros((S, A))
for ss in range(S):
Q_optimistic[ss, :] = agent._compute_q_vec(
agent.w_vec[0, :],
ss,
agent.bonus_scale_factor)
print(Q)
print(Q_optimistic)
assert (Q_optimistic - Q).min() >= -1e-5
def test_lsvi_ucb_matrix_inversion(FeatMapClass):
env = GridWorld(nrows=3, ncols=3, walls=())
def feature_map_fn(_env):
return FeatMapClass(_env.observation_space.n, _env.action_space.n)
reg_factor = 0.1
agent = LSVIUCBAgent(env, n_episodes=50,
feature_map_fn=feature_map_fn,
horizon=10,
reg_factor=reg_factor)
agent.fit()
assert np.allclose(np.linalg.inv(agent.lambda_mat), agent.lambda_mat_inv)
assert agent.episode == 50
agent.policy(env.observation_space.sample())
# Check counts
if FeatMapClass != OneHotFeatureMap:
return
S = env.observation_space.n
A = env.action_space.n
N_sa = np.zeros((S, A))
for state, action in zip(agent.state_hist, agent.action_hist):
N_sa[state, action] += 1.0
assert np.allclose(agent.lambda_mat_inv.diagonal(),
1.0/(N_sa.flatten()+reg_factor))
for ss in range(S):
for aa in range(A):
feat = agent.feature_map.map(ss, aa)
assert np.allclose(feat @ (agent.lambda_mat_inv.T @ feat),
1.0/(N_sa[ss, aa]+reg_factor))
def test_rlsvi(gamma, stage_dependent):
env = GridWorld(walls=(), nrows=5, ncols=5)
agent = RLSVIAgent(env,
horizon=11,
stage_dependent=stage_dependent,
gamma=gamma)
agent.fit(budget=50)
agent.policy(env.observation_space.sample())
def test_optql():
env = GridWorld(walls=(), nrows=5, ncols=5)
agent = OptQLAgent(env,
n_episodes=50,
horizon=11,
gamma=0.99,
bonus_scale_factor=0.1)
agent.fit()
agent.policy(env.observation_space.sample())
def test_gridworld_from_layout():
layout = """
IOOOO # OOOOO O OOOOR
OOOOO # OOOOO # OOOOO
OOOOO O OOOOO # OOTOO
OOOOO # OOOOO # OOOOO
IOOOO # OOOOO # OOOOr"""
env = GridWorld.from_layout(layout)
env.reset()
def test_discrete2onehot():
env = DiscreteToOneHotWrapper(GridWorld())
env.reseed(123)
assert isinstance(env.observation_space, spaces.Box)
for ii in range(env.unwrapped.observation_space.n):
initial_distr = np.zeros(env.unwrapped.observation_space.n)
initial_distr[ii] = 1.0
env.unwrapped.set_initial_state_distribution(initial_distr)
obs = env.reset()
assert np.array_equal(obs, initial_distr)
def test_ucbvi(gamma, stage_dependent, real_time_dp):
env = GridWorld(walls=(), nrows=5, ncols=5)
agent = UCBVIAgent(env,
n_episodes=50,
horizon=11,
stage_dependent=stage_dependent,
gamma=gamma,
real_time_dp=real_time_dp,
bonus_scale_factor=0.1)
agent.fit()
agent.policy(env.observation_space.sample())
def test_ucbvi(gamma, stage_dependent, bernoullized_reward):
env = GridWorld(walls=(), nrows=5, ncols=5)
agent = PSRLAgent(
env,
horizon=11,
bernoullized_reward=bernoullized_reward,
stage_dependent=stage_dependent,
gamma=gamma,
)
agent.fit(budget=50)
agent.policy(env.observation_space.sample())
def test_lsvi_ucb_matrix_inversion(FeatMapClass):
env = GridWorld(nrows=3, ncols=3, walls=())
def feature_map_fn():
return FeatMapClass(env.observation_space.n, env.action_space.n)
agent = LSVIUCBAgent(env, n_episodes=10,
feature_map_fn=feature_map_fn,
horizon=10)
agent.fit()
assert np.allclose(np.linalg.inv(agent.lambda_mat), agent.lambda_mat_inv)
assert agent.episode == 10
agent.policy(env.observation_space.sample())
def test_uncertainty_est_wrapper():
env = GridWorld()
def uncertainty_est_fn(observation_space, action_space):
return DiscreteCounter(observation_space, action_space)
w_env = UncertaintyEstimatorWrapper(env,
uncertainty_est_fn,
bonus_scale_factor=1.0)
for ii in range(10):
w_env.reset()
_, _, _, info = w_env.step(0)
nn = w_env.uncertainty_estimator.count(0, 0)
assert nn == ii + 1
assert info["exploration_bonus"] == pytest.approx(1 / np.sqrt(nn))
def test_mc_policy_eval(gamma, horizon, stationary_policy):
env = GridWorld(nrows=3,
ncols=3,
start_coord=(0, 0),
success_probability=1.0,
walls=(),
default_reward=0.0,
reward_at={(2, 2): 1.0})
agent = ValueIterationAgent(env, gamma=gamma, horizon=horizon)
agent.fit()
episode_rewards = mc_policy_evaluation(agent,
env,
n_sim=5,
gamma=gamma,
stationary_policy=stationary_policy)
assert episode_rewards.mean() == 1.0 * np.power(gamma, 4)
def _get_filled_replay(max_replay_size):
"""runs env for ~ 2 * max_replay_size timesteps."""
env = GridWorld(terminal_states=None)
env = TimeLimit(env, max_episode_steps=200)
env.reseed(123)
rng = np.random.default_rng(456)
buffer = replay.ReplayBuffer(
max_replay_size,
rng,
max_episode_steps=env._max_episode_steps,
enable_prioritized=True,
)
buffer.setup_entry("observations", np.float32)
buffer.setup_entry("actions", np.uint32)
buffer.setup_entry("rewards", np.float32)
buffer.setup_entry("dones", bool)
# Fill the replay buffer
total_time = 0
while True:
if total_time > 2 * buffer._max_replay_size:
break
done = False
obs = env.reset()
while not done:
total_time += 1
action = env.action_space.sample()
next_obs, reward, done, _ = env.step(action)
buffer.append({
"observations": obs,
"actions": action,
"rewards": reward,
"dones": done,
})
obs = next_obs
if done:
buffer.end_episode()
return buffer, env
def reset(self):
self.state = GridWorld.reset(self)
state_to_return = self.state
if self.array_observation:
state_to_return = self._convert_index_to_float_coord(self.state)
return state_to_return
from rlberry.agents.ucbvi import UCBVIAgent
from rlberry.agents.optql import OptQLAgent
from rlberry.envs.finite import GridWorld
from rlberry.stats import AgentStats, plot_episode_rewards
from rlberry.stats import MultipleStats
N_EP = 3000
HORIZON = 20
GAMMA = 1.0
env = GridWorld(nrows=5, ncols=10)
params = {}
params['ucbvi'] = {
'n_episodes': N_EP,
'horizon': HORIZON,
'stage_dependent': True,
'gamma': GAMMA,
'real_time_dp': True,
'bonus_scale_factor': 1.0,
}
params['optql'] = {
'n_episodes': N_EP,
'horizon': HORIZON,
'gamma': GAMMA,
'bonus_scale_factor': 1.0,
}
mstats = MultipleStats()
def test_lsvi_without_bonus():
seeding.set_global_seed(123)
def lsvi_debug_gather_data(agent):
"""
Function to gather data sampling uniformly
states and actions
"""
N = agent.n_episodes*agent.horizon
count = 0
while count < N:
state = agent.env.observation_space.sample()
action = agent.env.action_space.sample()
next_state, reward, done, info = agent.env.sample(state, action)
#
#
feat = agent.feature_map.map(state, action)
outer_prod = np.outer(feat, feat)
inv = agent.lambda_mat_inv
#
agent.lambda_mat += np.outer(feat, feat)
# update inverse
agent.lambda_mat_inv -= \
(inv @ outer_prod @ inv) / (1 + feat @ inv.T @ feat)
# update history
agent.reward_hist[count] = reward
agent.state_hist.append(state)
agent.action_hist.append(action)
agent.nstate_hist.append(next_state)
#
tt = agent.total_time_steps
agent.feat_hist[tt, :] = agent.feature_map.map(state, action)
for aa in range(agent.env.action_space.n):
agent.feat_ns_all_actions[tt, aa, :] = \
agent.feature_map.map(next_state, aa)
# increments
agent.total_time_steps += 1
count += 1
env = GridWorld(nrows=2, ncols=2, walls=(), success_probability=0.95)
def feature_map_fn(_env):
return OneHotFeatureMap(_env.observation_space.n, _env.action_space.n)
agent = LSVIUCBAgent(env, n_episodes=100,
feature_map_fn=feature_map_fn,
horizon=20,
gamma=0.99,
reg_factor=1e-5)
lsvi_debug_gather_data(agent)
# estimated Q
S = env.observation_space.n
Q_est = agent._run_lsvi(bonus_factor=0.0)[0, :].reshape((S, -1))
# near optimal Q
agent_opt = ValueIterationAgent(env, gamma=0.99, horizon=20)
agent_opt.fit()
Q = agent_opt.Q[0, :, :]
print(Q)
print("---")
print(Q_est)
print("-------")
print(np.abs(Q-Q_est))
# Check error
assert Q_est == pytest.approx(Q, rel=0.01)
from rlberry.envs.finite import GridWorld
env = GridWorld(7, 10, walls=((2, 2), (3, 3)))
env.enable_rendering()
for tt in range(50):
env.step(env.action_space.sample())
env.render()
from rlberry.agents.dynprog import ValueIterationAgent
from rlberry.envs.finite import GridWorld
env = GridWorld(7, 10, walls=((2, 2), (3, 3)))
agent = ValueIterationAgent(env, gamma=0.95)
info = agent.fit()
print(info)
env.enable_rendering()
state = env.reset()
for tt in range(200):
action = agent.policy(state)
next_s, _, done, _ = env.step(action)
if done:
break
state = next_s
env.save_video("gridworld.mp4", framerate=5)
import numpy as np
from rlberry.agents.features import FeatureMap
from rlberry.envs.finite import GridWorld
from rlberry.stats import AgentStats, plot_episode_rewards,\
compare_policies
from rlberry.agents.dynprog import ValueIterationAgent
from rlberry.agents.linear import LSVIUCBAgent
# Define environment
env = GridWorld(nrows=2, ncols=4, walls=(), success_probability=1.0)
# Create feature map
class OneHotFeatureMap(FeatureMap):
def __init__(self, S, A):
self.S = env.observation_space.n
self.A = env.action_space.n
self.shape = (S * A, )
def map(self, observation, action):
feat = np.zeros((self.S, self.A))
feat[observation, action] = 1.0
return feat.flatten()
# Function that returns an instance of a feature map
def feature_map_fn(env):
return OneHotFeatureMap(env.observation_space.n, env.action_space.n)
params = {
def test_gridworld_aux_functions():
env = GridWorld(nrows=5,
ncols=8,
walls=((1, 1), ),
reward_at={
(4, 4): 1,
(4, 3): -1
})
env.log() # from FiniteMDP
env.render_ascii() # from GridWorld
vals = np.arange(env.observation_space.n)
env.display_values(vals)
env.print_transition_at(0, 0, "up")
layout = env.get_layout_array(vals, fill_walls_with=np.inf)
for rr in range(env.nrows):
for cc in range(env.ncols):
if (rr, cc) in env.walls:
assert layout[rr, cc] == np.inf
else:
assert layout[rr, cc] == vals[env.coord2index[(rr, cc)]]
def __init__(
self,
nrooms=7,
reward_free=False,
array_observation=False,
room_size=5,
success_probability=0.95,
remove_walls=False,
initial_state_distribution="center",
include_traps=False,
):
assert nrooms > 0, "nrooms must be > 0"
assert initial_state_distribution in ("center", "uniform")
self.reward_free = reward_free
self.array_observation = array_observation
self.nrooms = nrooms
self.room_size = room_size
self.success_probability = success_probability
self.remove_walls = remove_walls
self.initial_state_distribution = initial_state_distribution
self.include_traps = include_traps
# Max number of rooms/columns per row
self.max_rooms_per_row = 5
# Room size (default = 5x5)
self.room_size = room_size
# Grid size
self.room_nrows = math.ceil(nrooms / self.max_rooms_per_row)
if self.room_nrows > 1:
self.room_ncols = self.max_rooms_per_row
else:
self.room_ncols = nrooms
nrows = self.room_size * self.room_nrows + (self.room_nrows - 1)
ncols = self.room_size * self.room_ncols + (self.room_ncols - 1)
# # walls
walls = []
for room_col in range(self.room_ncols - 1):
col = (room_col + 1) * (self.room_size + 1) - 1
for jj in range(nrows):
if (jj % (self.room_size + 1)) != (self.room_size // 2):
walls.append((jj, col))
for room_row in range(self.room_nrows - 1):
row = (room_row + 1) * (self.room_size + 1) - 1
for jj in range(ncols):
walls.append((row, jj))
# process each room
start_coord = None
terminal_state = None
self.traps = []
count = 0
for room_r in range(self.room_nrows):
if room_r % 2 == 0:
cols_iterator = range(self.room_ncols)
else:
cols_iterator = reversed(range(self.room_ncols))
for room_c in cols_iterator:
# existing rooms
if count < self.nrooms:
# remove top wall
if ((room_c == self.room_ncols - 1) and (room_r % 2 == 0)) or (
(room_c == 0) and (room_r % 2 == 1)
):
if room_r != self.room_nrows - 1:
wall_to_remove = self._convert_room_coord_to_global(
room_r, room_c, self.room_size, self.room_size // 2
)
if wall_to_remove in walls:
walls.remove(wall_to_remove)
# rooms to remove
else:
for ii in range(-1, self.room_size + 1):
for jj in range(-1, self.room_size + 1):
wall_to_include = self._convert_room_coord_to_global(
room_r, room_c, ii, jj
)
if (
wall_to_include[0] >= 0
and wall_to_include[0] < nrows
and wall_to_include[1] >= 0
and wall_to_include[1] < ncols
and (wall_to_include not in walls)
):
walls.append(wall_to_include)
pass
# start coord
if count == nrooms // 2:
start_coord = self._convert_room_coord_to_global(
room_r, room_c, self.room_size // 2, self.room_size // 2
)
# terminal state
if count == nrooms - 1:
terminal_state = self._convert_room_coord_to_global(
room_r, room_c, self.room_size // 2, self.room_size // 2
)
# trap
if include_traps:
self.traps.append(
self._convert_room_coord_to_global(
room_r,
room_c,
self.room_size // 2 + 1,
self.room_size // 2 + 1,
)
)
count += 1
terminal_states = (terminal_state,) + tuple(self.traps)
if self.reward_free:
reward_at = {}
else:
reward_at = {
terminal_state: 1.0,
start_coord: 0.01,
(self.room_size // 2, self.room_size // 2): 0.1,
}
# Check remove_walls
if remove_walls:
walls = ()
# Init base class
GridWorld.__init__(
self,
nrows=nrows,
ncols=ncols,
start_coord=start_coord,
terminal_states=terminal_states,
success_probability=success_probability,
reward_at=reward_at,
walls=walls,
default_reward=0.0,
)
# Check initial distribution
if initial_state_distribution == "uniform":
distr = np.ones(self.observation_space.n) / self.observation_space.n
self.set_initial_state_distribution(distr)
# spaces
if self.array_observation:
self.discrete_observation_space = self.observation_space
self.observation_space = spaces.Box(0.0, 1.0, shape=(2,))
Illustration of how to set up an MBQVI algorithm in rlberry.
The environment chosen here is GridWorld environment.
.. video:: ../../video_plot_mbqvi.mp4
:width: 600
"""
# sphinx_gallery_thumbnail_path = 'thumbnails/video_plot_mbqvi.jpg'
from rlberry.agents.mbqvi import MBQVIAgent
from rlberry.envs.finite import GridWorld
params = {}
params["n_samples"] = 100 # samples per state-action pair
params["gamma"] = 0.99
params["horizon"] = None
env = GridWorld(7, 10, walls=((2, 2), (3, 3)), success_probability=0.6)
agent = MBQVIAgent(env, **params)
info = agent.fit()
print(info)
# evaluate policy in a deterministic version of the environment
env_eval = GridWorld(7, 10, walls=((2, 2), (3, 3)), success_probability=1.0)
env_eval.enable_rendering()
state = env_eval.reset()
for tt in range(50):
action = agent.policy(state)
next_s, _, _, _ = env_eval.step(action)
state = next_s
video = env_eval.save_video("_video/video_plot_mbqvi.mp4")
from rlberry.agents.dynprog import ValueIterationAgent
from rlberry.envs.finite import GridWorld, Chain
for env in [Chain(), GridWorld(7, 10, walls=((2, 2), (3, 3)))]:
agent = ValueIterationAgent(env, gamma=0.95)
info = agent.fit()
print(info)
env.enable_rendering()
state = env.reset()
for tt in range(50):
action = agent.policy(state)
next_s, _, done, _ = env.step(action)
if done:
break
state = next_s
env.render()
