def _fit_worker(args):
"""Create and fit an agent instance"""
(
lock,
agent_handler,
agent_class,
fit_budget,
init_kwargs,
fit_kwargs,
writer,
worker_logging_level,
seeder,
) = args
# reseed external libraries
set_external_seed(seeder)
# logging level in thread
configure_logging(worker_logging_level)
# Using a lock when creating envs and agents, to avoid problems
# as here: https://github.com/openai/gym/issues/281
with lock:
if agent_handler.is_empty():
# create agent
agent = agent_class(**init_kwargs)
# seed agent
# TODO: check if extra reseeding here is necessary
agent.reseed(seeder)
agent_handler.set_instance(agent)
# set writer
if writer[0] is None:
agent_handler.set_writer(None)
elif (
writer[0] != "default"
): # 'default' corresponds to DefaultWriter created by Agent.__init__()
writer_fn = writer[0]
writer_kwargs = writer[1]
agent_handler.set_writer(writer_fn(**writer_kwargs))
# fit agent
agent_handler.fit(fit_budget, **fit_kwargs)
# Remove writer after fit (prevent pickle problems),
# unless the agent uses DefaultWriter
if not isinstance(agent_handler.writer, DefaultWriter):
agent_handler.set_writer(None)
# remove from memory to avoid pickle issues
agent_handler.dump()
# garbage collector
gc.collect()
return agent_handler
from rlberry.envs import Acrobot
from rlberry.agents import RSKernelUCBVIAgent
from rlberry.utils.logging import configure_logging
from rlberry.wrappers import RescaleRewardWrapper
configure_logging("DEBUG")
env = Acrobot()
# rescake rewards to [0, 1]
env = RescaleRewardWrapper(env, (0.0, 1.0))
agent = RSKernelUCBVIAgent(env,
n_episodes=500,
gamma=0.99,
horizon=300,
bonus_scale_factor=0.01,
min_dist=0.2,
bandwidth=0.05,
beta=1.0,
kernel_type="gaussian")
agent.fit()
env.enable_rendering()
state = env.reset()
time_before_done = 0
ended = False
for tt in range(4 * agent.horizon):
action = agent.policy(state)
next_state, reward, done, _ = env.step(action)
if not done and not ended:
__path__ = __import__('pkgutil').extend_path(__path__, __name__)
# Initialize seeding
from rlberry.seeding import seeding
seeding.set_global_seed()
# Initialize logging level
from rlberry.utils.logging import configure_logging
configure_logging(level="INFO")
import numpy as np
from rlberry.agents.cem import CEMAgent
from rlberry.envs.benchmarks.ball_exploration import PBall2D
import rlberry.seeding as seeding
from rlberry.utils.logging import configure_logging
configure_logging(level="DEBUG")
seeding.set_global_seed(123)
env = PBall2D(p=np.inf, reward_smoothness=np.array([0.8]),
reward_centers=[np.array([0.4, 0.4])])
n_episodes = 500
batch_size = 100
horizon = 25
gamma = 0.99
agent = CEMAgent(env, n_episodes, horizon, gamma, batch_size,
percentile=70, learning_rate=0.01)
agent.fit()
env.enable_rendering()
state = env.reset()
for tt in range(4*horizon):
action = agent.policy(state)
next_state, reward, done, _ = env.step(action)
state = next_state
env.render()