def test_vis2dwrapper():
env = MountainCar()
env = Vis2dWrapper(env, n_bins_obs=20, memory_size=200)
agent = RSUCBVIAgent(
env,
gamma=0.99,
horizon=200,
bonus_scale_factor=0.1,
copy_env=False,
min_dist=0.1,
)
agent.fit(budget=15)
env.plot_trajectories(show=False)
env.plot_trajectory_actions(show=False)
from rlberry.envs import Acrobot
from rlberry.agents import RSUCBVIAgent
from rlberry.utils.logging import configure_logging
from rlberry.wrappers import RescaleRewardWrapper
configure_logging("DEBUG")
env = Acrobot()
# rescale rewards to [0, 1]
env = RescaleRewardWrapper(env, (0.0, 1.0))
agent = RSUCBVIAgent(env,
n_episodes=500,
gamma=0.99,
horizon=300,
bonus_scale_factor=0.01,
min_dist=0.25)
agent.fit()
env.enable_rendering()
state = env.reset()
for tt in range(4 * agent.horizon):
action = agent.policy(state)
next_state, reward, done, _ = env.step(action)
state = next_state
env.save_video("acrobot.mp4")
from rlberry.agents import RSUCBVIAgent
from rlberry.envs.classic_control import MountainCar
from rlberry.envs.benchmarks.ball_exploration import PBall2D
from rlberry.utils.logging import configure_logging
configure_logging("DEBUG")
for env, horizon in zip([MountainCar(), PBall2D()], [170, 50]):
print("Running RS-UCBVI on %s" % env.name)
agent = RSUCBVIAgent(env,
n_episodes=1000,
gamma=0.99,
horizon=horizon,
bonus_scale_factor=0.1)
agent.fit()
env.enable_rendering()
state = env.reset()
for tt in range(200):
action = agent.policy(state)
next_state, reward, done, _ = env.step(action)
state = next_state
env.render()
"""
=====================
Demo: demo_gym_wrapper
=====================
"""
from rlberry.envs import gym_make
from rlberry.agents import RSUCBVIAgent
from rlberry.wrappers import RescaleRewardWrapper
env = gym_make("Acrobot-v1")
env.reward_range = (-1.0, 0.0) # missing in gym implementation
# rescake rewards to [0, 1]
env = RescaleRewardWrapper(env, (0.0, 1.0))
agent = RSUCBVIAgent(env,
gamma=0.99,
horizon=200,
bonus_scale_factor=0.1,
min_dist=0.2)
agent.fit(budget=10)
state = env.reset()
for tt in range(200):
action = agent.policy(state)
next_state, reward, done, _ = env.step(action)
state = next_state
env.render()
env.close()
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,
horizon=200,
bonus_scale_factor=0.1,
copy_env=False,
min_dist=0.1,
)
agent.fit(budget=100)
env.enable_rendering()
for ep in range(3):
state = env.reset()
for tt in range(agent.horizon):
action = agent.policy(state)
next_s, _, _, _ = env.step(action)
state = next_s
try:
"""
==================================================
A demo of RSUCBVI algorithm in MountainCar environment
==================================================
Illustration of how to set up an RSUCBVI algorithm in rlberry.
The environment chosen here is MountainCar environment.
.. video:: ../../video_plot_rsucbvi.mp4
:width: 600
"""
# sphinx_gallery_thumbnail_path = 'thumbnails/video_plot_rsucbvi.jpg'
from rlberry.agents import RSUCBVIAgent
from rlberry.envs.classic_control import MountainCar
env = MountainCar()
horizon = 170
print("Running RS-UCBVI on %s" % env.name)
agent = RSUCBVIAgent(env, gamma=0.99, horizon=horizon, bonus_scale_factor=0.1)
agent.fit(budget=500)
env.enable_rendering()
state = env.reset()
for tt in range(200):
action = agent.policy(state)
next_state, reward, done, _ = env.step(action)
state = next_state
video = env.save_video("_video/video_plot_rsucbvi.mp4")
# See https://numpy.org/doc/stable/reference/random/generator.html to check the
# methods available in rng.
seeder.rng.integers(5)
seeder.rng.normal()
print(type(seeder.rng))
# etc
# Environments and agents should be seeded using a single seeder,
# to ensure that their random number generators are independent.
from rlberry.envs import gym_make
from rlberry.agents import RSUCBVIAgent
env = gym_make("MountainCar-v0")
env.reseed(seeder)
agent = RSUCBVIAgent(env)
agent.reseed(seeder)
# Environments and Agents have their own seeder and rng.
# When writing your own agents and inheriring from the Agent class,
# you should use agent.rng whenever you need to generate random numbers;
# the same applies to your environments.
# This is necessary to ensure reproducibility.
print("env seeder: ", env.seeder)
print("random sample from env rng: ", env.rng.normal())
print("agent seeder: ", agent.seeder)
print("random sample from agent rng: ", agent.rng.normal())
# A seeder can spawn other seeders that are independent from it.
# This is useful to seed two different threads, using seeder1
# in the first thread, and seeder2 in the second thread.