from rlberry.agents import AVECPPOAgent
from rlberry.envs.classic_control import MountainCar
render = False
for env, n_episodes, horizon in zip([MountainCar()], [40000], [256]):
print("Running AVECPPO on %s" % env.name)
agent = AVECPPOAgent(env,
n_episodes=n_episodes,
horizon=horizon,
gamma=0.99,
learning_rate=0.00025,
eps_clip=0.2,
k_epochs=4)
agent.fit()
if render:
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()
"""
===============================
A demo of MountainCar environment
===============================
Illustration of MountainCar environment
.. video:: ../../video_plot_montain_car.mp4
:width: 600
"""
# sphinx_gallery_thumbnail_path = 'thumbnails/video_plot_montain_car.jpg'
from rlberry.agents.mbqvi import MBQVIAgent
from rlberry.envs.classic_control import MountainCar
from rlberry.wrappers import DiscretizeStateWrapper
_env = MountainCar()
env = DiscretizeStateWrapper(_env, 20)
agent = MBQVIAgent(env, n_samples=40, gamma=0.99)
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
video = env.save_video("_video/video_plot_montain_car.mp4")
from rlberry.agents import RSUCBVIAgent
from rlberry.envs.classic_control import MountainCar
from rlberry.envs.benchmarks.ball_exploration import PBall2D
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()
"""
==================================================
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")
from rlberry.agents import A2CAgent
from rlberry.envs.classic_control import MountainCar
from rlberry.envs.benchmarks.ball_exploration import PBall2D
from rlberry.seeding import seeding
from rlberry.utils.logging import configure_logging
render = True
seeding.set_global_seed(1223)
configure_logging("DEBUG")
for env, n_episodes, horizon in zip([PBall2D(), MountainCar()], [400, 40000],
[256, 512]):
print("Running A2C on %s" % env.name)
agent = A2CAgent(env,
n_episodes=n_episodes,
horizon=horizon,
gamma=0.99,
learning_rate=0.001,
k_epochs=4)
agent.fit()
if render:
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()