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
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)