def test_gym_copy_reseeding_2():
seeder = Seeder(123)
if _GYM_INSTALLED:
gym_env = gym.make("Acrobot-v1")
# nested wrapping
env = RescaleRewardWrapper(Wrapper(Wrapper(gym_env)), (0, 1))
env.reseed(seeder)
c_env = deepcopy(env)
c_env.reseed()
if deepcopy(env).is_online():
traj1 = get_env_trajectory(env, 500)
traj2 = get_env_trajectory(c_env, 500)
assert not compare_trajectories(traj1, traj2)
def test_rescale_wrapper_seeding(ModelClass):
seeding.set_global_seed(123)
env1 = RescaleRewardWrapper(ModelClass(), (0, 1))
seeding.set_global_seed(456)
env2 = RescaleRewardWrapper(ModelClass(), (0, 1))
seeding.set_global_seed(123)
env3 = RescaleRewardWrapper(ModelClass(), (0, 1))
if deepcopy(env1).is_online():
traj1 = get_env_trajectory(env1, 500)
traj2 = get_env_trajectory(env2, 500)
traj3 = get_env_trajectory(env3, 500)
assert not compare_trajectories(traj1, traj2)
assert compare_trajectories(traj1, traj3)
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:
"""
=====================
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()