def test_env_seeding(env_name):
seeding.set_global_seed(123)
env1 = gym_make(env_name)
seeding.set_global_seed(456)
env2 = gym_make(env_name)
seeding.set_global_seed(123)
env3 = gym_make(env_name)
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)
def test_env_seeding(env_name):
seeder1 = Seeder(123)
env1 = gym_make(env_name)
env1.reseed(seeder1)
seeder2 = Seeder(456)
env2 = gym_make(env_name)
env2.reseed(seeder2)
seeder3 = Seeder(123)
env3 = gym_make(env_name)
env3.reseed(seeder3)
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)
def test_dqn_agent(use_double_dqn, use_prioritized_replay):
env = gym_make("CartPole-v0")
agent = DQNAgent(
env,
learning_starts=5,
eval_interval=75,
train_interval=2,
gradient_steps=-1,
use_double_dqn=use_double_dqn,
use_prioritized_replay=use_prioritized_replay,
)
agent.fit(budget=500)
def test_dqn_agent():
env = gym_make("CartPole-v0")
params = {"n_episodes": 10}
agent = DQNAgent(env, **params)
agent.fit()
agent.policy(env.observation_space.sample())
from rlberry.envs import gym_make
from rlberry.agents import RSUCBVIAgent
from rlberry.utils.logging import configure_logging
from rlberry.wrappers import RescaleRewardWrapper
configure_logging("DEBUG")
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, n_episodes=10, gamma=0.99, horizon=200,
bonus_scale_factor=0.1, min_dist=0.2)
agent.fit()
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()
def fit(self, **kwargs):
result = self.wrapped.learn(**kwargs)
info = {} # possibly store something from results
return info
def policy(self, observation, **kwargs):
action, _state = self.wrapped.predict(observation, **kwargs)
return action
#
# Traning one agent
#
env = gym_make('CartPole-v1')
agent = A2CAgent(env, 'MlpPolicy', verbose=1)
agent.fit(total_timesteps=1000)
obs = env.reset()
for i in range(1000):
action = agent.policy(obs, deterministic=True)
obs, reward, done, info = env.step(action)
env.render()
if done:
break
env.close()
#
# Traning several agents and comparing different hyperparams
#
from rlberry.envs import gym_make
from pathlib import Path
from torch.utils.tensorboard import SummaryWriter
from rlberry.agents.dqn import DQNAgent
from rlberry.utils.logging import configure_logging
configure_logging(level="DEBUG")
env = gym_make("CartPole-v0")
agent = DQNAgent(env, n_episodes=50, exploration_kwargs={"tau": 1000})
agent.set_writer(SummaryWriter())
print(f"Running DQN on {env}")
print(f"Visualize with tensorboard by \
running:\n$tensorboard --logdir {Path(agent.writer.log_dir).parent}")
agent.fit()
for episode in range(3):
done = False
state = env.reset()
while not done:
action = agent.policy(state)
state, reward, done, _ = env.step(action)
env.render()
env.close()
seeder = Seeder(123)
# Each Seeder instance has a random number generator (rng)
# 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())