def main():
env = NormalizedGymEnv('BeamRiderNoFrameskip-v4', is_atari=True)
# agent = DDQNAgent(
# env=env,
# discount_factor=0.99,
# policy_network_params={'dueling': True},
# double_q=True,
# prioritized_replay=True,
# max_timesteps=int(10e6),
# seed=0)
# agent.train()
env = NormalizedGymEnv('CartPole-v0')
net = partial(mlp, hidden_units=[64])
agent = DDQNAgent(env=env,
policy_network=net,
policy_network_params={},
policy_learning_rate=1e-3,
max_timesteps=100000,
buffer_size=50000,
checkpoint_freq=1000,
exploration_fraction=0.3,
exploration_final_eps=0.05,
prioritized_replay=True,
double_q=True,
train_freq=1,
target_network_update_freq=500,
learning_start_timestep=1000,
seed=65)
agent.train()
def test_continuous_action_space_norm():
env = NormalizedGymEnv('MountainCarContinuous-v0')
env.reset()
env.step([0.1])
env.close()
assert GymEnv.env_action_space_is_discrete(env) is False
assert GymEnv.get_env_action_space_dim(env) == 1
def test_discrete_action_space_norm():
env = NormalizedGymEnv('CartPole-v0')
env.reset()
env.step(1)
env.close()
assert GymEnv.env_action_space_is_discrete(env) is True
assert GymEnv.get_env_action_space_dim(env) == 2
def _get_env(self, job_dir):
env_name = self._spec_dict['env']['name']
kwargs = {
k: v
for k, v in self._spec_dict['env'].items()
if k != 'name' and k != 'is_parallel'
}
kwargs['video'] = self._video
if not self._spec_dict['env'].get('is_parallel', False):
env = NormalizedGymEnv(env_name,
log_dir=job_dir,
force_reset=True,
**kwargs)
else:
env = ParallelEnvs(env_name,
log_dir=job_dir,
force_reset=True,
**kwargs)
if 'timestep_limit' in self._spec_dict['env']:
env.spec.timestep_limit = self._spec_dict['env']['timestep_limit']
return env
def main():
env = NormalizedGymEnv(
# 'MountainCarContinuous-v0'
'CartPole-v1'
# 'Acrobot-v1'
# 'Pendulum-v0'
# 'HalfCheetah-v1'
)
agent = CEMAgent(env,
discount_factor=0.99,
n_weight_samples=100,
init_var=1.,
best_pct=0.2,
policy_network=None,
policy_network_params={},
model_file_path=None,
min_std=1e-6,
init_std=1.0,
adaptive_std=False,
seed=5)
agent.train(100)
def main():
env = NormalizedGymEnv(
'MountainCarContinuous-v0',
# 'Walker2d-v1',
# 'Swimmer-v1',
# 'CartPole-v1',
# 'Acrobot-v1',
# 'Pendulum-v0',
# 'PongNoFrameskip-v4',
# normalize_obs=True,
# is_atari=True
)
print(env.action_space)
agent = TRPOAgent(env,
policy_network=mlp,
seed=123,
baseline_train_iters=5,
baseline_model_learning_rate=1e-3,
baseline_network=mlp)
agent.train(max_timesteps=1000000)
def main():
env = NormalizedGymEnv(
'MountainCarContinuous-v0'
# 'CartPole-v1'
# 'Walker2d-v1'
# 'Acrobot-v1'
# 'Pendulum-v0'
# 'HalfCheetah-v1'
# normalize_obs=True
)
print(env.action_space)
agent = REINFORCEAgent(
env=env, discount_factor=0.99,
policy_network=mlp,
policy_learning_rate=0.01,
entropy_weight=0,
baseline_train_iters=5,
baseline_model_learning_rate=1e-3,
baseline_network=mlp,
# baseline_network=None,
seed=5)
agent.train(10000, n_steps=1024)
"""
Regression tests for the REINFORCE agent on OpenAI gym environments
"""
import pytest
import numpy as np
import shutil
from yarlp.utils.env_utils import NormalizedGymEnv
from yarlp.agent.ddqn_agent import DDQNAgent
env = NormalizedGymEnv('PongNoFrameskip-v4', is_atari=True)
def test_ddqn():
agent = DDQNAgent(env,
max_timesteps=10,
learning_start_timestep=1,
train_freq=5,
batch_size=1)
agent.train()
def test_seed():
agent = DDQNAgent(env, seed=143, max_timesteps=2)
agent.train()
ob, *_ = agent.replay_buffer.sample(1)
agent = DDQNAgent(env, seed=143, max_timesteps=2)
agent.train()
ob2, *_ = agent.replay_buffer.sample(1)
def test_norm_reward():
env = NormalizedGymEnv('MountainCarContinuous-v0', normalize_rewards=True)
env.reset()
[env.step([0.01]) for _ in range(env.spec.timestep_limit + 1)]