def test_tnpg_inverted_pendulum(self):
"""Test TNPG with InvertedPendulum-v2 environment."""
with TFTrainer(snapshot_config, sess=self.sess) as trainer:
env = normalize(GymEnv('InvertedPendulum-v2'))
policy = GaussianMLPPolicy(name='policy',
env_spec=env.spec,
hidden_sizes=(32, 32))
baseline = LinearFeatureBaseline(env_spec=env.spec)
sampler = LocalSampler(
agents=policy,
envs=env,
max_episode_length=env.spec.max_episode_length,
is_tf_worker=True)
algo = TNPG(env_spec=env.spec,
policy=policy,
baseline=baseline,
sampler=sampler,
discount=0.99,
optimizer_args=dict(reg_coeff=5e-1))
trainer.setup(algo, env)
last_avg_ret = trainer.train(n_epochs=10, batch_size=10000)
assert last_avg_ret > 15
env.close()
def test_tnpg_cartpole(self):
"""Test TNPG with Cartpole environment."""
logger.reset()
env = TfEnv(normalize(CartpoleEnv()))
policy = GaussianMLPPolicy(name="policy",
env_spec=env.spec,
hidden_sizes=(32, 32))
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = TNPG(env=env,
policy=policy,
baseline=baseline,
batch_size=10000,
max_path_length=100,
n_itr=10,
discount=0.99,
optimizer_args=dict(reg_coeff=5e-2))
last_avg_ret = algo.train(sess=self.sess)
assert last_avg_ret > 40
def test_tnpg_inverted_pendulum(self):
"""Test TNPG with InvertedPendulum-v2 environment."""
env = TfEnv(normalize(gym.make("InvertedPendulum-v2")))
policy = GaussianMLPPolicy(name="policy",
env_spec=env.spec,
hidden_sizes=(32, 32))
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = TNPG(env=env,
policy=policy,
baseline=baseline,
batch_size=10000,
max_path_length=100,
n_itr=10,
discount=0.99,
optimizer_args=dict(reg_coeff=5e-1))
last_avg_ret = algo.train(sess=self.sess)
assert last_avg_ret > 30
env.close()
def test_tnpg_inverted_pendulum(self):
"""Test TNPG with InvertedPendulum-v2 environment."""
with LocalTFRunner(snapshot_config, sess=self.sess) as runner:
env = GarageEnv(normalize(gym.make('InvertedPendulum-v2')))
policy = GaussianMLPPolicy(name='policy',
env_spec=env.spec,
hidden_sizes=(32, 32))
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = TNPG(env_spec=env.spec,
policy=policy,
baseline=baseline,
max_path_length=100,
discount=0.99,
optimizer_args=dict(reg_coeff=5e-1))
runner.setup(algo, env)
last_avg_ret = runner.train(n_epochs=10, batch_size=10000)
assert last_avg_ret > 15
env.close()