def run_task(*_):
"""Implement the run_task method needed to run experiments with rllab."""
pass_params = (env_name, sumo_params, vehicles, env_params, net_params,
initial_config, scenario)
env = GymEnv(env_name, record_video=False, register_params=pass_params)
horizon = env.horizon
env = normalize(env)
policy = GaussianGRUPolicy(env_spec=env.spec, hidden_sizes=(64, ))
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = PPO(
env=env,
policy=policy,
baseline=baseline,
batch_size=horizon * 32 * 2,
max_path_length=horizon,
# whole_paths=True,
n_itr=400,
discount=0.999,
# step_size=0.01,
)
algo.train()
def run_task(*_):
# Please note that different environments with different action spaces may
# require different policies. For example with a Discrete action space, a
# CategoricalMLPPolicy works, but for a Box action space may need to use
# a GaussianMLPPolicy (see the trpo_gym_pendulum.py example)
env = normalize(
GymEnv(env_name="LunarLanderContinuous-v2", force_reset=True))
# policy = CategoricalMLPPolicy(
# env_spec=env.spec,
# # The neural network policy should have two hidden layers, each with 32 hidden units.
# hidden_sizes=(32, 32)
# )
policy = GaussianMLPPolicy(
env_spec=env.spec,
# The neural network policy should have two hidden layers, each with 32 hidden units.
hidden_sizes=(64, 64))
baseline = LinearFeatureBaseline(env_spec=env.spec)
# max_path_length = env.horizon
algo = PPO(
env=env,
policy=policy,
baseline=baseline,
batch_size=4000,
max_path_length=300,
n_itr=10000,
discount=0.99,
# step_size=0.02,
truncate_local_is_ratio=0.2
# Uncomment both lines (this and the plot parameter below) to enable plotting
# plot=True,
)
algo.train()
def main(exp_name, ent_wt=1.0):
register_custom_envs()
env_name = 'Acrobot-v2'
env = GymEnv(env_name)
policy = GaussianMLPPolicy(env_spec=env, hidden_sizes=(64, 64))
algo = PPO(env=env,
policy=policy,
n_itr=1500,
batch_size=8000,
max_path_length=1000,
discount=0.95,
store_paths=True,
entropy_weight=ent_wt,
baseline=LinearFeatureBaseline(env_spec=env))
data_path = 'data/acrobat_data_rllab_ppo/%s/' % exp_name
os.makedirs(data_path, exist_ok=True)
logger.set_snapshot_dir(data_path)
algo.train()
logger.set_snapshot_dir(None)
def main(exp_name, ent_wt=1.0):
register_custom_envs()
env_name = 'LunarLanderContinuous-v3'
env = GymEnv(env_name)
policy = GaussianMLPPolicy(env_spec=env, hidden_sizes=(64, 64))
baseline = GaussianMLPBaseline(env_spec=env)
algo = PPO(env=env,
policy=policy,
n_itr=1500,
batch_size=8000,
max_path_length=1000,
discount=0.99,
store_paths=True,
entropy_weight=ent_wt,
baseline=baseline)
data_path = 'data/%s_data_rllab_%s/%s/' % (env_name.replace(
'-', '_'), str(algo.__class__.__name__), exp_name)
os.makedirs(data_path, exist_ok=True)
logger.set_snapshot_dir(data_path)
algo.train()
logger.set_snapshot_dir(None)
def run_task(*_):
env = normalize(GymEnv(models[k]))
baseline = LinearFeatureBaseline(env_spec=env.spec)
learn_std = True
init_std = 1
# hidden_sizes = NN_sizes[i]
# hidden_sizes=(8,)
# hidden_sizes=(32, 32)
hidden_sizes = (100, 50, 25)
policy = GaussianMLPPolicy(env_spec=env.spec,
hidden_sizes=hidden_sizes,
learn_std=learn_std,
init_std=init_std)
# =======================
# Defining the algorithm
# =======================
batch_size = 5000
n_itr = 200
gamma = .99
step_size = 0.01
# max_path_length = 96,
algo = PPO(
env=env,
policy=policy,
baseline=baseline,
batch_size=batch_size,
# max_path_length=max_path_length,
n_itr=n_itr,
discount=gamma,
step_size=step_size)
algo.train()
#!/usr/bin/python3.5
from pre_maml.envs.quad_falling_down import QuadFallingDown
from rllab.envs.normalized_env import normalize
from rllab.baselines.linear_feature_baseline import LinearFeatureBaseline
from rllab.algos.ppo import PPO
from rllab.policies.gaussian_mlp_policy import GaussianMLPPolicy
env = normalize(env=QuadFallingDown())
policy = GaussianMLPPolicy(env_spec=env.spec, )
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = PPO(
env=env,
policy=policy,
baseline=baseline,
)
algo.train()
