def run_task(*_):
env = normalize(SimpleHumanoidEnv())
policy = DeterministicMLPPolicy(
env_spec=env.spec,
# The neural network policy should have two hidden layers, each with 32 hidden units.
hidden_sizes=(H_layer_first[h], H_layer_second[h]))
es = OUStrategy(env_spec=env.spec)
qf = ContinuousMLPQFunction(env_spec=env.spec)
algo = DDPG(
env=env,
policy=policy,
es=es,
qf=qf,
batch_size=size_of_batch,
max_path_length=100,
epoch_length=1000,
min_pool_size=10000,
n_epochs=number_of_episodes,
discount=discount_factor,
scale_reward=reward_scaling[r],
qf_learning_rate=critic_learning_rate[c],
policy_learning_rate=actor_learning_rate[c],
# Uncomment both lines (this and the plot parameter below) to enable plotting
# plot=True,
)
algo.train()
def run_task(*_):
env = normalize(SimpleHumanoidEnv())
policy = DeterministicMLPPolicy(
env_spec=env.spec,
# The neural network policy should have two hidden layers, each with 32 hidden units.
hidden_sizes=(100, 100)
)
es = OUStrategy(env_spec=env.spec)
qf = ContinuousMLPQFunction(
env_spec=env.spec
hidden_sizes =(100,100)
)
algo = DDPG(
env=env,
policy=policy,
es=es,
qf=qf,
batch_size=batch_size_value,
max_path_length=100,
epoch_length=1000,
min_pool_size=10000,
n_epochs=number_of_episodes,
discount=0.99,
scale_reward=0.01,
qf_learning_rate=0.001,
policy_learning_rate=0.0001,
# Uncomment both lines (this and the plot parameter below) to enable plotting
# plot=True,
)
algo.train()
def run_task(*_):
env = normalize(SimpleHumanoidEnv())
# env = SimpleHumanoidEnv()
policy = DeterministicMLPPolicy(
env_spec=env.spec,
# The neural network policy should have two hidden layers, each with 32 hidden units.
hidden_sizes=(32, 32))
es = OUStrategy(env_spec=env.spec)
qf = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=(32, 32))
"""
Persistence Length Exploration
"""
lp = Persistence_Length_Exploration(
env=env,
qf=qf,
policy=policy,
L_p=L_p_param[l_p_ind],
b_step_size=b_step_size[b_ind],
sigma=sigma_param[s_ind],
max_exploratory_steps=max_exploratory_steps_iters,
batch_size=batch_size_value,
n_epochs=num_episodes,
scale_reward=0.01,
epoch_length=steps_per_episode,
qf_learning_rate=0.001,
policy_learning_rate=0.0001,
)
"""
DDPG
"""
algo = DDPG(
env=env,
policy=policy,
es=es,
qf=qf,
lp=lp,
batch_size=batch_size_value,
max_path_length=100,
epoch_length=steps_per_episode,
min_pool_size=10000,
n_epochs=num_episodes,
discount=0.99,
scale_reward=0.01,
qf_learning_rate=0.001,
policy_learning_rate=0.0001,
# Uncomment both lines (this and the plot parameter below) to enable plotting
# plot=True,
)
algo.train()
def get(perm):
name = perm["problem"]
if name.lower() == "cartpole":
from rllab.envs.box2d.cartpole_env import CartpoleEnv
return normalize(CartpoleEnv())
elif name.lower() == "mountain car height bonus":
from rllab.envs.box2d.mountain_car_env import MountainCarEnv
return normalize(MountainCarEnv())
elif name.lower() == "mountain car":
from rllab.envs.box2d.mountain_car_env import MountainCarEnv
return normalize(MountainCarEnv(height_bonus=0))
elif name.lower() == "gym mountain car":
from rllab.envs.gym_env import GymEnv
return normalize(GymEnv("MountainCarContinuous-v0",
record_video=False))
elif name.lower() == "pendulum":
from rllab.envs.gym_env import GymEnv
return normalize(GymEnv("Pendulum-v0", record_video=False))
elif name.lower() == "mujoco double pendulum":
from rllab.envs.mujoco.inverted_double_pendulum_env import InvertedDoublePendulumEnv
return normalize(InvertedDoublePendulumEnv())
elif name.lower() == "double pendulum":
from rllab.envs.box2d.double_pendulum_env import DoublePendulumEnv
return normalize(DoublePendulumEnv())
elif name.lower() == "hopper":
from rllab.envs.mujoco.hopper_env import HopperEnv
return normalize(HopperEnv())
elif name.lower() == "swimmer":
from rllab.envs.mujoco.swimmer_env import SwimmerEnv
return normalize(SwimmerEnv())
elif name.lower() == "2d walker":
from rllab.envs.mujoco.walker2d_env import Walker2DEnv
return normalize(Walker2DEnv())
elif name.lower() == "half cheetah":
from rllab.envs.mujoco.half_cheetah_env import HalfCheetahEnv
return normalize(HalfCheetahEnv())
elif name.lower() == "ant":
from rllab.envs.mujoco.ant_env import AntEnv
return normalize(AntEnv())
elif name.lower() == "simple humanoid":
from rllab.envs.mujoco.simple_humanoid_env import SimpleHumanoidEnv
return normalize(SimpleHumanoidEnv())
elif name.lower() == "full humanoid":
from rllab.envs.mujoco.humanoid_env import HumanoidEnv
return normalize(HumanoidEnv())
else:
raise NotImplementedError(f"Environment {name} unknown")
from rllab.algos.ddpg import DDPG from rllab.envs.normalized_env import normalize from rllab.misc.instrument import run_experiment_lite from rllab.exploration_strategies.ou_strategy import OUStrategy from rllab.policies.deterministic_mlp_policy import DeterministicMLPPolicy from rllab.q_functions.continuous_mlp_q_function import ContinuousMLPQFunction from rllab.envs.mujoco.simple_humanoid_env import SimpleHumanoidEnv env = normalize(SimpleHumanoidEnv()) # H_layer_first = [32, 100, 400] # H_layer_second = [32, 100, 300] H_layer_first = [32] H_layer_second = [32] # reward_scaling = [0.01, 0.1, 1.0] reward_scaling = [0.01] # critic_learning_rate = [1e-3, 10e-3] # actor_learning_rate = [1e-4, 10e-4] critic_learning_rate = [0.001] actor_learning_rate = [0.0001] #0.99 was originally set by rllab discount_factor = 0.99 #originally : 32 set by rllab size_of_batch = 64