from trickster.agent import PPO
from trickster.rollout import MultiRolling, Trajectory
from trickster.utility import gymic
from trickster.model import mlp
envs = [gymic.rwd_scaled_env("LunarLander-v2") for _ in range(32)]
test_env = gymic.rwd_scaled_env("LunarLander-v2")
input_shape = envs[0].observation_space.shape
num_actions = envs[0].action_space.n
actor = mlp.wide_mlp_actor_categorical(input_shape, num_actions, adam_lr=1e-4)
critic = mlp.wide_mlp_critic_network(input_shape, output_dim=1, adam_lr=1e-4)
agent = PPO(actor,
critic,
action_space=num_actions,
discount_factor_gamma=0.99,
entropy_penalty_coef=0.05)
rollout = MultiRolling(agent, envs)
test_rollout = Trajectory(agent, test_env)
rollout.fit(episodes=1000,
updates_per_episode=1,
steps_per_update=32,
update_batch_size=32,
testing_rollout=test_rollout,
plot_curves=True)
test_rollout.render(repeats=10)
from trickster.agent import TD3
from trickster.rollout import Trajectory, Rolling, RolloutConfig
from trickster.utility import spaces, gymic
from trickster.experience import Experience
from trickster.model import mlp
env = gymic.rwd_scaled_env("LunarLanderContinuous-v2", reward_scale=0.01)
input_shape = env.observation_space.shape
num_actions = env.action_space.shape[0]
actor, critics = mlp.wide_ddpg_actor_critic(input_shape,
output_dim=num_actions,
action_range=2,
num_critics=2,
actor_lr=5e-4,
critic_lr=5e-4)
agent = TD3(actor,
critics,
action_space=spaces.CONTINUOUS,
memory=Experience(max_length=int(1e4)),
discount_factor_gamma=0.99,
action_noise_sigma=0.1,
action_noise_sigma_decay=1.,
action_minima=-2,
action_maxima=2,
target_noise_sigma=0.2,
target_noise_clip=0.5)
rollout = Rolling(agent, env)
from trickster.agent import DoubleDQN
from trickster.rollout import Rolling, Trajectory, RolloutConfig
from trickster.experience import Experience
from trickster.model import mlp
from trickster.utility import gymic
env = gymic.rwd_scaled_env("CartPole-v1")
test_env = gymic.rwd_scaled_env("CartPole-v1")
input_shape = env.observation_space.shape
num_actions = env.action_space.n
ann = mlp.wide_mlp_critic_network(input_shape, num_actions, adam_lr=1e-3)
agent = DoubleDQN(ann,
action_space=env.action_space,
memory=Experience(max_length=10000),
epsilon=1.,
epsilon_decay=0.99995,
epsilon_min=0.1,
discount_factor_gamma=0.98)
rollout = Rolling(agent, env, config=RolloutConfig(max_steps=300))
test_rollout = Trajectory(agent, test_env)
rollout.fit(episodes=500, updates_per_episode=32, step_per_update=2, update_batch_size=32,
testing_rollout=test_rollout, plot_curves=True)
test_rollout.render(repeats=10)
from trickster.agent import A2C
from trickster.rollout import Rolling, Trajectory, RolloutConfig
from trickster.model import mlp
from trickster.utility import gymic
env = gymic.rwd_scaled_env()
input_shape = env.observation_space.shape
num_actions = env.action_space.n
actor, critic = mlp.wide_pg_actor_critic(input_shape, num_actions)
agent = A2C(actor,
critic,
action_space=env.action_space,
discount_factor_gamma=0.98,
entropy_penalty_coef=0.01)
rollout = Rolling(agent, env, config=RolloutConfig(max_steps=300))
test_rollout = Trajectory(agent, gymic.rwd_scaled_env())
rollout.fit(episodes=1000, updates_per_episode=64, step_per_update=1, testing_rollout=test_rollout, plot_curves=True)
test_rollout.render(repeats=10)
from trickster.agent import DQN
from trickster.experience import Experience
from trickster.rollout import MultiRolling, RolloutConfig, Trajectory
from trickster.utility import gymic
from trickster.model import mlp
envs = [gymic.rwd_scaled_env("CartPole-v1") for _ in range(8)]
test_env = gymic.rwd_scaled_env("CartPole-v1")
input_shape = envs[0].observation_space.shape
num_actions = envs[0].action_space.n
ann = mlp.wide_mlp_critic_network(input_shape, num_actions, adam_lr=1e-4)
agent = DQN(ann,
action_space=2,
memory=Experience(max_length=10000),
epsilon=1.,
epsilon_decay=0.99995,
epsilon_min=0.1,
discount_factor_gamma=0.98,
use_target_network=True)
rollout = MultiRolling(agent,
envs,
rollout_configs=RolloutConfig(max_steps=300))
test_rollout = Trajectory(agent, test_env)
rollout.fit(episodes=500,
updates_per_episode=128,
steps_per_update=1,
from trickster.agent import REINFORCE from trickster.rollout import Trajectory, RolloutConfig from trickster.utility import gymic from trickster.model import mlp env = gymic.rwd_scaled_env() input_shape = env.observation_space.shape num_actions = env.action_space.n policy = mlp.wide_mlp_actor_categorical(input_shape, num_actions, adam_lr=1e-4) agent = REINFORCE(policy, action_space=num_actions) rollout = Trajectory(agent, env, config=RolloutConfig(max_steps=300)) rollout.fit(episodes=500, rollouts_per_update=1, update_batch_size=-1) rollout.render(repeats=10)
from trickster.agent import A2C
from trickster.rollout import MultiRolling, Trajectory
from trickster.utility import gymic
from trickster.model import mlp
envs = [gymic.rwd_scaled_env() for _ in range(8)]
input_shape = envs[0].observation_space.shape
num_actions = envs[0].action_space.n
actor, critic = mlp.wide_pg_actor_critic(input_shape,
num_actions,
critic_lr=5e-4)
agent = A2C(actor,
critic,
action_space=envs[0].action_space,
discount_factor_gamma=0.98,
entropy_penalty_coef=0.05)
rollout = MultiRolling(agent, envs)
test_rollout = Trajectory(agent, gymic.rwd_scaled_env())
rollout.fit(episodes=300,
updates_per_episode=128,
steps_per_update=1,
testing_rollout=test_rollout)
test_rollout.render()
import keras
from trickster.agent import REINFORCE
from trickster.rollout import Trajectory
from trickster.experience import Experience
from trickster.utility import gymic
from trickster.model import mlp
env = gymic.rwd_scaled_env("LunarLander-v2")
input_shape = env.observation_space.shape
num_actions = env.action_space.n
policy = mlp.wide_mlp_actor_categorical(input_shape, num_actions)
policy.compile(optimizer=keras.optimizers.SGD(lr=2e-4, momentum=0.9),
loss="categorical_crossentropy")
agent = REINFORCE(policy,
action_space=num_actions,
memory=Experience(max_length=10000),
discount_factor_gamma=0.99)
rollout = Trajectory(agent, env)
rollout.fit(episodes=1000, rollouts_per_update=16, update_batch_size=-1)
rollout.render(repeats=10)