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)
ALGO = "DQN"
NUM_ENVS = 4
TRAJECTORY_MAX_STEPS = 200
STEPS_PER_UPDATE = 1
UPDATES_PER_EPOCH = 64
EPOCHS = 200
UPDATE_BATCH_SIZE = 100
envs = [gym.make(ENV_NAME) for _ in range(NUM_ENVS)]
test_env = gym.make(ENV_NAME)
algo = {"DQN": DQN, "DoubleDQN": DoubleDQN}[ALGO]
agent = algo.from_environment(envs[0])
rollout = MultiRolling(agent, envs, TRAJECTORY_MAX_STEPS)
test_rollout = Trajectory(agent, test_env, TRAJECTORY_MAX_STEPS)
rollout.fit(epochs=EPOCHS,
updates_per_epoch=UPDATES_PER_EPOCH,
steps_per_update=STEPS_PER_UPDATE,
update_batch_size=UPDATE_BATCH_SIZE,
warmup_buffer=True,
callbacks=[
callbacks.TrajectoryEvaluator(testing_rollout=test_rollout,
repeats=4),
callbacks.ProgressPrinter(rollout.progress_keys)
])
test_rollout.render(repeats=10)
critic = Sequential([
Dense(16, activation="relu", input_shape=input_shape),
Dense(16, activation="relu"),
Dense(1, activation="linear")
])
critic.compile(loss="mse", optimizer=Adam(CRITIC_ADAM_LR))
agent = PPO(actor,
critic,
action_space=test_env.action_space,
memory=Experience(max_length=EXPERIENCE_SIZE),
reward_discount_factor_gamma=DISCOUNT_FACTOR_GAMMA,
entropy_penalty_coef=ENTROPY_PENALTY_BETA)
rollout = MultiRolling(agent,
envs,
rollout_configs=RolloutConfig(max_steps=MAX_TIMESTEPS))
test_rollout = Trajectory(agent, gym.make("CartPole-v1"))
rewards = []
actor_loss = []
actor_utility = []
actor_kld = []
actor_entropy = []
critic_loss = []
for episode in range(1, NUM_EPISODES + 1):
roll_history = rollout.roll(steps=ROLL_TIMESTEPS,
verbose=0,
push_experience=True)
agent_history = agent.fit(epochs=FIT_EPOCHS,
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,
update_batch_size=32,
testing_rollout=test_rollout,
plot_curves=True)
test_rollout.render()
input_shape = envs[0].observation_space.shape
num_actions = envs[0].action_space.n
qnet = Sequential([Dense(16, activation="relu", input_shape=input_shape),
Dense(16, activation="relu"),
Dense(num_actions, activation="linear")])
qnet.compile(loss="mse", optimizer=Adam(1e-3))
agent = DQN(qnet,
action_space=2,
memory=Experience(max_length=10000),
epsilon=1.,
discount_factor_gamma=0.98,
use_target_network=True)
rollout = MultiRolling(agent, envs, rollout_configs=RolloutConfig(max_steps=300))
test_rollout = Trajectory(agent, gym.make("CartPole-v1"))
rewards = []
losses = []
for episode in range(1, 501):
episode_losses = []
for update in range(32):
rollout.roll(steps=4, verbose=0, push_experience=True)
agent_history = agent.fit(batch_size=32, verbose=0)
episode_losses.append(agent_history["loss"])
test_history = test_rollout.rollout(verbose=0, push_experience=False, render=False)
rewards.append(test_history["reward_sum"])
action_space=MOVES,
memory=Experience(max_length=10000),
discount_factor_gamma=0.995,
entropy_penalty_coef=0.0,
state_preprocessor=lambda state: state / 255.)
episode = 1
reward_memory = deque(maxlen=10)
step_lengths = deque(maxlen=10)
critic_losses = deque(maxlen=10)
actor_losses = deque(maxlen=10)
actor_utility = deque(maxlen=10)
actor_entropy = deque(maxlen=10)
rollout = MultiRolling(agent, envs, rollout_configs=RolloutConfig(max_steps=512, skipframes=2))
test_rollout = Trajectory(agent, test_env, config=RolloutConfig(max_steps=512, skipframes=2))
while 1:
episode_a_losses = []
episode_a_utility = []
episode_a_entropy = []
episode_c_losses = []
for update in range(32):
rollout.roll(steps=2, verbose=0, push_experience=True)
agent_history = agent.fit(batch_size=-1, verbose=0, reset_memory=True)
episode_a_losses.append(agent_history["actor_loss"])
episode_a_utility.append(agent_history["actor_utility"])
episode_a_entropy.append(agent_history["actor_entropy"])
ALGO = "SAC"
TRAJECTORY_MAX_STEPS = 100
STEPS_PER_UPDATE = 1
UPDATES_PER_EPOCH = 32
EPOCHS = 1000
UPDATE_BATCH_SIZE = 64
NUM_ENVS = 2
envs = [gym.make(ENV_NAME) for _ in range(NUM_ENVS)]
test_env = gym.make(ENV_NAME)
algo = {"DDPG": DDPG, "TD3": TD3, "SAC": SAC}[ALGO]
agent = algo.from_environment(envs[0])
rollout = MultiRolling(agent, envs, TRAJECTORY_MAX_STEPS)
test_rollout = Trajectory(agent, test_env, TRAJECTORY_MAX_STEPS)
cbs = [
callbacks.TrajectoryEvaluator(testing_rollout=test_rollout),
callbacks.ProgressPrinter(keys=rollout.progress_keys),
callbacks.TrajectoryRenderer(testing_rollout=test_rollout),
callbacks.TensorBoard(experiment_name=rollout.experiment_name)
]
rollout.fit(epochs=EPOCHS,
updates_per_epoch=UPDATES_PER_EPOCH,
steps_per_update=STEPS_PER_UPDATE,
update_batch_size=UPDATE_BATCH_SIZE,
warmup_buffer=1000)
Conv2D(16, 3, padding="same", activation="relu"), # 20
Conv2D(32, 3, strides=2, padding="same", activation="relu"), # 10
GlobalAveragePooling2D(), # 32
Dense(1, activation="linear")
])
critic.compile(loss="mse", optimizer=Adam(CRITIC_ADAM_LR))
agent = PPO(actor,
critic,
action_space=test_env.action_space,
discount_factor_gamma=DISCOUNT_GAMMA,
gae_factor_lambda=GAE_LAMBDA,
entropy_penalty_coef=ENTROPY_PENALTY_BETA)
rollout = MultiRolling(agent.dispatch_workers(NUM_PARALLEL_ENVS), envs)
test_rollout = Trajectory(agent, test_env)
rewards = []
actor_loss = []
actor_utility = []
actor_std = []
actor_kld = []
actor_entropy = []
critic_loss = []
for episode in range(1, NUM_EPISODES + 1):
roll_history = rollout.roll(steps=ROLL_TIMESTEPS,
verbose=0,
push_experience=True)
agent_history = agent.fit(epochs=FIT_EPOCHS,
from trickster.agent import A2C
from trickster.rollout import MultiRolling, Trajectory, RolloutConfig
from trickster.model import mlp
cfg = MatchConfig(canvas_size=(100, 100),
players_per_side=2,
learning_type=MatchConfig.LEARNING_TYPE_SINGLE_AGENT,
observation_type=MatchConfig.OBSERVATION_TYPE_VECTOR)
envs = [Match(cfg) for _ in range(8)]
test_env = Match(cfg)
actor, critic = mlp.wide_pg_actor_critic(envs[0].observation_space.shape,
envs[0].action_space.n,
actor_lr=1e-4,
critic_lr=1e-4)
agent = A2C(actor, critic, test_env.action_space, entropy_penalty_coef=0.1)
rcfg = RolloutConfig(max_steps=512, skipframes=2)
training_rollout = MultiRolling(agent, envs, rcfg)
testing_rollout = Trajectory(agent, test_env, rcfg)
training_rollout.fit(episodes=1000,
updates_per_episode=512,
steps_per_update=1,
testing_rollout=testing_rollout)
testing_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()