def test_dqn_doesnt_store_invalid_transitions(self):
STEPS = 55
env = DummyEnv()
test_env = DummyEnv()
model = tf.keras.Sequential([tf.keras.layers.Dense(2, input_dim=2)])
model.compile(tf.keras.optimizers.SGD(learning_rate=0.), loss="mse")
agent = DQN.from_environment(env,
model,
discount_gamma=0.,
use_target_network=False)
rollout = Rolling(agent, env)
test_rollout = Trajectory(agent, test_env)
rollout.fit(epochs=10,
updates_per_epoch=12,
steps_per_update=STEPS,
update_batch_size=8,
testing_rollout=test_rollout,
buffer_warmup=False)
data = agent.memory_sampler.sample(-1)
np.testing.assert_array_less(data["state"], 10)
np.testing.assert_equal(
(data["state_next"] - data["state"]).sum(axis=1), 1)
def test_td3_doesnt_store_invalid_transitions(self):
STEPS = 55
env = DummyEnv(action_space="continuous")
test_env = DummyEnv(action_space="continuous")
actor_out = tf.convert_to_tensor([[0., 0.]])
critic_out = tf.convert_to_tensor([1.])
actor = arch.TestingModel(actor_out)
critic = arch.TestingModel(critic_out)
critic2 = arch.TestingModel(critic_out)
actor_target = arch.TestingModel(actor_out)
critic_target = arch.TestingModel(critic_out)
critic2_target = arch.TestingModel(critic_out)
actor.optimizer = tf.keras.optimizers.SGD(0)
critic.optimizer = tf.keras.optimizers.SGD(0)
critic2.optimizer = tf.keras.optimizers.SGD(0)
agent = TD3(actor,
actor_target,
critic,
critic_target,
critic2,
critic2_target,
discount_gamma=0.,
polyak_tau=0.,
action_minima=-1.,
action_maxima=1.,
update_actor_every=1)
rollout = Rolling(agent, env)
test_rollout = Trajectory(agent, test_env)
rollout.fit(epochs=10,
updates_per_epoch=12,
steps_per_update=STEPS,
update_batch_size=8,
testing_rollout=test_rollout,
buffer_warmup=False)
data = agent.memory_sampler.sample(-1)
np.testing.assert_array_less(data["state"], 10)
np.testing.assert_equal(
(data["state_next"] - data["state"]).sum(axis=1), 1)
def test_a2c_doesnt_store_invalid_transitions_in_td_setting(self):
STEPS = 35
env = DummyEnv()
agent = A2C.from_environment(env, discount_gamma=0.)
rollout = Rolling(agent, env)
rollout.roll(STEPS, verbose=0, push_experience=True)
data = agent.memory_sampler.sample(-1)
self.assertEqual(agent.episodes, 3)
np.testing.assert_array_less(data["state"], 10)
self.assertEqual(len(data["state"]), STEPS - 4)
activation="relu",
input_shape=input_shape,
kernel_initializer="he_uniform"),
Dense(16, activation="relu", kernel_initializer="he_uniform"),
Dense(1, activation="linear", kernel_initializer="he_uniform")
])
critic.compile(loss="mse", optimizer=Adam(5e-4))
agent = A2C(actor,
critic,
action_space=env.action_space,
memory=Experience(max_length=10000),
discount_factor_gamma=0.98,
entropy_penalty_coef=0.01)
rollout = Rolling(agent, env, config=RolloutConfig(max_steps=300))
test_rollout = Trajectory(agent, gym.make("CartPole-v1"))
rewards = []
actor_loss = []
actor_utility = []
actor_entropy = []
critic_loss = []
for episode in range(1, 1001):
episode_actor_loss = []
episode_actor_utility = []
episode_actor_entropy = []
episode_critic_loss = []
for update in range(32):
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)
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)
test_rollout = Trajectory(agent, env, RolloutConfig(testing_rollout=True))
rollout.fit(episodes=1000,
updates_per_episode=64,
step_per_update=1,
update_batch_size=32,
testing_rollout=test_rollout)
test_rollout.render(repeats=10)
test_env = Match(cfg)
ann = mlp.wide_dueling_q_network(env.observation_space.shape,
env.action_space.n,
adam_lr=1e-4)
experience = Experience(10000)
agent = DoubleDQN(ann,
env.action_space,
experience,
epsilon=1.,
epsilon_decay=1.,
epsilon_min=0.1)
rcfg = RolloutConfig(max_steps=1024, skipframes=2)
training_rollout = Rolling(agent, env, rcfg)
testing_rollout = Trajectory(agent, test_env, rcfg)
print("Filling experience...")
while experience.N < 10000:
training_rollout.roll(steps=32, verbose=0, push_experience=True)
print(f"\r{experience.N/10000:.2%} 10000/{experience.N}", end="")
print()
agent.epsilon_decay = 0.99995
logger = history.History("reward_sum", *agent.history_keys, "epsilon")
for episode in range(1, 501):
for update in range(32):
training_rollout.roll(steps=32, verbose=0, push_experience=True)
keras.layers.BatchNormalization(),
keras.layers.ReLU(),
keras.layers.MaxPool2D(), # 5
keras.layers.GlobalAveragePooling2D(), # 16
keras.layers.Dense(4, kernel_initializer="he_uniform"),
keras.layers.BatchNormalization(),
keras.layers.ReLU(),
keras.layers.Dense(2, kernel_initializer="he_uniform")
])
qnet.compile(keras.optimizers.Adam(1e-3), "mse")
agent = DQN(qnet, 2, Experience(max_length=10_000), discount_factor_gamma=0.99,
epsilon=1.0, epsilon_decay=0.99999, epsilon_min=0.3, use_target_network=True,
state_preprocessor=None)
rollout = Rolling(agent, env, RolloutConfig(skipframes=2))
test_rollout = Trajectory(agent, test_env)
rewards = deque(maxlen=100)
losses = deque(maxlen=100)
episode = 0
while 1:
episode += 1
rollout.roll(steps=4, verbose=0, push_experience=True)
if agent.memory.N < 1000:
print(f"\rFilling memory... {agent.memory.N}/1000", end="")
continue
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)
critic = Sequential([
Dense(400, activation="relu", input_shape=input_shape),
Dense(300, activation="relu"),
Dense(1, activation="linear")
])
critic.compile(loss="mse", optimizer=Adam(1e-4))
agent = A2C(actor,
critic,
action_space=num_actions,
absolute_memory_limit=10000,
discount_factor_gamma=0.99,
entropy_penalty_coef=0.05)
rollout = Rolling(agent.create_workers(1)[0], env)
test_rollout = Trajectory(agent, Lunar())
def train():
hst = history.History("reward_sum", "actor_loss", "actor_utility",
"actor_utility_std", "actor_entropy", "values",
"advantages", "critic_loss")
for episode in range(1, 1001):
for update in range(1, 4):
rollout.roll(steps=128, verbose=0, push_experience=True)
agent_history = agent.fit(batch_size=-1,
verbose=0,
reset_memory=True)
Dense(24, activation="relu", input_shape=input_shape),
Dense(24, activation="relu"),
Dense(num_actions, activation="linear")
])
policy.compile(loss="mse", optimizer=Adam(1e-4))
agent = DQN(policy,
action_space=num_actions,
memory=Experience(max_length=10000),
epsilon=1.,
epsilon_decay=1.,
epsilon_min=0.1,
discount_factor_gamma=0.99,
use_target_network=True)
rollout = Rolling(agent, env)
test_rollout = Trajectory(agent, env)
hst = history.History("reward_sum", "loss", "epsilon")
for warmup in range(1, 33):
rollout.roll(32, verbose=0, push_experience=True)
agent.epsilon_decay = 0.99999
for episode in range(1, 1001):
rollout.roll(steps=32, verbose=0, push_experience=True)
agent_history = agent.fit(updates=10, batch_size=64, verbose=0)
test_history = test_rollout.rollout(verbose=0,
push_experience=False,
render=False)
hst.record(reward_sum=test_history["reward_sum"],
