class Tester:
def __init__(self, env_name):
self.env_name = env_name
self.action_space = gym.make(self.env_name).action_space.n
self.q_network = QNetwork(self.action_space)
self.define_network()
def define_network(self):
env = gym.make(self.env_name)
state = env.reset()
self.q_network(np.atleast_2d(state))
def test_play(self, current_weights, epsilon):
self.q_network.set_weights(current_weights)
env = gym.make(self.env_name)
state = env.reset()
episode_rewards = 0
done = False
while not done:
action = self.q_network.sample_action(state, epsilon)
next_state, reward, done, _ = env.step(action)
episode_rewards += reward
state = next_state
return episode_rewards
class Learner:
def __init__(self, gamma, env_name):
self.env_name = env_name
self.action_space = gym.make(self.env_name).action_space.n
self.q_network = QNetwork(self.action_space)
self.target_q_network = QNetwork(self.action_space)
self.gamma = gamma
self.optimizer = tf.keras.optimizers.Adam(lr=0.001)
def define_network(self):
env = gym.make(self.env_name)
state = env.reset()
self.q_network(np.atleast_2d(state))
self.target_q_network(np.atleast_2d(state))
self.target_q_network.set_weights(self.q_network.get_weights())
current_weights = self.q_network.get_weights()
return current_weights
def update_network(self, minibatchs):
indices_all = []
td_errors_all = []
for (indices, weights, transitions) in minibatchs:
states, actions, rewards, next_states, dones = zip(*transitions)
states = np.vstack(states)
actions = np.array(actions)
rewards = np.vstack(rewards)
next_states = np.vstack(next_states)
dones = np.vstack(dones)
next_qvalues = self.q_network(next_states)
next_actions = tf.cast(tf.argmax(next_qvalues, axis=1), tf.int32)
next_actions_onehot = tf.one_hot(next_actions, self.action_space)
next_maxQ = tf.reduce_sum(
next_qvalues * next_actions_onehot, axis=1, keepdims=True)
TQ = rewards + self.gamma * (1 - dones) * next_maxQ
with tf.GradientTape() as tape:
qvalues = self.q_network(states)
actions_onehot = tf.one_hot(actions, self.action_space)
Q = tf.reduce_sum(qvalues * actions_onehot, axis=1, keepdims=True)
td_errors = tf.square(TQ - Q)
loss = tf.reduce_mean(weights * td_errors)
grads = tape.gradient(loss, self.q_network.trainable_variables)
grads, _ = tf.clip_by_global_norm(grads, 40.0)
self.optimizer.apply_gradients(
zip(grads, self.q_network.trainable_variables))
indices_all += indices
td_errors_all += td_errors.numpy().flatten().tolist()
current_weights = self.q_network.get_weights()
return current_weights, indices_all, td_errors_all
class Agent():
"""Interacts with and learns from the environment."""
def __init__(self, state_shape, action_size, buffer_size, batch_size,
gamma, tau, learning_rate, update_every, seed):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
buffer_size (int): replay buffer size
batch_size (int): minibatch size
gamma (float): discount factor
tau (float): used for soft update of target parameters
learning_rate (float): learning rate
update_every (int): how many steps between network updates
device (torch.Device): pytorch device
seed (int): random seed
"""
self.state_shape = state_shape
self.action_size = action_size
self.batch_size = batch_size
self.gamma = gamma
self.tau = tau
self.update_every = update_every
self.seed = random.seed(seed)
# Q-Network
self.qnetwork_local = QNetwork(action_size)
self.qnetwork_target = QNetwork(action_size)
#self.optimizer = tf.keras.optimizers.Adam(learning_rate=learning_rate)
self.optimizer = tf.keras.optimizers.RMSprop(learning_rate=0.00025,
momentum=0.95)
self.loss_fn = tf.keras.losses.MeanSquaredError()
# Replay memory
self.memory = ReplayBuffer(action_size, buffer_size, batch_size, seed)
# Initialize time step (for updating every self.update_every steps)
self.t_step = 0
def step(self, state, action, reward, next_state, done):
"""Adds new experience to the replay buffer and learns from a subset of memories.
Params
======
state (array_like): initial state
action (int): chosen action
reward (int): reward given
next_state (array_like): next state
done (bool): True if the episode is finished
"""
# Save experience in replay memory
self.memory.add(state, action, reward, next_state, done)
# Learn every self.update_every time steps.
self.t_step = (self.t_step + 1) % self.update_every
if self.t_step == 0:
# If enough samples are available in memory, get random subset and learn
if len(self.memory) > self.batch_size:
experiences = self.memory.sample()
self.learn(experiences)
def act(self, state, epsilon=0.):
"""Returns actions for given state as per current policy.
Params
======
state (array_like): current state
epsilon (float): epsilon, for epsilon-greedy action selection
"""
#state = tf.expand_dims(state, axis=0)
action_values = self.qnetwork_local(state)
# Epsilon-greedy action selection
if random.random() > epsilon:
return np.argmax(action_values.numpy())
else:
return random.choice(np.arange(self.action_size))
def learn(self, experiences):
"""Update value parameters using given batch of experience tuples.
Params
======
experiences (Tuple[torch.Variable]): tuple of (s, a, r, s', done) tuples
"""
states, actions, rewards, next_states, dones = experiences
# Get max predicted Q values (for next states) from target model
Q_targets_next = self.qnetwork_target(next_states)
Q_targets_next = tf.math.reduce_max(Q_targets_next, axis=1)
Q_targets_next = tf.expand_dims(Q_targets_next, axis=1)
# Compute Q targets for current states
Q_targets = rewards + (self.gamma * Q_targets_next * (1 - dones))
with tf.GradientTape() as tape:
# Get expected Q values from local model
Q_expected = self.qnetwork_local(states)
Q_expected = tf.gather(Q_expected,
indices=actions,
axis=1,
batch_dims=1)
loss = self.loss_fn(y_true=Q_targets, y_pred=Q_expected)
gradients = tape.gradient(loss, self.qnetwork_local.trainable_weights)
self.optimizer.apply_gradients(
zip(gradients, self.qnetwork_local.trainable_weights))
# ------------------- update target network ------------------- #
#self.soft_update()
def hard_update(self):
self.qnetwork_target.set_weights(self.qnetwork_local.get_weights())
def soft_update(self):
"""Soft update model parameters.
Params
======
local_model (PyTorch model): weights will be copied from
target_model (PyTorch model): weights will be copied to
"""
for target_layer, local_layer in zip(self.qnetwork_target.layers,
self.qnetwork_local.layers):
if target_layer.trainable:
for i in range(len(target_layer.trainable_weights)):
target_layer.trainable_weights[
i] = self.tau * local_layer.trainable_weights[i] + (
1.0 - self.tau) * target_layer.trainable_weights[i]
class Agent:
"""Interacts with and learns from the environment."""
def __init__(self,
state_size,
action_size,
buffer_size,
batch_size,
gamma,
tau,
lr,
update_every):
"""Initialize an Agent object.
Args:
state_size: Integer. Dimension of each state
action_size: Integer. Dimension of each action
buffer_size: Integer. Replay buffer size
batch_size: Integer. Mini-batch size
gamma: Float. Discount factor
tau: Float. For soft update of target parameters
lr: Float. Learning rate
update_every: Integer. How often to update the network
"""
# Environment parameters
self.state_size = state_size
self.action_size = action_size
# Q-Learning
self.gamma = gamma
# Q-Network
self.model_local = QNetwork(state_size, action_size)
self.model_target = QNetwork(state_size, action_size)
self.optimizer = tf.keras.optimizers.Adam(learning_rate=lr)
self.loss_fn = tf.keras.losses.MeanSquaredError(name="mse")
self.tau = tau
self.update_every = update_every
self.batch_size = batch_size
# Replay memory
self.memory = ReplayBuffer(action_size, buffer_size, batch_size)
# Initialize time step (for updating every update_every steps)
self.t_step = 0
def __str__(self):
return 'RL_Agent_Class'
def __repr__(self):
return 'RL_Agent_Class'
def step(self, state, action, reward, next_state, done):
"""Save state on buffer and trigger learn according to update_every
Args:
state: The previous state of the environment
action: Integer. Previous action selected by the agent
reward: Float. Reward value
next_state: The current state of the environment
done: Boolean. Whether the episode is complete
"""
# Save experience in replay memory
self.memory.add(state, action, reward, next_state, done)
# Learn every update_every time steps.
self.t_step = (self.t_step + 1) % self.update_every
if self.t_step == 0:
# If enough samples are available in memory, get random subset and learn
if len(self.memory) > self.batch_size:
experiences = self.memory.sample()
self.learn(experiences)
def act(self, state, eps=0.):
"""Returns actions for given state as per current policy.
Args:
state: A array like object or list with states
eps: Float. Random value for epsilon-greedy action selection
Returns:
An action selected by the network or by the epsilon-greedy method
"""
# Reshape state
state = np.expand_dims(state, 0)
# Predict action
action_values = self.model_local(state)
# Epsilon-greedy action selection
if random.random() > eps:
return np.argmax(action_values)
else:
return random.choice(np.arange(self.action_size))
def learn(self, experiences):
"""Update value parameters using given batch of experience tuples.
Args:
experiences: Tuple. Content of tuple (s, a, r, s', done)
"""
states, actions, rewards, next_states, dones = experiences
# Create mask to actions
mask = tf.one_hot(actions.reshape(-1), self.action_size)
with tf.GradientTape(persistent=True) as tape:
# Get expected Q values from local model
q_expected = tf.reduce_sum(self.model_local(states) * mask, axis=1, keepdims=True)
# Get max predicted Q values (for next states) from target model
q_targets_next = tf.reduce_max(self.model_target(next_states, training=True), axis=1)
# Compute Q targets for current states
q_targets = tf.add(
rewards, tf.multiply(
self.gamma, q_targets_next, tf.subtract(1.0, dones))
)
# Compute loss
loss = self.loss_fn(q_expected, q_targets)
# Minimize the loss
gradients = tape.gradient(loss, self.model_local.trainable_variables)
self.optimizer.apply_gradients(zip(gradients, self.model_local.trainable_variables))
# Update target network
self.soft_update()
def soft_update(self):
"""Soft update model parameters.
The model is update using:
θ_target = τ * θ_local + (1 - τ) * θ_target
"""
# Instantiate weight list
new_weights = []
# Apply soft update
for weights in self.model_local.get_weights():
new_weights.append(self.tau * weights + (1.0 - self.tau) * weights)
# Set new weights
self.model_target.set_weights(new_weights)
class Actor:
def __init__(self, pid, epsilon, gamma, env_name):
self.pid = pid
self.env_name = env_name
self.env = gym.make(self.env_name)
self.action_space = self.env.action_space.n
self.q_network = QNetwork(self.action_space)
self.epsilon = epsilon
self.gamma = gamma
self.buffer = []
self.state = self.env.reset()
self.define_network()
self.episode_rewards = 0
def define_network(self):
tf.config.set_visible_devices([], 'GPU')
env = gym.make(self.env_name)
state = env.reset()
self.q_network(np.atleast_2d(state))
def rollout(self, current_weights):
#: グローバルQ関数と重みを同期
self.q_network.set_weights(current_weights)
#: rollout 100step
for _ in range(100):
state = self.state
action = self.q_network.sample_action(state, self.epsilon)
next_state, reward, done, _ = self.env.step(action)
self.episode_rewards += reward
transition = (state, action, reward, next_state, done)
self.buffer.append(transition)
if done:
print(self.episode_rewards)
self.state = self.env.reset()
self.episode_rewards = 0
else:
self.state = next_state
#: 初期優先度の計算
states = np.vstack([transition[0] for transition in self.buffer])
actions = np.array([transition[1] for trainsition in self.buffer])
rewards = np.vstack([transition[2] for trainsition in self.buffer])
next_states = np.vstack([transition[3] for transition in self.buffer])
dones = np.vstack([transition[4] for transition in self.buffer])
next_qvalues = self.q_network(next_states)
next_actions = tf.cast(tf.argmax(next_qvalues, axis=1), tf.int32)
next_actions_onehot = tf.one_hot(next_actions, self.action_space)
next_maxQ = tf.reduce_sum(
next_qvalues * next_actions_onehot, axis=1, keepdims=True)
TQ = rewards + self.gamma * (1 - dones) * next_maxQ
qvalues = self.q_network(states)
actions_onehot = tf.one_hot(actions, self.action_space)
Q = tf.reduce_sum(qvalues * actions_onehot, axis=1, keepdims=True)
td_errors = (TQ - Q).numpy().flatten()
transitions = self.buffer
self.buffer = []
return td_errors, transitions, self.pid
class DQNAgent:
def __init__(self,
env_name="BreakoutDeterministic-v4",
gamma=0.99,
batch_size=32,
lr=0.00025,
update_period=4,
target_update_period=10000,
n_frames=4):
self.env_name = env_name
self.gamma = gamma
self.batch_size = batch_size
self.epsilon_scheduler = (
lambda steps: max(1.0 - 0.9 * steps / 1000000, 0.1))
self.update_period = update_period
self.target_update_period = target_update_period
env = gym.make(self.env_name)
self.action_space = env.action_space.n
self.qnet = QNetwork(self.action_space)
self.target_qnet = QNetwork(self.action_space)
self.optimizer = Adam(lr=lr, epsilon=0.01 / self.batch_size)
self.n_frames = n_frames
self.use_reward_clipping = True
self.huber_loss = tf.keras.losses.Huber()
def learn(self, n_episodes, buffer_size=1000000, logdir="log"):
logdir = Path(__file__).parent / logdir
if logdir.exists():
shutil.rmtree(logdir)
self.summary_writer = tf.summary.create_file_writer(str(logdir))
self.replay_buffer = ReplayBuffer(max_len=buffer_size)
steps = 0
for episode in range(1, n_episodes + 1):
env = gym.make(self.env_name)
frame = preprocess_frame(env.reset())
frames = collections.deque([frame] * self.n_frames,
maxlen=self.n_frames)
episode_rewards = 0
episode_steps = 0
done = False
lives = 5
while not done:
steps, episode_steps = steps + 1, episode_steps + 1
epsilon = self.epsilon_scheduler(steps)
state = np.stack(frames, axis=2)[np.newaxis, ...]
action = self.qnet.sample_action(state, epsilon=epsilon)
next_frame, reward, done, info = env.step(action)
episode_rewards += reward
frames.append(preprocess_frame(next_frame))
next_state = np.stack(frames, axis=2)[np.newaxis, ...]
if info["ale.lives"] != lives:
lives = info["ale.lives"]
transition = (state, action, reward, next_state, True)
else:
transition = (state, action, reward, next_state, done)
self.replay_buffer.push(transition)
if len(self.replay_buffer) > 50000:
if steps % self.update_period == 0:
loss = self.update_network()
with self.summary_writer.as_default():
tf.summary.scalar("loss", loss, step=steps)
tf.summary.scalar("epsilon", epsilon, step=steps)
tf.summary.scalar("buffer_size",
len(self.replay_buffer),
step=steps)
tf.summary.scalar("train_score",
episode_rewards,
step=steps)
tf.summary.scalar("train_steps",
episode_steps,
step=steps)
if steps % self.target_update_period == 0:
self.target_qnet.set_weights(self.qnet.get_weights())
if done:
break
print(
f"Episode: {episode}, score: {episode_rewards}, steps: {episode_steps}"
)
if episode % 20 == 0:
test_scores, test_steps = self.test_play(n_testplay=1)
with self.summary_writer.as_default():
tf.summary.scalar("test_score", test_scores[0], step=steps)
tf.summary.scalar("test_step", test_steps[0], step=steps)
if episode % 1000 == 0:
self.qnet.save_weights("checkpoints/qnet")
def update_network(self):
#: ミニバッチの作成
(states, actions, rewards, next_states,
dones) = self.replay_buffer.get_minibatch(self.batch_size)
if self.use_reward_clipping:
rewards = np.clip(rewards, -1, 1)
next_actions, next_qvalues = self.target_qnet.sample_actions(
next_states)
next_actions_onehot = tf.one_hot(next_actions, self.action_space)
max_next_qvalues = tf.reduce_sum(next_qvalues * next_actions_onehot,
axis=1,
keepdims=True)
target_q = rewards + self.gamma * (1 - dones) * max_next_qvalues
with tf.GradientTape() as tape:
qvalues = self.qnet(states)
actions_onehot = tf.one_hot(actions.flatten().astype(np.int32),
self.action_space)
q = tf.reduce_sum(qvalues * actions_onehot, axis=1, keepdims=True)
loss = self.huber_loss(target_q, q)
grads = tape.gradient(loss, self.qnet.trainable_variables)
self.optimizer.apply_gradients(
zip(grads, self.qnet.trainable_variables))
return loss
def test_play(self, n_testplay=1, monitor_dir=None, checkpoint_path=None):
if checkpoint_path:
env = gym.make(self.env_name)
frame = preprocess_frame(env.reset())
frames = collections.deque([frame] * self.n_frames,
maxlen=self.n_frames)
state = np.stack(frames, axis=2)[np.newaxis, ...]
self.qnet(state)
self.qnet.load_weights(checkpoint_path)
if monitor_dir:
monitor_dir = Path(monitor_dir)
if monitor_dir.exists():
shutil.rmtree(monitor_dir)
monitor_dir.mkdir()
env = gym.wrappers.Monitor(gym.make(self.env_name),
monitor_dir,
force=True,
video_callable=(lambda ep: True))
else:
env = gym.make(self.env_name)
scores = []
steps = []
for _ in range(n_testplay):
frame = preprocess_frame(env.reset())
frames = collections.deque([frame] * self.n_frames,
maxlen=self.n_frames)
done = False
episode_steps = 0
episode_rewards = 0
while not done:
state = np.stack(frames, axis=2)[np.newaxis, ...]
action = self.qnet.sample_action(state, epsilon=0.05)
next_frame, reward, done, _ = env.step(action)
frames.append(preprocess_frame(next_frame))
episode_rewards += reward
episode_steps += 1
if episode_steps > 500 and episode_rewards < 3:
#: ゲーム開始(action: 0)しないまま停滞するケースへの対処
break
scores.append(episode_rewards)
steps.append(episode_steps)
return scores, steps