kernel_initializer=initializer),
Conv2D(16, 2, activation='elu', padding='valid',
input_shape=dummy_env.observation_space.shape,
kernel_initializer=initializer),
Flatten(),
Dropout(0.5),
Dense(512, activation='elu', kernel_initializer=initializer)
])
# Exploration and learning rate decay after each epoch
eps = 0.2
eps_decay = 0.9
learning_rate = 3e-3
learning_decay = 0.9
explore_policy = EpsGreedy(eps, gym_2048.check_valid)
exploit_policy = Greedy(gym_2048.check_valid)
# Create Deep Q-Learning Network agent
agent = DQN(model, actions=dummy_env.action_space.n, gamma=0.99,
batch_size=64, nsteps=50, enable_double_dqn=True,
enable_dueling_network=True, target_update=10,
test_policy=exploit_policy)
def plot_rewards_show(episode_rewards, episode_steps, done=False,
title='Rewards'):
plt.clf()
plt.xlabel('Step')
plt.ylabel('Reward')
plt.title('DQN rewards')
for i, (ed, steps) in enumerate(zip(episode_rewards, episode_steps)):
plt.plot(steps, ed, alpha=0.5 if i == 0 else 0.2,
def __init__(self,
model,
actions,
optimizer=None,
policy=None,
test_policy=None,
gamma=0.99,
instances=8,
nsteps=1,
value_loss=0.5,
entropy_loss=0.01):
"""
TODO: Describe parameters
"""
self.actions = actions
self.optimizer = Adam(lr=3e-3) if optimizer is None else optimizer
self.memory = memory.OnPolicy(steps=nsteps, instances=instances)
if policy is None:
# Create one policy per instance, with varying exploration parameters
self.policy = [Greedy()] + [
GaussianEpsGreedy(eps, 0.1)
for eps in np.arange(0, 1, 1 / (instances - 1))
]
else:
self.policy = policy
self.test_policy = Greedy() if test_policy is None else test_policy
self.gamma = gamma
self.instances = instances
self.nsteps = nsteps
self.value_loss = value_loss
self.entropy_loss = entropy_loss
self.training = True
# Create output model layers based on number of actions
raw_output = model.layers[-1].output
actor = Dense(actions, activation='softmax')(
raw_output) # Actor (Policy Network)
critic = Dense(1, activation='linear')(
raw_output) # Critic (Value Network)
output_layer = Concatenate()([actor, critic])
self.model = Model(inputs=model.input, outputs=output_layer)
def a2c_loss(targets_actions, y_pred):
# Unpack input
targets, actions = targets_actions[:,
0], targets_actions[:,
1:] # Unpack
probs, values = y_pred[:, :-1], y_pred[:, -1]
# Compute advantages and logprobabilities
adv = targets - values
logprob = tf.math.log(
tf.reduce_sum(probs * actions, axis=1, keepdims=False) + 1e-10)
# Compute composite loss
loss_policy = -adv * logprob
loss_value = self.value_loss * tf.square(adv)
entropy = self.entropy_loss * tf.reduce_sum(
probs * tf.math.log(probs + 1e-10), axis=1, keepdims=False)
return tf.reduce_mean(loss_policy + loss_value + entropy)
self.model.compile(optimizer=self.optimizer, loss=a2c_loss)
class A2C(Agent):
"""Advantage Actor-Critic (A2C)
A2C is a synchronous version of A3C which gives equal or better performance.
For more information on A2C refer to the OpenAI blog post: https://blog.openai.com/baselines-acktr-a2c/.
The A3C algorithm is described in "Asynchronous Methods for Deep Reinforcement Learning" (Mnih et al., 2016)
Since this algorithm is on-policy, it can and should be trained with multiple simultaneous environment instances.
The parallelism decorrelates the agents' data into a more stationary process which aids learning.
"""
def __init__(self,
model,
actions,
optimizer=None,
policy=None,
test_policy=None,
gamma=0.99,
instances=8,
nsteps=1,
value_loss=0.5,
entropy_loss=0.01):
"""
TODO: Describe parameters
"""
self.actions = actions
self.optimizer = Adam(lr=3e-3) if optimizer is None else optimizer
self.memory = memory.OnPolicy(steps=nsteps, instances=instances)
if policy is None:
# Create one policy per instance, with varying exploration parameters
self.policy = [Greedy()] + [
GaussianEpsGreedy(eps, 0.1)
for eps in np.arange(0, 1, 1 / (instances - 1))
]
else:
self.policy = policy
self.test_policy = Greedy() if test_policy is None else test_policy
self.gamma = gamma
self.instances = instances
self.nsteps = nsteps
self.value_loss = value_loss
self.entropy_loss = entropy_loss
self.training = True
# Create output model layers based on number of actions
raw_output = model.layers[-1].output
actor = Dense(actions, activation='softmax')(
raw_output) # Actor (Policy Network)
critic = Dense(1, activation='linear')(
raw_output) # Critic (Value Network)
output_layer = Concatenate()([actor, critic])
self.model = Model(inputs=model.input, outputs=output_layer)
def a2c_loss(targets_actions, y_pred):
# Unpack input
targets, actions = targets_actions[:,
0], targets_actions[:,
1:] # Unpack
probs, values = y_pred[:, :-1], y_pred[:, -1]
# Compute advantages and logprobabilities
adv = targets - values
logprob = tf.math.log(
tf.reduce_sum(probs * actions, axis=1, keepdims=False) + 1e-10)
# Compute composite loss
loss_policy = -adv * logprob
loss_value = self.value_loss * tf.square(adv)
entropy = self.entropy_loss * tf.reduce_sum(
probs * tf.math.log(probs + 1e-10), axis=1, keepdims=False)
return tf.reduce_mean(loss_policy + loss_value + entropy)
self.model.compile(optimizer=self.optimizer, loss=a2c_loss)
def save(self, filename, overwrite=False):
"""Saves the model parameters to the specified file."""
self.model.save_weights(filename, overwrite=overwrite)
def act(self, state, instance=0):
"""Returns the action to be taken given a state."""
qvals = self.model.predict(np.array([state]))[0][:-1]
if self.training:
return self.policy[instance].act(qvals) if isinstance(
self.policy, list) else self.policy.act(qvals)
else:
return self.test_policy[instance].act(qvals) if isinstance(
self.test_policy, list) else self.test_policy.act(qvals)
def push(self, transition, instance=0):
"""Stores the transition in memory."""
self.memory.put(transition, instance)
def train(self, step):
"""Trains the agent for one step."""
if len(self.memory) < self.instances:
return
state_batch, action_batch, reward_batches, end_state_batch, not_done_mask = self.memory.get(
)
# Compute the value of the last next states
target_qvals = np.zeros(self.instances)
non_final_last_next_states = [
es for es in end_state_batch if es is not None
]
if len(non_final_last_next_states) > 0:
non_final_mask = list(map(lambda s: s is not None,
end_state_batch))
target_qvals[non_final_mask] = self.model.predict_on_batch(
np.array(non_final_last_next_states))[:, -1].squeeze()
# Compute n-step discounted return
# If episode ended within any sampled nstep trace - zero out remaining rewards
for n in reversed(range(self.nsteps)):
rewards = np.array([b[n] for b in reward_batches])
target_qvals *= np.array([t[n] for t in not_done_mask])
target_qvals = rewards + (self.gamma * target_qvals)
# Prepare loss data: target Q-values and actions taken (as a mask)
ran = np.arange(self.instances)
targets_actions = np.zeros((self.instances, self.actions + 1))
targets_actions[ran, 0] = target_qvals
targets_actions[ran, np.array(action_batch) + 1] = 1
self.model.train_on_batch(np.array(state_batch), targets_actions)
# Optimizer with sheduled learning rate decay
optimizer = Adam(lr=3e-3, decay=1e-5)
# Run multiple instances
instances = 8
# Exploration and learning rate decay after each epoch
eps_max = 0.2
eps_decay = 0.9
learning_rate = 3e-3
learning_decay = 0.9
# Create Advantage Actor-Critic agent
agent = A2C(model,
actions=dummy_env.action_space.n,
nsteps=20,
instances=instances,
optimizer=optimizer,
test_policy=Greedy(gym_2048.check_valid))
# Run epochs
for epoch in range(20):
# Create a policy for each instance with a different eps
policy = [Greedy(gym_2048.check_valid)] + [
EpsGreedy(eps, gym_2048.check_valid)
for eps in np.arange(0, eps_max, eps_max / (instances - 1))
]
# Update agent
agent.policy = policy
agent.model.optimizer.lr = learning_rate
# Run epoch
print(f'Epoch {epoch}')
run_epoch(create_env,
agent,
def __init__(self,
model,
actions,
optimizer=None,
policy=None,
test_policy=None,
memsize=100000,
target_update=10,
gamma=0.99,
batch_size=32,
nsteps=0,
enable_double_dqn=False,
enable_dueling_network=False,
dueling_type='avg'):
"""
TODO: Describe parameters
"""
self.actions = actions
self.optimizer = Adam(lr=3e-3) if optimizer is None else optimizer
self.policy = EpsGreedy(0.1) if policy is None else policy
self.test_policy = Greedy() if test_policy is None else test_policy
self.memsize = memsize
self.memory = PrioritizedExperienceReplay(memsize, nsteps)
self.target_update = target_update
self.gamma = gamma
self.batch_size = batch_size
self.nsteps = nsteps
self.training = True
# Extension options
self.enable_double_dqn = enable_double_dqn
self.enable_dueling_network = enable_dueling_network
self.dueling_type = dueling_type
# Create output layer based on number of actions and (optionally) a dueling architecture
raw_output = model.layers[-1].output
if self.enable_dueling_network:
# "Dueling Network Architectures for Deep Reinforcement Learning" (Wang et al., 2016)
# Output the state value (V) and the action-specific advantages (A) separately then
# compute the Q values: Q = A + V
dueling_layer = Dense(self.actions + 1,
activation='linear')(raw_output)
if self.dueling_type == 'avg': f = lambda a: tf.expand_dims(a[:,0], -1) + a[:,1:] - \
tf.reduce_mean(a[:,1:], axis=1, keepdims=True)
elif self.dueling_type == 'max': f = lambda a: tf.expand_dims(a[:,0], -1) + a[:,1:] - \
tf.reduce_max(a[:,1:], axis=1, keepdims=True)
elif self.dueling_type == 'naive':
f = lambda a: tf.expand_dims(a[:, 0], -1) + a[:, 1:]
else:
raise HkException(
"dueling_type must be one of {'avg','max','naive'}")
output_layer = Lambda(f,
output_shape=(self.actions, ))(dueling_layer)
else:
output_layer = Dense(self.actions, activation='linear')(raw_output)
self.model = Model(inputs=model.input, outputs=output_layer)
# Define loss function that computes the MSE between target Q-values and cumulative discounted rewards
# If using PrioritizedExperienceReplay, the loss function also computes the TD error
# and updates the trace priorities
def masked_q_loss(data, y_pred):
"""Computes the MSE between the Q-values of the actions that were taken and the cumulative discounted
rewards obtained after taking those actions. Updates trace priorities if using PrioritizedExperienceReplay.
"""
action_batch, target_qvals = data[:, 0], data[:, 1]
seq = tf.cast(tf.range(0, tf.shape(action_batch)[0]), tf.int32)
action_idxs = tf.transpose(
tf.stack([seq, tf.cast(action_batch, tf.int32)]))
qvals = tf.gather_nd(y_pred, action_idxs)
if isinstance(self.memory, PrioritizedExperienceReplay):
def update_priorities(_qvals, _target_qvals, _traces_idxs):
"""Computes the TD error and updates memory priorities."""
td_error = np.abs((_target_qvals - _qvals).numpy())
_traces_idxs = (tf.cast(_traces_idxs, tf.int32)).numpy()
self.memory.update_priorities(_traces_idxs, td_error)
return _qvals
qvals = tf.py_function(func=update_priorities,
inp=[qvals, target_qvals, data[:, 2]],
Tout=tf.float32)
return tf.keras.losses.mse(qvals, target_qvals)
self.model.compile(optimizer=self.optimizer, loss=masked_q_loss)
# Clone model to use for delayed Q targets
self.target_model = tf.keras.models.clone_model(self.model)
self.target_model.set_weights(self.model.get_weights())
class DQN(Agent):
"""Deep Q-Learning Network
Base implementation:
"Playing Atari with Deep Reinforcement Learning" (Mnih et al., 2013)
Extensions:
Multi-step returns: "Reinforcement Learning: An Introduction" 2nd ed. (Sutton & Barto, 2018)
Double Q-Learning: "Deep Reinforcement Learning with Double Q-learning" (van Hasselt et al., 2015)
Dueling Q-Network: "Dueling Network Architectures for Deep Reinforcement Learning" (Wang et al., 2016)
"""
def __init__(self,
model,
actions,
optimizer=None,
policy=None,
test_policy=None,
memsize=100000,
target_update=10,
gamma=0.99,
batch_size=32,
nsteps=0,
enable_double_dqn=False,
enable_dueling_network=False,
dueling_type='avg'):
"""
TODO: Describe parameters
"""
self.actions = actions
self.optimizer = Adam(lr=3e-3) if optimizer is None else optimizer
self.policy = EpsGreedy(0.1) if policy is None else policy
self.test_policy = Greedy() if test_policy is None else test_policy
self.memsize = memsize
self.memory = PrioritizedExperienceReplay(memsize, nsteps)
self.target_update = target_update
self.gamma = gamma
self.batch_size = batch_size
self.nsteps = nsteps
self.training = True
# Extension options
self.enable_double_dqn = enable_double_dqn
self.enable_dueling_network = enable_dueling_network
self.dueling_type = dueling_type
# Create output layer based on number of actions and (optionally) a dueling architecture
raw_output = model.layers[-1].output
if self.enable_dueling_network:
# "Dueling Network Architectures for Deep Reinforcement Learning" (Wang et al., 2016)
# Output the state value (V) and the action-specific advantages (A) separately then
# compute the Q values: Q = A + V
dueling_layer = Dense(self.actions + 1,
activation='linear')(raw_output)
if self.dueling_type == 'avg': f = lambda a: tf.expand_dims(a[:,0], -1) + a[:,1:] - \
tf.reduce_mean(a[:,1:], axis=1, keepdims=True)
elif self.dueling_type == 'max': f = lambda a: tf.expand_dims(a[:,0], -1) + a[:,1:] - \
tf.reduce_max(a[:,1:], axis=1, keepdims=True)
elif self.dueling_type == 'naive':
f = lambda a: tf.expand_dims(a[:, 0], -1) + a[:, 1:]
else:
raise HkException(
"dueling_type must be one of {'avg','max','naive'}")
output_layer = Lambda(f,
output_shape=(self.actions, ))(dueling_layer)
else:
output_layer = Dense(self.actions, activation='linear')(raw_output)
self.model = Model(inputs=model.input, outputs=output_layer)
# Define loss function that computes the MSE between target Q-values and cumulative discounted rewards
# If using PrioritizedExperienceReplay, the loss function also computes the TD error
# and updates the trace priorities
def masked_q_loss(data, y_pred):
"""Computes the MSE between the Q-values of the actions that were taken and the cumulative discounted
rewards obtained after taking those actions. Updates trace priorities if using PrioritizedExperienceReplay.
"""
action_batch, target_qvals = data[:, 0], data[:, 1]
seq = tf.cast(tf.range(0, tf.shape(action_batch)[0]), tf.int32)
action_idxs = tf.transpose(
tf.stack([seq, tf.cast(action_batch, tf.int32)]))
qvals = tf.gather_nd(y_pred, action_idxs)
if isinstance(self.memory, PrioritizedExperienceReplay):
def update_priorities(_qvals, _target_qvals, _traces_idxs):
"""Computes the TD error and updates memory priorities."""
td_error = np.abs((_target_qvals - _qvals).numpy())
_traces_idxs = (tf.cast(_traces_idxs, tf.int32)).numpy()
self.memory.update_priorities(_traces_idxs, td_error)
return _qvals
qvals = tf.py_function(func=update_priorities,
inp=[qvals, target_qvals, data[:, 2]],
Tout=tf.float32)
return tf.keras.losses.mse(qvals, target_qvals)
self.model.compile(optimizer=self.optimizer, loss=masked_q_loss)
# Clone model to use for delayed Q targets
self.target_model = tf.keras.models.clone_model(self.model)
self.target_model.set_weights(self.model.get_weights())
def save(self, filename, overwrite=False):
"""Saves the model parameters to the specified file."""
self.model.save_weights(filename, overwrite=overwrite)
def act(self, state, instance=0):
"""Returns the action to be taken given a state."""
qvals = self.model.predict(np.array([state]))[0]
return self.policy.act(
qvals) if self.training else self.test_policy.act(qvals)
def push(self, transition, instance=0):
"""Stores the transition in memory."""
self.memory.put(transition)
def train(self, step):
"""Trains the agent for one step."""
if len(self.memory) == 0:
return
# Update target network
if self.target_update >= 1 and step % self.target_update == 0:
# Perform a hard update
self.target_model.set_weights(self.model.get_weights())
elif self.target_update < 1:
# Perform a soft update
mw = np.array(self.model.get_weights())
tmw = np.array(self.target_model.get_weights())
self.target_model.set_weights(self.target_update * mw +
(1 - self.target_update) * tmw)
# Train even when memory has fewer than the specified batch_size
batch_size = min(len(self.memory), self.batch_size)
# Sample batch_size traces from memory
state_batch, action_batch, reward_batches, end_state_batch, not_done_mask = self.memory.get(
batch_size)
# Compute the value of the last next states
target_qvals = np.zeros(batch_size)
non_final_last_next_states = [
es for es in end_state_batch if es is not None
]
if len(non_final_last_next_states) > 0:
if self.enable_double_dqn:
# "Deep Reinforcement Learning with Double Q-learning" (van Hasselt et al., 2015)
# The online network predicts the actions while the target network is used to estimate the Q-values
q_values = self.model.predict_on_batch(
np.array(non_final_last_next_states))
actions = np.argmax(q_values, axis=1)
# Estimate Q-values using the target network but select the values with the
# highest Q-value wrt to the online model (as computed above).
target_q_values = self.target_model.predict_on_batch(
np.array(non_final_last_next_states))
selected_target_q_vals = target_q_values[
range(len(target_q_values)), actions]
else:
# Use delayed target network to compute target Q-values
selected_target_q_vals = self.target_model.predict_on_batch(
np.array(non_final_last_next_states)).max(1)
non_final_mask = list(map(lambda s: s is not None,
end_state_batch))
target_qvals[non_final_mask] = selected_target_q_vals
# Compute n-step discounted return
# If episode ended within any sampled nstep trace - zero out remaining rewards
for n in reversed(range(self.nsteps)):
rewards = np.array([b[n] for b in reward_batches])
target_qvals *= np.array([t[n] for t in not_done_mask])
target_qvals = rewards + (self.gamma * target_qvals)
# Compile information needed by the custom loss function
loss_data = [action_batch, target_qvals]
# If using PrioritizedExperienceReplay then we need to provide the trace indexes
# to the loss function as well so we can update the priorities of the traces
if isinstance(self.memory, PrioritizedExperienceReplay):
loss_data.append(self.memory.last_traces_idxs())
# Train model
self.model.train_on_batch(np.array(state_batch),
np.stack(loss_data).transpose())