def initialize(self, custom_getter):
super(QModel, self).initialize(custom_getter)
# TEMP: Random sampling fix
if self.random_sampling_fix:
self.next_state_inputs = dict()
for name, state in self.states_spec.items():
self.next_state_inputs[name] = tf.placeholder(
dtype=util.tf_dtype(state['type']),
shape=(None, ) + tuple(state['shape']),
name=('next-' + name))
# Target network
self.target_network = Network.from_spec(
spec=self.network_spec,
kwargs=dict(scope='target', summary_labels=self.summary_labels))
# Target network optimizer
self.target_optimizer = Synchronization(
sync_frequency=self.target_sync_frequency,
update_weight=self.target_update_weight)
# Target network distributions
self.target_distributions = self.generate_distributions(
self.actions_spec, self.distributions_spec, self.summary_labels)
def setup_components_and_tf_funcs(self, custom_getter=None):
custom_getter = super(
DPGTargetModel, self).setup_components_and_tf_funcs(custom_getter)
# Target network
self.target_network = Network.from_spec(
spec=self.network_spec,
kwargs=dict(scope='target-network',
summary_labels=self.summary_labels))
# Target network optimizer
self.target_network_optimizer = Synchronization(
sync_frequency=self.target_sync_frequency,
update_weight=self.target_update_weight)
# Target network distributions
self.target_distributions = self.create_distributions()
# Critic
#print ("type of self.critic_network_spec[]:")
#print (type(self.critic_network_spec))
#for element in self.critic_network_spec:
# print (element)
# oliver: CHANGES HERE!!
size_t0 = self.critic_network_spec[0]['size']
size_t1 = self.critic_network_spec[1]['size']
self.critic = DDPGCriticNetwork(scope='critic',
size_t0=size_t0,
size_t1=size_t1)
self.critic_optimizer = Optimizer.from_spec(
spec=self.critic_optimizer_spec,
kwargs=dict(summary_labels=self.summary_labels))
self.target_critic = DDPGCriticNetwork(scope='target-critic',
size_t0=size_t0,
size_t1=size_t1)
# Target critic optimizer
self.target_critic_optimizer = Synchronization(
sync_frequency=self.target_sync_frequency,
update_weight=self.target_update_weight)
self.fn_target_actions_and_internals = tf.make_template(
name_='target-actions-and-internals',
func_=self.tf_target_actions_and_internals,
custom_getter_=custom_getter)
self.fn_predict_target_q = tf.make_template(
name_='predict-target-q',
func_=self.tf_predict_target_q,
custom_getter_=custom_getter)
return custom_getter
def initialize(self, custom_getter):
super(QModel, self).initialize(custom_getter)
# Target network
self.target_network = Network.from_spec(
spec=self.network_spec,
kwargs=dict(scope='target', summary_labels=self.summary_labels))
# Target network optimizer
self.target_optimizer = Synchronization(
sync_frequency=self.target_sync_frequency,
update_weight=self.target_update_weight)
def setup_components_and_tf_funcs(self, custom_getter=None):
custom_getter = super(DPGTargetModel, self).setup_components_and_tf_funcs(custom_getter)
# Target network
self.target_network = Network.from_spec(
spec=self.network_spec,
kwargs=dict(scope='target-network', summary_labels=self.summary_labels)
)
# Target network optimizer
self.target_network_optimizer = Synchronization(
sync_frequency=self.target_sync_frequency,
update_weight=self.target_update_weight
)
# Target network distributions
self.target_distributions = self.create_distributions()
# critic
self.critic_network = Network.from_spec(
spec=self.critic_network_spec,
kwargs=dict(scope='critic')
)
self.target_critic_network = Network.from_spec(
spec=self.critic_network_spec,
kwargs=dict(scope='target-critic')
)
self.critic_optimizer = Optimizer.from_spec(
spec=self.critic_optimizer_spec,
kwargs=dict(summary_labels=self.summary_labels)
)
# Target critic optimizer
self.target_critic_optimizer = Synchronization(
sync_frequency=self.target_sync_frequency,
update_weight=self.target_update_weight
)
self.fn_target_actions_and_internals = tf.make_template(
name_='target-actions-and-internals',
func_=self.tf_target_actions_and_internals,
custom_getter_=custom_getter
)
self.fn_predict_target_q = tf.make_template(
name_='predict-target-q',
func_=self.tf_predict_target_q,
custom_getter_=custom_getter
)
return custom_getter
def initialize(self, custom_getter):
super(DPGTargetModel, self).initialize(custom_getter)
# Target network
self.target_network = Network.from_spec(
spec=self.network_spec,
kwargs=dict(scope='target-network',
summary_labels=self.summary_labels))
# Target network optimizer
self.target_network_optimizer = Synchronization(
sync_frequency=self.target_sync_frequency,
update_weight=self.target_update_weight)
# Target network distributions
self.target_distributions = self.create_distributions()
# Critic
size_t0 = self.critic_network_spec['size_t0']
size_t1 = self.critic_network_spec['size_t1']
self.critic = DDPGCriticNetwork(scope='critic',
size_t0=size_t0,
size_t1=size_t1)
self.critic_optimizer = Optimizer.from_spec(
spec=self.critic_optimizer_spec,
kwargs=dict(summary_labels=self.summary_labels))
self.target_critic = DDPGCriticNetwork(scope='target-critic',
size_t0=size_t0,
size_t1=size_t1)
# Target critic optimizer
self.target_critic_optimizer = Synchronization(
sync_frequency=self.target_sync_frequency,
update_weight=self.target_update_weight)
self.fn_target_actions_and_internals = tf.make_template(
name_='target-actions-and-internals',
func_=self.tf_target_actions_and_internals,
custom_getter_=custom_getter)
self.fn_predict_target_q = tf.make_template(
name_='predict-target-q',
func_=self.tf_predict_target_q,
custom_getter_=custom_getter)
def __init__(self, states_spec, actions_spec, network_spec, config):
with tf.name_scope(name=config.scope):
# Target network
self.target_network = Network.from_spec(
spec=network_spec, kwargs=dict(scope='target'))
# Target network optimizer
self.target_optimizer = Synchronization(
sync_frequency=config.target_sync_frequency,
update_weight=config.target_update_weight)
self.double_q_model = config.double_q_model
assert config.huber_loss is None or config.huber_loss > 0.0
self.huber_loss = config.huber_loss
super(QModel, self).__init__(states_spec=states_spec,
actions_spec=actions_spec,
network_spec=network_spec,
config=config)
class QModel(DistributionModel):
"""
Q-value model.
"""
def __init__(
self,
states_spec,
actions_spec,
network_spec,
device,
scope,
saver_spec,
summary_spec,
distributed_spec,
optimizer,
discount,
normalize_rewards,
variable_noise,
distributions_spec,
entropy_regularization,
target_sync_frequency,
target_update_weight,
double_q_model,
huber_loss,
# TEMP: Random sampling fix
random_sampling_fix):
self.target_sync_frequency = target_sync_frequency
self.target_update_weight = target_update_weight
self.double_q_model = double_q_model
assert huber_loss is None or huber_loss > 0.0
self.huber_loss = huber_loss
# TEMP: Random sampling fix
self.random_sampling_fix = random_sampling_fix
super(QModel, self).__init__(
states_spec=states_spec,
actions_spec=actions_spec,
network_spec=network_spec,
device=device,
scope=scope,
saver_spec=saver_spec,
summary_spec=summary_spec,
distributed_spec=distributed_spec,
optimizer=optimizer,
discount=discount,
normalize_rewards=normalize_rewards,
variable_noise=variable_noise,
distributions_spec=distributions_spec,
entropy_regularization=entropy_regularization,
)
def initialize(self, custom_getter):
super(QModel, self).initialize(custom_getter)
# TEMP: Random sampling fix
if self.random_sampling_fix:
self.next_state_inputs = dict()
for name, state in self.states_spec.items():
self.next_state_inputs[name] = tf.placeholder(
dtype=util.tf_dtype(state['type']),
shape=(None, ) + tuple(state['shape']),
name=('next-' + name))
# Target network
self.target_network = Network.from_spec(
spec=self.network_spec,
kwargs=dict(scope='target', summary_labels=self.summary_labels))
# Target network optimizer
self.target_optimizer = Synchronization(
sync_frequency=self.target_sync_frequency,
update_weight=self.target_update_weight)
# Target network distributions
self.target_distributions = self.generate_distributions(
self.actions_spec, self.distributions_spec, self.summary_labels)
def tf_q_value(self, embedding, distr_params, action, name):
# Mainly for NAF.
return self.distributions[name].state_action_value(
distr_params=distr_params, action=action)
def tf_q_delta(self, q_value, next_q_value, terminal, reward):
"""
Creates the deltas (or advantage) of the Q values.
:return: A list of deltas per action
"""
for _ in range(util.rank(q_value) - 1):
terminal = tf.expand_dims(input=terminal, axis=1)
reward = tf.expand_dims(input=reward, axis=1)
multiples = (1, ) + util.shape(q_value)[1:]
terminal = tf.tile(input=terminal, multiples=multiples)
reward = tf.tile(input=reward, multiples=multiples)
zeros = tf.zeros_like(tensor=next_q_value)
next_q_value = tf.where(condition=terminal,
x=zeros,
y=(self.discount * next_q_value))
return reward + next_q_value - q_value # tf.stop_gradient(q_target)
def tf_loss_per_instance(self, states, internals, actions, terminal,
reward, update):
# TEMP: Random sampling fix
if self.random_sampling_fix:
next_states = self.get_states(states=self.next_state_inputs)
next_states = {
name: tf.stop_gradient(input=state)
for name, state in next_states.items()
}
embedding, next_internals = self.network.apply(
x=states,
internals=internals,
update=update,
return_internals=True)
# Both networks can use the same internals, could that be a problem?
# Otherwise need to handle internals indices correctly everywhere
target_embedding = self.target_network.apply(
x=next_states, internals=next_internals, update=update)
else:
embedding = self.network.apply(
x={name: state[:-1]
for name, state in states.items()},
internals=[internal[:-1] for internal in internals],
update=update)
# Both networks can use the same internals, could that be a problem?
# Otherwise need to handle internals indices correctly everywhere
target_embedding = self.target_network.apply(
x={name: state[1:]
for name, state in states.items()},
internals=[internal[1:] for internal in internals],
update=update)
actions = {name: action[:-1] for name, action in actions.items()}
terminal = terminal[:-1]
reward = reward[:-1]
deltas = list()
for name, distribution in self.distributions.items():
target_distribution = self.target_distributions[name]
distr_params = distribution.parameterize(x=embedding)
target_distr_params = target_distribution.parameterize(
x=target_embedding)
q_value = self.tf_q_value(embedding=embedding,
distr_params=distr_params,
action=actions[name],
name=name)
if self.double_q_model:
action_taken = distribution.sample(distr_params=distr_params,
deterministic=True)
else:
action_taken = target_distribution.sample(
distr_params=target_distr_params, deterministic=True)
next_q_value = target_distribution.state_action_value(
distr_params=target_distr_params, action=action_taken)
delta = self.tf_q_delta(q_value=q_value,
next_q_value=next_q_value,
terminal=terminal,
reward=reward)
collapsed_size = util.prod(util.shape(delta)[1:])
delta = tf.reshape(tensor=delta, shape=(-1, collapsed_size))
deltas.append(delta)
# Surrogate loss as the mean squared error between actual observed rewards and expected rewards
loss_per_instance = tf.reduce_mean(input_tensor=tf.concat(
values=deltas, axis=1),
axis=1)
# Optional Huber loss
if self.huber_loss is not None and self.huber_loss > 0.0:
return tf.where(
condition=(tf.abs(x=loss_per_instance) <= self.huber_loss),
x=(0.5 * tf.square(x=loss_per_instance)),
y=(self.huber_loss *
(tf.abs(x=loss_per_instance) - 0.5 * self.huber_loss)))
else:
return tf.square(x=loss_per_instance)
def tf_optimization(self, states, internals, actions, terminal, reward,
update):
optimization = super(QModel, self).tf_optimization(states=states,
internals=internals,
actions=actions,
terminal=terminal,
reward=reward,
update=update)
network_distributions_variables = self.get_distributions_variables(
self.distributions)
target_distributions_variables = self.get_distributions_variables(
self.target_distributions)
target_optimization = self.target_optimizer.minimize(
time=self.timestep,
variables=self.target_network.get_variables() +
target_distributions_variables,
source_variables=self.network.get_variables() +
network_distributions_variables)
return tf.group(optimization, target_optimization)
def get_variables(self, include_non_trainable=False):
model_variables = super(
QModel,
self).get_variables(include_non_trainable=include_non_trainable)
if include_non_trainable:
# Target network and optimizer variables only included if 'include_non_trainable' set
target_variables = self.target_network.get_variables(
include_non_trainable=include_non_trainable)
target_distributions_variables = self.get_distributions_variables(
self.target_distributions)
target_optimizer_variables = self.target_optimizer.get_variables()
return model_variables + target_variables + target_optimizer_variables + target_distributions_variables
else:
return model_variables
def get_summaries(self):
target_distributions_summaries = self.get_distributions_summaries(
self.target_distributions)
return super(QModel,
self).get_summaries() + self.target_network.get_summaries(
) + target_distributions_summaries
# TEMP: Random sampling fix
def update(self,
states,
internals,
actions,
terminal,
reward,
return_loss_per_instance=False):
fetches = [self.optimization]
# Optionally fetch loss per instance
if return_loss_per_instance:
fetches.append(self.loss_per_instance)
terminal = np.asarray(terminal)
batched = (terminal.ndim == 1)
if batched:
# TEMP: Random sampling fix
if self.random_sampling_fix:
feed_dict = {
state_input: states[name][0]
for name, state_input in self.state_inputs.items()
}
feed_dict.update({
state_input: states[name][1]
for name, state_input in self.next_state_inputs.items()
})
else:
feed_dict = {
state_input: states[name]
for name, state_input in self.state_inputs.items()
}
feed_dict.update({
internal_input: internals[n]
for n, internal_input in enumerate(self.internal_inputs)
})
feed_dict.update({
action_input: actions[name]
for name, action_input in self.action_inputs.items()
})
feed_dict[self.terminal_input] = terminal
feed_dict[self.reward_input] = reward
else:
# TEMP: Random sampling fix
if self.random_sampling_fix:
raise TensorForceError("Unbatched version not covered by fix.")
else:
feed_dict = {
state_input: (states[name], )
for name, state_input in self.state_inputs.items()
}
feed_dict.update({
internal_input: (internals[n], )
for n, internal_input in enumerate(self.internal_inputs)
})
feed_dict.update({
action_input: (actions[name], )
for name, action_input in self.action_inputs.items()
})
feed_dict[self.terminal_input] = (terminal, )
feed_dict[self.reward_input] = (reward, )
feed_dict[self.deterministic_input] = True
feed_dict[self.update_input] = True
fetched = self.monitored_session.run(fetches=fetches,
feed_dict=feed_dict)
if return_loss_per_instance:
return fetched[1]
class QModel(DistributionModel):
"""
Q-value model.
"""
def __init__(self, states_spec, actions_spec, network_spec, config):
self.target_sync_frequency = config.target_sync_frequency
self.target_update_weight = config.target_update_weight
self.double_q_model = config.double_q_model
assert config.huber_loss is None or config.huber_loss > 0.0
self.huber_loss = config.huber_loss
super(QModel, self).__init__(states_spec=states_spec,
actions_spec=actions_spec,
network_spec=network_spec,
config=config)
def initialize(self, custom_getter):
super(QModel, self).initialize(custom_getter)
# Target network
self.target_network = Network.from_spec(
spec=self.network_spec,
kwargs=dict(scope='target', summary_labels=self.summary_labels))
# Target network optimizer
self.target_optimizer = Synchronization(
sync_frequency=self.target_sync_frequency,
update_weight=self.target_update_weight)
def tf_q_value(self, embedding, distr_params, action, name):
# Mainly for NAF.
return self.distributions[name].state_action_value(
distr_params=distr_params, action=action)
def tf_q_delta(self, q_value, next_q_value, terminal, reward):
"""
Creates the deltas (or advantage) of the Q values.
:return: A list of deltas per action
"""
for _ in range(util.rank(q_value) - 1):
terminal = tf.expand_dims(input=terminal, axis=1)
reward = tf.expand_dims(input=reward, axis=1)
multiples = (1, ) + util.shape(q_value)[1:]
terminal = tf.tile(input=terminal, multiples=multiples)
reward = tf.tile(input=reward, multiples=multiples)
zeros = tf.zeros_like(tensor=next_q_value)
next_q_value = tf.where(condition=terminal,
x=zeros,
y=(self.discount * next_q_value))
return reward + next_q_value - q_value # tf.stop_gradient(q_target)
def tf_loss_per_instance(self, states, internals, actions, terminal,
reward, update):
embedding = self.network.apply(
x={name: state[:-1]
for name, state in states.items()},
internals=[internal[:-1] for internal in internals],
update=update)
# Both networks can use the same internals, could that be a problem?
# Otherwise need to handle internals indices correctly everywhere
target_embedding = self.target_network.apply(
x={name: state[1:]
for name, state in states.items()},
internals=[internal[1:] for internal in internals],
update=update)
deltas = list()
for name, distribution in self.distributions.items():
distr_params = distribution.parameterize(x=embedding)
target_distr_params = distribution.parameterize(
x=target_embedding) # TODO: separate distribution parameters?
q_value = self.tf_q_value(embedding=embedding,
distr_params=distr_params,
action=actions[name][:-1],
name=name)
if self.double_q_model:
action_taken = distribution.sample(distr_params=distr_params,
deterministic=True)
else:
action_taken = distribution.sample(
distr_params=target_distr_params, deterministic=True)
next_q_value = distribution.state_action_value(
distr_params=target_distr_params, action=action_taken)
delta = self.tf_q_delta(q_value=q_value,
next_q_value=next_q_value,
terminal=terminal[:-1],
reward=reward[:-1])
collapsed_size = util.prod(util.shape(delta)[1:])
delta = tf.reshape(tensor=delta, shape=(-1, collapsed_size))
deltas.append(delta)
# Surrogate loss as the mean squared error between actual observed rewards and expected rewards
loss_per_instance = tf.reduce_mean(input_tensor=tf.concat(
values=deltas, axis=1),
axis=1)
# Optional Huber loss
if self.huber_loss is not None and self.huber_loss > 0.0:
return tf.where(
condition=(tf.abs(x=loss_per_instance) <= self.huber_loss),
x=(0.5 * tf.square(x=loss_per_instance)),
y=(self.huber_loss *
(tf.abs(x=loss_per_instance) - 0.5 * self.huber_loss)))
else:
return tf.square(x=loss_per_instance)
def tf_optimization(self, states, internals, actions, terminal, reward,
update):
optimization = super(QModel, self).tf_optimization(states=states,
internals=internals,
actions=actions,
terminal=terminal,
reward=reward,
update=update)
target_optimization = self.target_optimizer.minimize(
time=self.timestep,
variables=self.target_network.get_variables(),
source_variables=self.network.get_variables())
return tf.group(optimization, target_optimization)
def get_variables(self, include_non_trainable=False):
model_variables = super(
QModel,
self).get_variables(include_non_trainable=include_non_trainable)
if include_non_trainable:
# Target network and optimizer variables only included if 'include_non_trainable' set
target_variables = self.target_network.get_variables(
include_non_trainable=include_non_trainable)
target_optimizer_variables = self.target_optimizer.get_variables()
return model_variables + target_variables + target_optimizer_variables
else:
return model_variables
def get_summaries(self):
return super(
QModel,
self).get_summaries() + self.target_network.get_summaries()
class DPGTargetModel(DistributionModel):
"""
Policy gradient model log likelihood model with target network (e.g. DDPG)
"""
COMPONENT_CRITIC = "critic"
COMPONENT_TARGET_NETWORK = "target_network"
COMPONENT_TARGET_DISTRIBUTION = "target_distribution"
def __init__(self, states, actions, scope, device, saver, summarizer,
execution, batching_capacity, variable_noise,
states_preprocessing, actions_exploration,
reward_preprocessing, update_mode, memory, optimizer,
discount, network, distributions, entropy_regularization,
critic_network, critic_optimizer, target_sync_frequency,
target_update_weight):
self.critic_network_spec = critic_network
self.critic_optimizer_spec = critic_optimizer
self.target_sync_frequency = target_sync_frequency
self.target_update_weight = target_update_weight
# self.network is the actor, self.critic is the critic
self.target_network = None
self.target_network_optimizer = None
self.critic_network = None
self.critic_optimizer = None
self.target_critic_network = None
self.target_critic_optimizer = None
super(DPGTargetModel,
self).__init__(states=states,
actions=actions,
scope=scope,
device=device,
saver=saver,
summarizer=summarizer,
execution=execution,
batching_capacity=batching_capacity,
variable_noise=variable_noise,
states_preprocessing=states_preprocessing,
actions_exploration=actions_exploration,
reward_preprocessing=reward_preprocessing,
update_mode=update_mode,
memory=memory,
optimizer=optimizer,
discount=discount,
network=network,
distributions=distributions,
entropy_regularization=entropy_regularization,
requires_deterministic=True)
assert self.memory_spec["include_next_states"]
assert self.requires_deterministic
def setup_components_and_tf_funcs(self, custom_getter=None):
custom_getter = super(
DPGTargetModel, self).setup_components_and_tf_funcs(custom_getter)
# Target network
self.target_network = Network.from_spec(
spec=self.network_spec,
kwargs=dict(scope='target-network',
summary_labels=self.summary_labels))
# Target network optimizer
self.target_network_optimizer = Synchronization(
sync_frequency=self.target_sync_frequency,
update_weight=self.target_update_weight)
# Target network distributions
self.target_distributions = self.create_distributions()
# Critic
size_t0 = self.critic_network_spec['size_t0']
size_t1 = self.critic_network_spec['size_t1']
self.critic_network = DDPGCriticNetwork(scope='critic',
size_t0=size_t0,
size_t1=size_t1)
self.critic_optimizer = Optimizer.from_spec(
spec=self.critic_optimizer_spec,
kwargs=dict(summary_labels=self.summary_labels))
self.target_critic_network = DDPGCriticNetwork(scope='target-critic',
size_t0=size_t0,
size_t1=size_t1)
# Target critic optimizer
self.target_critic_optimizer = Synchronization(
sync_frequency=self.target_sync_frequency,
update_weight=self.target_update_weight)
self.fn_target_actions_and_internals = tf.make_template(
name_='target-actions-and-internals',
func_=self.tf_target_actions_and_internals,
custom_getter_=custom_getter)
self.fn_predict_target_q = tf.make_template(
name_='predict-target-q',
func_=self.tf_predict_target_q,
custom_getter_=custom_getter)
return custom_getter
def tf_target_actions_and_internals(self,
states,
internals,
deterministic=True):
embedding, internals = self.target_network.apply(
x=states,
internals=internals,
update=tf.constant(value=False),
return_internals=True)
actions = dict()
for name in sorted(self.target_distributions):
distribution = self.target_distributions[name]
distr_params = distribution.parameterize(x=embedding)
actions[name] = distribution.sample(
distr_params=distr_params,
deterministic=tf.logical_or(x=deterministic,
y=self.requires_deterministic))
return actions, internals
def tf_loss_per_instance(self,
states,
internals,
actions,
terminal,
reward,
next_states,
next_internals,
update,
reference=None):
q = self.critic_network.apply(dict(states=states, actions=actions),
internals=internals,
update=update)
return -q
def tf_predict_target_q(self, states, internals, terminal, actions, reward,
update):
q_value = self.target_critic_network.apply(dict(states=states,
actions=actions),
internals=internals,
update=update)
return reward + (
1. - tf.cast(terminal, dtype=tf.float32)) * self.discount * q_value
def tf_optimization(self,
states,
internals,
actions,
terminal,
reward,
next_states=None,
next_internals=None):
update = tf.constant(value=True)
# Predict actions from target actor
next_target_actions, next_target_internals = self.fn_target_actions_and_internals(
states=next_states, internals=next_internals, deterministic=True)
# Predicted Q value of next states
predicted_q = self.fn_predict_target_q(states=next_states,
internals=next_internals,
actions=next_target_actions,
terminal=terminal,
reward=reward,
update=update)
predicted_q = tf.stop_gradient(input=predicted_q)
real_q = self.critic_network.apply(dict(states=states,
actions=actions),
internals=internals,
update=update)
# Update critic
def fn_critic_loss(predicted_q, real_q):
return tf.reduce_mean(tf.square(real_q - predicted_q))
critic_optimization = self.critic_optimizer.minimize(
time=self.timestep,
variables=self.critic_network.get_variables(),
arguments=dict(predicted_q=predicted_q, real_q=real_q),
fn_loss=fn_critic_loss)
# Update actor
predicted_actions, predicted_internals = self.fn_actions_and_internals(
states=states, internals=internals, deterministic=True)
optimization = super(DPGTargetModel, self).tf_optimization(
states=states,
internals=internals,
actions=predicted_actions,
terminal=terminal,
reward=reward,
next_states=next_states,
next_internals=next_internals)
# Update target actor (network) and critic
network_distributions_variables = [
variable for name in sorted(self.distributions)
for variable in self.distributions[name].get_variables(
include_nontrainable=False)
]
target_distributions_variables = [
variable for name in sorted(self.target_distributions)
for variable in self.target_distributions[name].get_variables(
include_nontrainable=False)
]
target_optimization = self.target_network_optimizer.minimize(
time=self.timestep,
variables=self.target_network.get_variables() +
target_distributions_variables,
source_variables=self.network.get_variables() +
network_distributions_variables)
target_critic_optimization = self.target_critic_optimizer.minimize(
time=self.timestep,
variables=self.target_critic_network.get_variables(),
source_variables=self.critic_network.get_variables())
return tf.group(critic_optimization, optimization, target_optimization,
target_critic_optimization)
def get_variables(self,
include_submodules=False,
include_nontrainable=False):
model_variables = super(DPGTargetModel, self).get_variables(
include_submodules=include_submodules,
include_nontrainable=include_nontrainable)
critic_variables = self.critic_network.get_variables(
include_nontrainable=include_nontrainable)
model_variables += critic_variables
if include_nontrainable:
critic_optimizer_variables = self.critic_optimizer.get_variables()
for variable in critic_optimizer_variables:
if variable in model_variables:
model_variables.remove(variable)
model_variables += critic_optimizer_variables
if include_submodules:
target_variables = self.target_network.get_variables(
include_nontrainable=include_nontrainable)
model_variables += target_variables
target_distributions_variables = [
variable for name in sorted(self.target_distributions)
for variable in self.target_distributions[name].get_variables(
include_nontrainable=include_nontrainable)
]
model_variables += target_distributions_variables
target_critic_variables = self.target_critic_network.get_variables(
include_nontrainable=include_nontrainable)
model_variables += target_critic_variables
if include_nontrainable:
target_optimizer_variables = self.target_network_optimizer.get_variables(
)
model_variables += target_optimizer_variables
target_critic_optimizer_variables = self.target_critic_optimizer.get_variables(
)
model_variables += target_critic_optimizer_variables
return model_variables
def get_components(self):
result = dict(super(DPGTargetModel, self).get_components())
result[DPGTargetModel.COMPONENT_CRITIC] = self.critic_network
result[DPGTargetModel.COMPONENT_TARGET_NETWORK] = self.target_network
for name in sorted(self.target_distributions):
result["%s_%s" % (DPGTargetModel.COMPONENT_TARGET_DISTRIBUTION,
name)] = self.target_distributions[name]
if len(self.target_distributions) == 1:
result[DPGTargetModel.
COMPONENT_TARGET_DISTRIBUTION] = self.target_distributions[
next(iter(sorted(self.target_distributions)))]
return result