def __init__(self, clip_ratio, memory_spec=None, **kwargs):
"""
Args:
memory_spec (Optional[dict,Memory]): The spec for the Memory to use for the PPO algorithm.
"""
super(PPOAgent, self).__init__(name=kwargs.pop("name", "ppo-agent"),
**kwargs)
self.train_time_steps = 0
# PPO uses a ring buffer.
self.memory = Memory.from_spec(memory_spec)
self.record_space = Dict(states=self.state_space,
actions=self.action_space,
rewards=float,
terminals=BoolBox(),
add_batch_rank=False)
self.policy = Policy(network_spec=self.neural_network,
action_adapter_spec=None)
self.merger = DictMerger(output_space=self.record_space)
splitter_input_space = copy.deepcopy(self.record_space)
self.splitter = ContainerSplitter(input_space=splitter_input_space)
self.loss_function = PPOLossFunction(clip_ratio=clip_ratio,
discount=self.discount)
self.define_graph_api()
if self.auto_build:
self._build_graph()
self.graph_built = True
def __init__(self,
state_space,
action_space,
discount=0.98,
preprocessing_spec=None,
network_spec=None,
internal_states_space=None,
policy_spec=None,
value_function_spec=None,
execution_spec=None,
optimizer_spec=None,
value_function_optimizer_spec=None,
observe_spec=None,
update_spec=None,
summary_spec=None,
saver_spec=None,
auto_build=True,
name="sac-agent",
double_q=True,
initial_alpha=1.0,
gumbel_softmax_temperature=1.0,
target_entropy=None,
memory_spec=None,
value_function_sync_spec=None):
"""
This is an implementation of the Soft-Actor Critic algorithm.
Paper: http://arxiv.org/abs/1801.01290
Args:
state_space (Union[dict,Space]): Spec dict for the state Space or a direct Space object.
action_space (Union[dict,Space]): Spec dict for the action Space or a direct Space object.
preprocessing_spec (Optional[list,PreprocessorStack]): The spec list for the different necessary states
preprocessing steps or a PreprocessorStack object itself.
discount (float): The discount factor (gamma).
network_spec (Optional[list,NeuralNetwork]): Spec list for a NeuralNetwork Component or the NeuralNetwork
object itself.
internal_states_space (Optional[Union[dict,Space]]): Spec dict for the internal-states Space or a direct
Space object for the Space(s) of the internal (RNN) states.
policy_spec (Optional[dict]): An optional dict for further kwargs passing into the Policy c'tor.
value_function_spec (list, dict, ValueFunction): Neural network specification for baseline or instance
of ValueFunction.
execution_spec (Optional[dict,Execution]): The spec-dict specifying execution settings.
optimizer_spec (Optional[dict,Optimizer]): The spec-dict to create the Optimizer for this Agent.
value_function_optimizer_spec (dict): Optimizer config for value function optimizer. If None, the optimizer
spec for the policy is used (same learning rate and optimizer type).
observe_spec (Optional[dict]): Spec-dict to specify `Agent.observe()` settings.
update_spec (Optional[dict]): Spec-dict to specify `Agent.update()` settings.
summary_spec (Optional[dict]): Spec-dict to specify summary settings.
saver_spec (Optional[dict]): Spec-dict to specify saver settings.
auto_build (Optional[bool]): If True (default), immediately builds the graph using the agent's
graph builder. If false, users must separately call agent.build(). Useful for debugging or analyzing
components before building.
name (str): Some name for this Agent object.
double_q (bool): Whether to train two q networks independently.
initial_alpha (float): "The temperature parameter α determines the
relative importance of the entropy term against the reward".
gumbel_softmax_temperature (float): Temperature parameter for the Gumbel-Softmax distribution used
for discrete actions.
memory_spec (Optional[dict,Memory]): The spec for the Memory to use for the DQN algorithm.
update_spec (dict): Here we can have sync_interval or sync_tau (for the value network update).
"""
# If VF spec is a network spec, wrap with SAC vf type. The VF must concatenate actions and states,
# which can require splitting the network in the case of e.g. conv-inputs.
if isinstance(value_function_spec, list):
value_function_spec = dict(type="sac_value_function",
network_spec=value_function_spec)
self.logger.info("Using default SAC value function.")
elif isinstance(value_function_spec, ValueFunction):
self.logger.info(
"Using value function object {}".format(ValueFunction))
if policy_spec is None:
# Continuous action space: Use squashed normal.
# Discrete: Gumbel-softmax.
policy_spec = dict(
deterministic=False,
distributions_spec=dict(
bounded_distribution_type="squashed",
discrete_distribution_type="gumbel_softmax",
gumbel_softmax_temperature=gumbel_softmax_temperature))
super(SACAgent, self).__init__(
state_space=state_space,
action_space=action_space,
discount=discount,
preprocessing_spec=preprocessing_spec,
network_spec=network_spec,
internal_states_space=internal_states_space,
policy_spec=policy_spec,
value_function_spec=value_function_spec,
execution_spec=execution_spec,
optimizer_spec=optimizer_spec,
value_function_optimizer_spec=value_function_optimizer_spec,
observe_spec=observe_spec,
update_spec=update_spec,
summary_spec=summary_spec,
saver_spec=saver_spec,
auto_build=auto_build,
name=name)
self.double_q = double_q
self.target_entropy = target_entropy
self.initial_alpha = initial_alpha
# Assert that the synch interval is a multiple of the update_interval.
if "sync_interval" in self.update_spec:
if self.update_spec["sync_interval"] / self.update_spec["update_interval"] != \
self.update_spec["sync_interval"] // self.update_spec["update_interval"]:
raise RLGraphError(
"ERROR: sync_interval ({}) must be multiple of update_interval "
"({})!".format(self.update_spec["sync_interval"],
self.update_spec["update_interval"]))
elif "sync_tau" in self.update_spec:
if self.update_spec["sync_tau"] <= 0 or self.update_spec[
"sync_tau"] > 1.0:
raise RLGraphError(
"sync_tau ({}) must be in interval (0.0, 1.0]!".format(
self.update_spec["sync_tau"]))
else:
self.update_spec[
"sync_tau"] = 0.005 # The value mentioned in the paper
# Extend input Space definitions to this Agent's specific API-methods.
preprocessed_state_space = self.preprocessed_state_space.with_batch_rank(
)
reward_space = FloatBox(add_batch_rank=True)
terminal_space = BoolBox(add_batch_rank=True)
#self.iterations = self.update_spec["num_iterations"]
self.batch_size = self.update_spec["batch_size"]
float_action_space = self.action_space.with_batch_rank().map(
mapping=lambda flat_key, space: space.as_one_hot_float_space()
if isinstance(space, IntBox) else space)
self.input_spaces.update(
dict(env_actions=self.action_space.with_batch_rank(),
actions=float_action_space,
preprocessed_states=preprocessed_state_space,
rewards=reward_space,
terminals=terminal_space,
next_states=preprocessed_state_space,
states=self.state_space.with_batch_rank(add_batch_rank=True),
batch_size=int,
importance_weights=FloatBox(add_batch_rank=True),
deterministic=bool,
weights="variables:{}".format(self.policy.scope)))
if value_function_sync_spec is None:
value_function_sync_spec = SyncSpecification(
sync_interval=self.update_spec["sync_interval"] //
self.update_spec["update_interval"],
sync_tau=self.update_spec["sync_tau"]
if "sync_tau" in self.update_spec else 5e-3)
self.memory = Memory.from_spec(memory_spec)
self.alpha_optimizer = self.optimizer.copy(
scope="alpha-" +
self.optimizer.scope) if self.target_entropy is not None else None
self.root_component = SACAgentComponent(
agent=self,
policy=self.policy,
q_function=self.value_function,
preprocessor=self.preprocessor,
memory=self.memory,
discount=self.discount,
initial_alpha=self.initial_alpha,
target_entropy=target_entropy,
optimizer=self.optimizer,
vf_optimizer=self.value_function_optimizer,
alpha_optimizer=self.alpha_optimizer,
q_sync_spec=value_function_sync_spec,
num_q_functions=2 if self.double_q is True else 1)
extra_optimizers = [self.value_function_optimizer]
if self.alpha_optimizer is not None:
extra_optimizers.append(self.alpha_optimizer)
self.build_options = dict(optimizers=extra_optimizers)
if self.auto_build:
self._build_graph([self.root_component],
self.input_spaces,
optimizer=self.optimizer,
batch_size=self.update_spec["batch_size"],
build_options=self.build_options)
self.graph_built = True
def __init__(self,
state_space,
action_space,
discount=0.98,
preprocessing_spec=None,
network_spec=None,
internal_states_space=None,
policy_spec=None,
value_function_spec=None,
execution_spec=None,
optimizer_spec=None,
value_function_optimizer_spec=None,
observe_spec=None,
update_spec=None,
summary_spec=None,
saver_spec=None,
auto_build=True,
name="actor-critic-agent",
gae_lambda=1.0,
clip_rewards=0.0,
sample_episodes=False,
weight_entropy=None,
memory_spec=None):
"""
Args:
state_space (Union[dict,Space]): Spec dict for the state Space or a direct Space object.
action_space (Union[dict,Space]): Spec dict for the action Space or a direct Space object.
preprocessing_spec (Optional[list,PreprocessorStack]): The spec list for the different necessary states
preprocessing steps or a PreprocessorStack object itself.
discount (float): The discount factor (gamma).
network_spec (Optional[list,NeuralNetwork]): Spec list for a NeuralNetwork Component or the NeuralNetwork
object itself.
internal_states_space (Optional[Union[dict,Space]]): Spec dict for the internal-states Space or a direct
Space object for the Space(s) of the internal (RNN) states.
policy_spec (Optional[dict]): An optional dict for further kwargs passing into the Policy c'tor.
value_function_spec (list, dict, ValueFunction): Neural network specification for baseline or instance
of ValueFunction.
execution_spec (Optional[dict,Execution]): The spec-dict specifying execution settings.
optimizer_spec (Optional[dict,Optimizer]): The spec-dict to create the Optimizer for this Agent.
value_function_optimizer_spec (dict): Optimizer config for value function optimizer. If None, the optimizer
spec for the policy is used (same learning rate and optimizer type).
observe_spec (Optional[dict]): Spec-dict to specify `Agent.observe()` settings.
update_spec (Optional[dict]): Spec-dict to specify `Agent.update()` settings.
summary_spec (Optional[dict]): Spec-dict to specify summary settings.
saver_spec (Optional[dict]): Spec-dict to specify saver settings.
auto_build (Optional[bool]): If True (default), immediately builds the graph using the agent's
graph builder. If false, users must separately call agent.build(). Useful for debugging or analyzing
components before building.
name (str): Some name for this Agent object.
gae_lambda (float): Lambda for generalized advantage estimation.
clip_rewards (float): Reward clip value. If not 0, rewards will be clipped into this range.
sample_episodes (bool): If true, the update method interprets the batch_size as the number of
episodes to fetch from the memory. If false, batch_size will refer to the number of time-steps. This
is especially relevant for environments where episode lengths may vastly differ throughout training. For
example, in CartPole, a losing episode is typically 10 steps, and a winning episode 200 steps.
weight_entropy (float): The coefficient used for the entropy regularization term (L[E]).
memory_spec (Optional[dict,Memory]): The spec for the Memory to use. Should typically be
a ring-buffer.
"""
# Set policy to stochastic.
if policy_spec is not None:
policy_spec["deterministic"] = False
else:
policy_spec = dict(deterministic=False)
super(ActorCriticAgent, self).__init__(
state_space=state_space,
action_space=action_space,
discount=discount,
preprocessing_spec=preprocessing_spec,
network_spec=network_spec,
internal_states_space=internal_states_space,
policy_spec=policy_spec,
value_function_spec=value_function_spec,
execution_spec=execution_spec,
optimizer_spec=optimizer_spec,
value_function_optimizer_spec=value_function_optimizer_spec,
observe_spec=observe_spec,
update_spec=update_spec,
summary_spec=summary_spec,
saver_spec=saver_spec,
name=name,
auto_build=auto_build)
self.sample_episodes = sample_episodes
# Extend input Space definitions to this Agent's specific API-methods.
preprocessed_state_space = self.preprocessed_state_space.with_batch_rank(
)
reward_space = FloatBox(add_batch_rank=True)
terminal_space = BoolBox(add_batch_rank=True)
self.input_spaces.update(
dict(actions=self.action_space.with_batch_rank(),
policy_weights="variables:{}".format(self.policy.scope),
deterministic=bool,
preprocessed_states=preprocessed_state_space,
rewards=reward_space,
terminals=terminal_space,
sequence_indices=BoolBox(add_batch_rank=True)))
# The merger to merge inputs into one record Dict going into the memory.
self.merger = ContainerMerger("states", "actions", "rewards",
"terminals")
self.memory = Memory.from_spec(memory_spec)
assert isinstance(self.memory, RingBuffer), \
"ERROR: Actor-critic memory must be ring-buffer for episode-handling."
# The splitter for splitting up the records coming from the memory.
self.splitter = ContainerSplitter("states", "actions", "rewards",
"terminals")
self.gae_function = GeneralizedAdvantageEstimation(
gae_lambda=gae_lambda,
discount=self.discount,
clip_rewards=clip_rewards)
self.loss_function = ActorCriticLossFunction(
weight_entropy=weight_entropy)
# Add all our sub-components to the core.
sub_components = [
self.preprocessor, self.merger, self.memory, self.splitter,
self.policy, self.loss_function, self.optimizer,
self.value_function, self.value_function_optimizer,
self.gae_function
]
self.root_component.add_components(*sub_components)
# Define the Agent's (root-Component's) API.
self.define_graph_api()
self.build_options = dict(vf_optimizer=self.value_function_optimizer)
if self.auto_build:
self._build_graph([self.root_component],
self.input_spaces,
optimizer=self.optimizer,
batch_size=self.update_spec["batch_size"],
build_options=self.build_options)
self.graph_built = True
def __init__(self, expert_margin=0.5, supervised_weight=1.0, double_q=True, dueling_q=True,
huber_loss=False, n_step=1, shared_container_action_target=True,
memory_spec=None, demo_memory_spec=None,
demo_sample_ratio=0.2, store_last_memory_batch=False, store_last_q_table=False, **kwargs):
# TODO Most of this is DQN duplicate but the way the loss function is instantiated, inheriting
# from DQN does not work well.
"""
Args:
expert_margin (float): The expert margin enforces a distance in Q-values between expert action and
all other actions.
supervised_weight (float): Indicates weight of the expert loss.
double_q (bool): Whether to use the double DQN loss function (see [2]).
dueling_q (bool): Whether to use a dueling layer in the ActionAdapter (see [3]).
huber_loss (bool) : Whether to apply a Huber loss. (see [4]).
n_step (Optional[int]): n-step adjustment to discounting.
memory_spec (Optional[dict,Memory]): The spec for the Memory to use.
demo_memory_spec (Optional[dict,Memory]): The spec for the Demo-Memory to use.
store_last_memory_batch (bool): Whether to store the last pulled batch from the memory in
`self.last_memory_batch` for debugging purposes.
Default: False.
store_last_q_table (bool): Whether to store the Q(s,a) values for the last received batch
(memory or external) in `self.last_q_table` for debugging purposes.
Default: False.
"""
# Fix action-adapter before passing it to the super constructor.
policy_spec = kwargs.pop("policy_spec", dict())
# Use a DuelingPolicy (instead of a basic Policy) if option is set.
if dueling_q is True:
policy_spec["type"] = "dueling-policy"
# Give us some default state-value nodes.
if "units_state_value_stream" not in policy_spec:
policy_spec["units_state_value_stream"] = 128
super(DQFDAgent, self).__init__(
policy_spec=policy_spec, name=kwargs.pop("name", "dqfd-agent"), **kwargs
)
# Assert that the synch interval is a multiple of the update_interval.
if self.update_spec["sync_interval"] / self.update_spec["update_interval"] != \
self.update_spec["sync_interval"] // self.update_spec["update_interval"]:
raise RLGraphError(
"ERROR: sync_interval ({}) must be multiple of update_interval "
"({})!".format(self.update_spec["sync_interval"], self.update_spec["update_interval"])
)
self.double_q = double_q
self.dueling_q = dueling_q
self.huber_loss = huber_loss
self.demo_batch_size = int(demo_sample_ratio * self.update_spec['batch_size'] / (1.0 - demo_sample_ratio))
self.shared_container_action_target = shared_container_action_target
# Debugging tools.
self.store_last_memory_batch = store_last_memory_batch
self.last_memory_batch = None
self.store_last_q_table = store_last_q_table
self.last_q_table = None
# Extend input Space definitions to this Agent's specific API-methods.
preprocessed_state_space = self.preprocessed_state_space.with_batch_rank()
reward_space = FloatBox(add_batch_rank=True)
terminal_space = BoolBox(add_batch_rank=True)
weight_space = FloatBox(add_batch_rank=True)
self.input_spaces.update(dict(
actions=self.action_space.with_batch_rank(),
policy_weights="variables:{}".format(self.policy.scope),
time_step=int,
use_exploration=bool,
demo_batch_size=int,
apply_demo_loss=bool,
preprocessed_states=preprocessed_state_space,
rewards=reward_space,
terminals=terminal_space,
next_states=preprocessed_state_space,
preprocessed_next_states=preprocessed_state_space,
importance_weights=weight_space
))
# The merger to merge inputs into one record Dict going into the memory.
self.merger = ContainerMerger("states", "actions", "rewards", "next_states", "terminals")
# The replay memory.
self.memory = Memory.from_spec(memory_spec)
# Cannot have same default name.
demo_memory_spec["scope"] = "demo-memory"
self.demo_memory = Memory.from_spec(demo_memory_spec)
# The splitter for splitting up the records from the memories.
self.splitter = ContainerSplitter("states", "actions", "rewards", "terminals", "next_states")
# Copy our Policy (target-net), make target-net synchronizable.
self.target_policy = self.policy.copy(scope="target-policy", trainable=False)
# Number of steps since the last target-net synching from the main policy.
self.steps_since_target_net_sync = 0
use_importance_weights = isinstance(self.memory, PrioritizedReplay)
self.loss_function = DQFDLossFunction(
expert_margin=expert_margin, supervised_weight=supervised_weight,
discount=self.discount, double_q=self.double_q, huber_loss=self.huber_loss,
shared_container_action_target=shared_container_action_target,
importance_weights=use_importance_weights, n_step=n_step
)
# Add all our sub-components to the core.
self.root_component.add_components(
self.preprocessor, self.merger, self.memory, self.demo_memory, self.splitter, self.policy,
self.target_policy, self.exploration, self.loss_function, self.optimizer
)
# Define the Agent's (root-Component's) API.
self.define_graph_api()
if self.auto_build:
self._build_graph([self.root_component], self.input_spaces, optimizer=self.optimizer,
batch_size=self.update_spec["batch_size"])
self.graph_built = True
def __init__(self, double_q=True, dueling_q=True, huber_loss=False, n_step=1, memory_spec=None,
store_last_memory_batch=False, store_last_q_table=False, **kwargs):
"""
Args:
double_q (bool): Whether to use the double DQN loss function (see [2]).
dueling_q (bool): Whether to use a dueling layer in the ActionAdapter (see [3]).
huber_loss (bool) : Whether to apply a Huber loss. (see [4]).
n_step (Optional[int]): n-step adjustment to discounting.
memory_spec (Optional[dict,Memory]): The spec for the Memory to use for the DQN algorithm.
store_last_memory_batch (bool): Whether to store the last pulled batch from the memory in
`self.last_memory_batch` for debugging purposes.
Default: False.
store_last_q_table (bool): Whether to store the Q(s,a) values for the last received batch
(memory or external) in `self.last_q_table` for debugging purposes.
Default: False.
"""
# Fix action-adapter before passing it to the super constructor.
action_adapter_spec = kwargs.pop("action_adapter_spec", dict())
# Use a DuelingActionAdapter (instead of a basic ActionAdapter) if option is set.
if dueling_q is True:
action_adapter_spec["type"] = "dueling-action-adapter"
assert "units_state_value_stream" in action_adapter_spec
assert "units_advantage_stream" in action_adapter_spec
super(DQNAgent, self).__init__(
action_adapter_spec=action_adapter_spec, name=kwargs.pop("name", "dqn-agent"), **kwargs
)
self.double_q = double_q
self.dueling_q = dueling_q
self.huber_loss = huber_loss
# Debugging tools.
self.store_last_memory_batch = store_last_memory_batch
self.last_memory_batch = None
self.store_last_q_table = store_last_q_table
self.last_q_table = None
# Extend input Space definitions to this Agent's specific API-methods.
preprocessed_state_space = self.preprocessed_state_space.with_batch_rank()
reward_space = FloatBox(add_batch_rank=True)
terminal_space = BoolBox(add_batch_rank=True)
weight_space = FloatBox(add_batch_rank=True)
self.input_spaces.update(dict(
actions=self.action_space.with_batch_rank(),
weights="variables:policy",
time_step=int,
use_exploration=bool,
preprocessed_states=preprocessed_state_space,
rewards=reward_space,
terminals=terminal_space,
next_states=preprocessed_state_space,
preprocessed_next_states=preprocessed_state_space,
importance_weights=weight_space,
# TODO: This is currently necessary for multi-GPU handling (as the update_from_external_batch
# TODO: gets overridden by a generic function with args=*inputs)
#inputs=[preprocessed_state_space, self.action_space.with_batch_rank(), reward_space, terminal_space,
# preprocessed_state_space, weight_space]
))
# The merger to merge inputs into one record Dict going into the memory.
self.merger = DictMerger("states", "actions", "rewards", "next_states", "terminals")
# The replay memory.
self.memory = Memory.from_spec(memory_spec)
# The splitter for splitting up the records coming from the memory.
self.splitter = ContainerSplitter("states", "actions", "rewards", "terminals", "next_states")
# Copy our Policy (target-net), make target-net synchronizable.
self.target_policy = self.policy.copy(scope="target-policy", trainable=False)
self.target_policy.add_components(Synchronizable(), expose_apis="sync")
# Number of steps since the last target-net synching from the main policy.
self.steps_since_target_net_sync = 0
use_importance_weights = isinstance(self.memory, PrioritizedReplay)
self.loss_function = DQNLossFunction(
discount=self.discount, double_q=self.double_q, huber_loss=self.huber_loss,
importance_weights=use_importance_weights, n_step=n_step
)
# Add all our sub-components to the core.
sub_components = [self.preprocessor, self.merger, self.memory, self.splitter, self.policy,
self.target_policy, self.exploration, self.loss_function, self.optimizer]
self.root_component.add_components(*sub_components)
# Define the Agent's (root-Component's) API.
self.define_graph_api("policy", "preprocessor-stack", self.optimizer.scope, *sub_components)
# markup = get_graph_markup(self.graph_builder.root_component)
# print(markup)
if self.auto_build:
self._build_graph([self.root_component], self.input_spaces, optimizer=self.optimizer,
batch_size=self.update_spec["batch_size"])
self.graph_built = True
def __init__(self,
state_space,
action_space,
discount=0.98,
preprocessing_spec=None,
network_spec=None,
internal_states_space=None,
policy_spec=None,
value_function_spec=None,
execution_spec=None,
optimizer_spec=None,
value_function_optimizer_spec=None,
observe_spec=None,
update_spec=None,
summary_spec=None,
saver_spec=None,
auto_build=True,
name="ppo-agent",
clip_ratio=0.2,
gae_lambda=1.0,
clip_rewards=0.0,
value_function_clipping=None,
standardize_advantages=False,
sample_episodes=True,
weight_entropy=None,
memory_spec=None):
"""
Args:
state_space (Union[dict,Space]): Spec dict for the state Space or a direct Space object.
action_space (Union[dict,Space]): Spec dict for the action Space or a direct Space object.
preprocessing_spec (Optional[list,PreprocessorStack]): The spec list for the different necessary states
preprocessing steps or a PreprocessorStack object itself.
discount (float): The discount factor (gamma).
network_spec (Optional[list,NeuralNetwork]): Spec list for a NeuralNetwork Component or the NeuralNetwork
object itself.
internal_states_space (Optional[Union[dict,Space]]): Spec dict for the internal-states Space or a direct
Space object for the Space(s) of the internal (RNN) states.
policy_spec (Optional[dict]): An optional dict for further kwargs passing into the Policy c'tor.
value_function_spec (list, dict, ValueFunction): Neural network specification for baseline or instance
of ValueFunction.
execution_spec (Optional[dict,Execution]): The spec-dict specifying execution settings.
optimizer_spec (Optional[dict,Optimizer]): The spec-dict to create the Optimizer for this Agent.
value_function_optimizer_spec (dict): Optimizer config for value function optimizer. If None, the optimizer
spec for the policy is used (same learning rate and optimizer type).
observe_spec (Optional[dict]): Spec-dict to specify `Agent.observe()` settings.
update_spec (Optional[dict]): Spec-dict to specify `Agent.update()` settings.
summary_spec (Optional[dict]): Spec-dict to specify summary settings.
saver_spec (Optional[dict]): Spec-dict to specify saver settings.
auto_build (Optional[bool]): If True (default), immediately builds the graph using the agent's
graph builder. If false, users must separately call agent.build(). Useful for debugging or analyzing
components before building.
name (str): Some name for this Agent object.
clip_ratio (float): Clipping parameter for likelihood ratio.
gae_lambda (float): Lambda for generalized advantage estimation.
clip_rewards (float): Reward clipping value. If not 0, rewards will be clipped within a +/- `clip_rewards`
range.
value_function_clipping (Optional[float]): If not None, uses clipped value function objective. If None,
uses simple value function objective.
standardize_advantages (bool): If true, standardize advantage values in update.
sample_episodes (bool): If True, the update method interprets the batch_size as the number of
episodes to fetch from the memory. If False, batch_size will refer to the number of time-steps. This
is especially relevant for environments where episode lengths may vastly differ throughout training. For
example, in CartPole, a losing episode is typically 10 steps, and a winning episode 200 steps.
weight_entropy (float): The coefficient used for the entropy regularization term (L[E]).
memory_spec (Optional[dict,Memory]): The spec for the Memory to use. Should typically be
a ring-buffer.
"""
if policy_spec is not None:
policy_spec["deterministic"] = False
else:
policy_spec = dict(deterministic=False)
super(PPOAgent, self).__init__(
state_space=state_space,
action_space=action_space,
discount=discount,
preprocessing_spec=preprocessing_spec,
network_spec=network_spec,
internal_states_space=internal_states_space,
policy_spec=policy_spec,
value_function_spec=value_function_spec,
execution_spec=execution_spec,
optimizer_spec=optimizer_spec,
value_function_optimizer_spec=value_function_optimizer_spec,
observe_spec=observe_spec,
update_spec=update_spec,
summary_spec=summary_spec,
saver_spec=saver_spec,
name=name,
auto_build=auto_build)
self.sample_episodes = sample_episodes
# TODO: Have to manually set it here for multi-GPU synchronizer to know its number
# TODO: of return values when calling _graph_fn_calculate_update_from_external_batch.
# self.root_component.graph_fn_num_outputs["_graph_fn_update_from_external_batch"] = 4
# Extend input Space definitions to this Agent's specific API-methods.
preprocessed_state_space = self.preprocessed_state_space.with_batch_rank(
)
reward_space = FloatBox(add_batch_rank=True)
terminal_space = BoolBox(add_batch_rank=True)
self.input_spaces.update(
dict(actions=self.action_space.with_batch_rank(),
policy_weights="variables:policy",
value_function_weights="variables:value-function",
deterministic=bool,
preprocessed_states=preprocessed_state_space,
rewards=reward_space,
terminals=terminal_space,
sequence_indices=BoolBox(add_batch_rank=True),
apply_postprocessing=bool))
# The merger to merge inputs into one record Dict going into the memory.
self.merger = ContainerMerger("states", "actions", "rewards",
"terminals")
self.memory = Memory.from_spec(memory_spec)
assert isinstance(
self.memory, RingBuffer
), "ERROR: PPO memory must be ring-buffer for episode-handling!"
# Make sure the python buffer is not larger than our memory capacity.
assert self.observe_spec["buffer_size"] <= self.memory.capacity, \
"ERROR: Buffer's size ({}) in `observe_spec` must be smaller or equal to the memory's capacity ({})!". \
format(self.observe_spec["buffer_size"], self.memory.capacity)
# The splitter for splitting up the records coming from the memory.
self.standardize_advantages = standardize_advantages
self.gae_function = GeneralizedAdvantageEstimation(
gae_lambda=gae_lambda,
discount=self.discount,
clip_rewards=clip_rewards)
self.loss_function = PPOLossFunction(
clip_ratio=clip_ratio,
value_function_clipping=value_function_clipping,
weight_entropy=weight_entropy)
self.iterations = self.update_spec["num_iterations"]
self.sample_size = self.update_spec["sample_size"]
self.batch_size = self.update_spec["batch_size"]
# Add all our sub-components to the core.
self.root_component.add_components(
self.preprocessor, self.merger, self.memory, self.policy,
self.exploration, self.loss_function, self.optimizer,
self.value_function, self.value_function_optimizer,
self.vars_merger, self.vars_splitter, self.gae_function)
# Define the Agent's (root-Component's) API.
self.define_graph_api()
self.build_options = dict(vf_optimizer=self.value_function_optimizer)
if self.auto_build:
self._build_graph(
[self.root_component],
self.input_spaces,
optimizer=self.optimizer,
# Important: Use sample-size, not batch-size as the sub-samples (from a batch) are the ones that get
# multi-gpu-split.
batch_size=self.update_spec["sample_size"],
build_options=self.build_options)
self.graph_built = True
def __init__(self,
gae_lambda=1.0,
sample_episodes=False,
weight_entropy=None,
memory_spec=None,
**kwargs):
"""
Args:
gae_lambda (float): Lambda for generalized advantage estimation.
sample_episodes (bool): If true, the update method interprets the batch_size as the number of
episodes to fetch from the memory. If false, batch_size will refer to the number of time-steps. This
is especially relevant for environments where episode lengths may vastly differ throughout training. For
example, in CartPole, a losing episode is typically 10 steps, and a winning episode 200 steps.
weight_entropy (float): The coefficient used for the entropy regularization term (L[E]).
memory_spec (Optional[dict,Memory]): The spec for the Memory to use. Should typically be
a ring-buffer.
"""
super(ActorCriticAgent, self).__init__(
policy_spec=dict(deterministic=False), # Set policy to stochastic.
name=kwargs.pop("name", "actor-critic-agent"),
**kwargs)
self.sample_episodes = sample_episodes
# Extend input Space definitions to this Agent's specific API-methods.
preprocessed_state_space = self.preprocessed_state_space.with_batch_rank(
)
reward_space = FloatBox(add_batch_rank=True)
terminal_space = BoolBox(add_batch_rank=True)
self.input_spaces.update(
dict(actions=self.action_space.with_batch_rank(),
policy_weights="variables:{}".format(self.policy.scope),
deterministic=bool,
preprocessed_states=preprocessed_state_space,
rewards=reward_space,
terminals=terminal_space,
sequence_indices=BoolBox(add_batch_rank=True)))
# The merger to merge inputs into one record Dict going into the memory.
self.merger = DictMerger("states", "actions", "rewards", "terminals")
self.memory = Memory.from_spec(memory_spec)
assert isinstance(self.memory, RingBuffer),\
"ERROR: Actor-critic memory must be ring-buffer for episode-handling."
# The splitter for splitting up the records coming from the memory.
self.splitter = ContainerSplitter("states", "actions", "rewards",
"terminals")
self.loss_function = ActorCriticLossFunction(
discount=self.discount,
gae_lambda=gae_lambda,
weight_entropy=weight_entropy)
# Add all our sub-components to the core.
sub_components = [
self.preprocessor, self.merger, self.memory, self.splitter,
self.policy, self.loss_function, self.optimizer,
self.value_function, self.value_function_optimizer
]
self.root_component.add_components(*sub_components)
# Define the Agent's (root-Component's) API.
self.define_graph_api()
if self.auto_build:
self._build_graph(
[self.root_component],
self.input_spaces,
optimizer=self.optimizer,
batch_size=self.update_spec["batch_size"],
build_options=dict(vf_optimizer=self.value_function_optimizer))
self.graph_built = True
def __init__(self,
double_q=True,
dueling_q=True,
huber_loss=False,
n_step=1,
shared_container_action_target=True,
memory_spec=None,
store_last_memory_batch=False,
store_last_q_table=False,
**kwargs):
"""
Args:
double_q (bool): Whether to use the double DQN loss function (see [2]).
dueling_q (bool): Whether to use a dueling layer in the ActionAdapter (see [3]).
huber_loss (bool) : Whether to apply a Huber loss. (see [4]).
n_step (Optional[int]): n-step adjustment to discounting.
memory_spec (Optional[dict,Memory]): The spec for the Memory to use for the DQN algorithm.
store_last_memory_batch (bool): Whether to store the last pulled batch from the memory in
`self.last_memory_batch` for debugging purposes.
Default: False.
store_last_q_table (bool): Whether to store the Q(s,a) values for the last received batch
(memory or external) in `self.last_q_table` for debugging purposes.
Default: False.
"""
# Fix action-adapter before passing it to the super constructor.
policy_spec = kwargs.pop("policy_spec", dict())
# Use a DuelingPolicy (instead of a basic Policy) if option is set.
if dueling_q is True:
policy_spec["type"] = "dueling-policy"
# Give us some default state-value nodes.
if "units_state_value_stream" not in policy_spec:
policy_spec["units_state_value_stream"] = 128
super(DQNAgent, self).__init__(policy_spec=policy_spec,
name=kwargs.pop("name", "dqn-agent"),
**kwargs)
# TODO: Have to manually set it here for multi-GPU synchronizer to know its number
# TODO: of return values when calling _graph_fn_calculate_update_from_external_batch.
#self.root_component.graph_fn_num_outputs["_graph_fn_update_from_external_batch"] = 4
# Assert that the synch interval is a multiple of the update_interval.
if self.update_spec["sync_interval"] / self.update_spec["update_interval"] != \
self.update_spec["sync_interval"] // self.update_spec["update_interval"]:
raise RLGraphError(
"ERROR: sync_interval ({}) must be multiple of update_interval "
"({})!".format(self.update_spec["sync_interval"],
self.update_spec["update_interval"]))
self.double_q = double_q
self.dueling_q = dueling_q
self.huber_loss = huber_loss
self.shared_container_action_target = shared_container_action_target
# Debugging tools.
self.store_last_memory_batch = store_last_memory_batch
self.last_memory_batch = None
self.store_last_q_table = store_last_q_table
self.last_q_table = None
# Extend input Space definitions to this Agent's specific API-methods.
preprocessed_state_space = self.preprocessed_state_space.with_batch_rank(
)
reward_space = FloatBox(add_batch_rank=True)
terminal_space = BoolBox(add_batch_rank=True)
weight_space = FloatBox(add_batch_rank=True)
self.input_spaces.update(
dict(
actions=self.action_space.with_batch_rank(),
# weights will have a Space derived from the vars of policy.
policy_weights="variables:{}".format(self.policy.scope),
time_step=int,
use_exploration=bool,
preprocessed_states=preprocessed_state_space,
rewards=reward_space,
terminals=terminal_space,
next_states=preprocessed_state_space,
preprocessed_next_states=preprocessed_state_space,
importance_weights=weight_space,
))
if self.value_function is not None:
self.input_spaces[
"value_function_weights"] = "variables:{}".format(
self.value_function.scope),
# The merger to merge inputs into one record Dict going into the memory.
self.merger = DictMerger("states", "actions", "rewards", "next_states",
"terminals")
# The replay memory.
self.memory = Memory.from_spec(memory_spec)
# The splitter for splitting up the records coming from the memory.
self.splitter = ContainerSplitter("states", "actions", "rewards",
"terminals", "next_states")
# Make sure the python buffer is not larger than our memory capacity.
assert self.observe_spec["buffer_size"] <= self.memory.capacity,\
"ERROR: Buffer's size ({}) in `observe_spec` must be smaller or equal to the memory's capacity ({})!".\
format(self.observe_spec["buffer_size"], self.memory.capacity)
# Copy our Policy (target-net), make target-net synchronizable.
self.target_policy = self.policy.copy(scope="target-policy",
trainable=False)
# Number of steps since the last target-net synching from the main policy.
self.steps_since_target_net_sync = 0
use_importance_weights = isinstance(self.memory, PrioritizedReplay)
self.loss_function = DQNLossFunction(
discount=self.discount,
double_q=self.double_q,
huber_loss=self.huber_loss,
shared_container_action_target=shared_container_action_target,
importance_weights=use_importance_weights,
n_step=n_step)
self.root_component.add_components(
self.preprocessor,
self.merger,
self.memory,
self.splitter,
self.policy,
self.target_policy,
self.value_function,
self.value_function_optimizer, # <- should both be None for DQN
self.exploration,
self.loss_function,
self.optimizer,
self.vars_merger,
self.vars_splitter)
# Define the Agent's (root-Component's) API.
self.define_graph_api()
# markup = get_graph_markup(self.graph_builder.root_component)
# print(markup)
if self.auto_build:
self._build_graph([self.root_component],
self.input_spaces,
optimizer=self.optimizer,
batch_size=self.update_spec["batch_size"])
self.graph_built = True
def __init__(
self,
state_space,
action_space,
discount=0.98,
preprocessing_spec=None,
network_spec=None,
internal_states_space=None,
policy_spec=None,
exploration_spec=None,
execution_spec=None,
optimizer_spec=None,
observe_spec=None,
update_spec=None,
summary_spec=None,
saver_spec=None,
auto_build=True,
name="dqfd-agent",
expert_margin=0.5,
supervised_weight=1.0,
double_q=True,
dueling_q=True,
huber_loss=False,
n_step=1,
shared_container_action_target=False,
memory_spec=None,
demo_memory_spec=None,
demo_sample_ratio=0.2,
):
"""
Args:
state_space (Union[dict,Space]): Spec dict for the state Space or a direct Space object.
action_space (Union[dict,Space]): Spec dict for the action Space or a direct Space object.
preprocessing_spec (Optional[list,PreprocessorStack]): The spec list for the different necessary states
preprocessing steps or a PreprocessorStack object itself.
discount (float): The discount factor (gamma).
network_spec (Optional[list,NeuralNetwork]): Spec list for a NeuralNetwork Component or the NeuralNetwork
object itself.
internal_states_space (Optional[Union[dict,Space]]): Spec dict for the internal-states Space or a direct
Space object for the Space(s) of the internal (RNN) states.
policy_spec (Optional[dict]): An optional dict for further kwargs passing into the Policy c'tor.
exploration_spec (Optional[dict]): The spec-dict to create the Exploration Component.
execution_spec (Optional[dict,Execution]): The spec-dict specifying execution settings.
optimizer_spec (Optional[dict,Optimizer]): The spec-dict to create the Optimizer for this Agent.
observe_spec (Optional[dict]): Spec-dict to specify `Agent.observe()` settings.
update_spec (Optional[dict]): Spec-dict to specify `Agent.update()` settings.
summary_spec (Optional[dict]): Spec-dict to specify summary settings.
saver_spec (Optional[dict]): Spec-dict to specify saver settings.
auto_build (Optional[bool]): If True (default), immediately builds the graph using the agent's
graph builder. If false, users must separately call agent.build(). Useful for debugging or analyzing
components before building.
name (str): Some name for this Agent object.
expert_margin (float): The expert margin enforces a distance in Q-values between expert action and
all other actions.
supervised_weight (float): Indicates weight of the expert loss.
double_q (bool): Whether to use the double DQN loss function (see [2]).
dueling_q (bool): Whether to use a dueling layer in the ActionAdapter (see [3]).
huber_loss (bool) : Whether to apply a Huber loss. (see [4]).
n_step (Optional[int]): n-step adjustment to discounting.
memory_spec (Optional[dict,Memory]): The spec for the Memory to use.
demo_memory_spec (Optional[dict,Memory]): The spec for the Demo-Memory to use.
"""
# Fix action-adapter before passing it to the super constructor.
# Use a DuelingPolicy (instead of a basic Policy) if option is set.
if dueling_q is True:
if policy_spec is None:
policy_spec = {}
policy_spec["type"] = "dueling-policy"
# Give us some default state-value nodes.
if "units_state_value_stream" not in policy_spec:
policy_spec["units_state_value_stream"] = 128
super(DQFDAgent, self).__init__(
state_space=state_space,
action_space=action_space,
discount=discount,
preprocessing_spec=preprocessing_spec,
network_spec=network_spec,
internal_states_space=internal_states_space,
policy_spec=policy_spec,
exploration_spec=exploration_spec,
execution_spec=execution_spec,
optimizer_spec=optimizer_spec,
observe_spec=observe_spec,
update_spec=update_spec,
summary_spec=summary_spec,
saver_spec=saver_spec,
auto_build=auto_build,
name=name
)
# Assert that the synch interval is a multiple of the update_interval.
if self.update_spec["sync_interval"] / self.update_spec["update_interval"] != \
self.update_spec["sync_interval"] // self.update_spec["update_interval"]:
raise RLGraphError(
"ERROR: sync_interval ({}) must be multiple of update_interval "
"({})!".format(self.update_spec["sync_interval"], self.update_spec["update_interval"])
)
self.double_q = double_q
self.dueling_q = dueling_q
self.huber_loss = huber_loss
self.expert_margin = expert_margin
self.batch_size = self.update_spec["batch_size"]
self.default_margins = np.asarray([self.expert_margin] * self.batch_size)
self.demo_batch_size = int(demo_sample_ratio * self.update_spec["batch_size"] / (1.0 - demo_sample_ratio))
self.demo_margins = np.asarray([self.expert_margin] * self.demo_batch_size)
self.shared_container_action_target = shared_container_action_target
# Extend input Space definitions to this Agent's specific API-methods.
preprocessed_state_space = self.preprocessed_state_space.with_batch_rank()
reward_space = FloatBox(add_batch_rank=True)
terminal_space = BoolBox(add_batch_rank=True)
weight_space = FloatBox(add_batch_rank=True)
self.input_spaces.update(dict(
actions=self.action_space.with_batch_rank(),
policy_weights="variables:{}".format(self.policy.scope),
time_step=int,
use_exploration=bool,
demo_batch_size=int,
apply_demo_loss=bool,
preprocessed_states=preprocessed_state_space,
rewards=reward_space,
terminals=terminal_space,
expert_margins=FloatBox(add_batch_rank=True),
next_states=preprocessed_state_space,
preprocessed_next_states=preprocessed_state_space,
importance_weights=weight_space
))
# The merger to merge inputs into one record Dict going into the memory.
self.merger = ContainerMerger("states", "actions", "rewards", "next_states", "terminals")
# The replay memory.
self.memory = Memory.from_spec(memory_spec)
# Cannot have same default name.
demo_memory_spec["scope"] = "demo-memory"
self.demo_memory = Memory.from_spec(demo_memory_spec)
# The splitter for splitting up the records from the memories.
self.splitter = ContainerSplitter("states", "actions", "rewards", "terminals", "next_states")
# Copy our Policy (target-net), make target-net synchronizable.
self.target_policy = self.policy.copy(scope="target-policy", trainable=False)
# Number of steps since the last target-net synching from the main policy.
self.steps_since_target_net_sync = 0
self.use_importance_weights = isinstance(self.memory, PrioritizedReplay)
self.loss_function = DQFDLossFunction(
supervised_weight=supervised_weight,
discount=self.discount, double_q=self.double_q, huber_loss=self.huber_loss,
shared_container_action_target=shared_container_action_target,
importance_weights=self.use_importance_weights, n_step=n_step
)
# Add all our sub-components to the core.
self.root_component.add_components(
self.preprocessor, self.merger, self.memory, self.demo_memory, self.splitter, self.policy,
self.target_policy, self.exploration, self.loss_function, self.optimizer
)
# Define the Agent's (root-Component's) API.
self.define_graph_api()
if self.auto_build:
self._build_graph([self.root_component], self.input_spaces, optimizer=self.optimizer,
batch_size=self.update_spec["batch_size"])
self.graph_built = True
def __init__(self,
double_q=True,
initial_alpha=1.0,
gumbel_softmax_temperature=1.0,
target_entropy=None,
memory_spec=None,
value_function_sync_spec=None,
**kwargs):
"""
This is an implementation of the Soft-Actor Critic algorithm.
Paper: http://arxiv.org/abs/1801.01290
Args:
double_q (bool): Whether to train two q networks independently.
initial_alpha (float): "The temperature parameter α determines the
relative importance of the entropy term against the reward".
gumbel_softmax_temperature (float): Temperature parameter for the Gumbel-Softmax distribution used
for discrete actions.
memory_spec (Optional[dict,Memory]): The spec for the Memory to use for the DQN algorithm.
update_spec (dict): Here we can have sync_interval or sync_tau (for the value network update).
"""
value_function_spec = kwargs.pop("value_function_spec")
value_function_spec = dict(type="sac_value_function",
network_spec=value_function_spec)
super(SACAgent, self).__init__(
# Continuous action space: Use squashed normal.
# Discrete: Gumbel-softmax.
policy_spec=dict(
deterministic=False,
distributions_spec=dict(
bounded_distribution_type="squashed",
discrete_distribution_type="gumbel_softmax",
gumbel_softmax_temperature=gumbel_softmax_temperature)),
name=kwargs.pop("name", "sac-agent"),
value_function_spec=value_function_spec,
**kwargs)
self.double_q = double_q
self.target_entropy = target_entropy
self.initial_alpha = initial_alpha
# Assert that the synch interval is a multiple of the update_interval.
if "sync_interval" in self.update_spec:
if self.update_spec["sync_interval"] / self.update_spec["update_interval"] != \
self.update_spec["sync_interval"] // self.update_spec["update_interval"]:
raise RLGraphError(
"ERROR: sync_interval ({}) must be multiple of update_interval "
"({})!".format(self.update_spec["sync_interval"],
self.update_spec["update_interval"]))
elif "sync_tau" in self.update_spec:
if self.update_spec["sync_tau"] <= 0 or self.update_spec[
"sync_tau"] > 1.0:
raise RLGraphError(
"sync_tau ({}) must be in interval (0.0, 1.0]!".format(
self.update_spec["sync_tau"]))
else:
self.update_spec[
"sync_tau"] = 0.005 # The value mentioned in the paper
# Extend input Space definitions to this Agent's specific API-methods.
preprocessed_state_space = self.preprocessed_state_space.with_batch_rank(
)
reward_space = FloatBox(add_batch_rank=True)
terminal_space = BoolBox(add_batch_rank=True)
self.iterations = self.update_spec["num_iterations"]
self.batch_size = self.update_spec["batch_size"]
float_action_space = self.action_space.with_batch_rank()
if isinstance(self.action_space, Dict):
for name, space in float_action_space.flatten(
scope_separator_at_start=False).items():
if isinstance(space, IntBox):
float_action_space[name] = space.as_one_hot_float_space()
elif isinstance(self.action_space, IntBox):
float_action_space = float_action_space.as_one_hot_float_space()
self.input_spaces.update(
dict(env_actions=self.action_space.with_batch_rank(),
actions=float_action_space,
preprocessed_states=preprocessed_state_space,
rewards=reward_space,
terminals=terminal_space,
next_states=preprocessed_state_space,
states=self.state_space.with_batch_rank(add_batch_rank=True),
batch_size=int,
importance_weights=FloatBox(add_batch_rank=True),
deterministic=bool,
weights="variables:{}".format(self.policy.scope)))
if value_function_sync_spec is None:
value_function_sync_spec = SyncSpecification(
sync_interval=self.update_spec["sync_interval"] //
self.update_spec["update_interval"],
sync_tau=self.update_spec["sync_tau"]
if "sync_tau" in self.update_spec else 5e-3)
self.memory = Memory.from_spec(memory_spec)
self.alpha_optimizer = self.optimizer.copy(
scope="alpha-" +
self.optimizer.scope) if self.target_entropy is not None else None
self.root_component = SACAgentComponent(
agent=self,
policy=self.policy,
q_function=self.value_function,
preprocessor=self.preprocessor,
memory=self.memory,
discount=self.discount,
initial_alpha=self.initial_alpha,
target_entropy=target_entropy,
optimizer=self.optimizer,
vf_optimizer=self.value_function_optimizer,
alpha_optimizer=self.alpha_optimizer,
q_sync_spec=value_function_sync_spec,
num_q_functions=2 if self.double_q is True else 1)
extra_optimizers = [self.value_function_optimizer]
if self.alpha_optimizer is not None:
extra_optimizers.append(self.alpha_optimizer)
self.build_options = dict(optimizers=extra_optimizers)
if self.auto_build:
self._build_graph([self.root_component],
self.input_spaces,
optimizer=self.optimizer,
batch_size=self.update_spec["batch_size"],
build_options=self.build_options)
self.graph_built = True
def __init__(self,
clip_ratio=0.2,
gae_lambda=1.0,
clip_rewards=0.0,
standardize_advantages=False,
sample_episodes=True,
weight_entropy=None,
memory_spec=None,
**kwargs):
"""
Args:
clip_ratio (float): Clipping parameter for likelihood ratio.
gae_lambda (float): Lambda for generalized advantage estimation.
clip_rewards (float): Reward clip value. If not 0, rewards will be clipped into this range.
standardize_advantages (bool): If true, standardize advantage values in update.
sample_episodes (bool): If True, the update method interprets the batch_size as the number of
episodes to fetch from the memory. If False, batch_size will refer to the number of time-steps. This
is especially relevant for environments where episode lengths may vastly differ throughout training. For
example, in CartPole, a losing episode is typically 10 steps, and a winning episode 200 steps.
weight_entropy (float): The coefficient used for the entropy regularization term (L[E]).
memory_spec (Optional[dict,Memory]): The spec for the Memory to use. Should typically be
a ring-buffer.
"""
if "policy_spec" in kwargs:
policy_spec = kwargs.pop("policy_spec")
policy_spec["deterministic"] = False
else:
policy_spec = dict(deterministic=False)
super(PPOAgent, self).__init__(
policy_spec=policy_spec, # Set policy to stochastic.
name=kwargs.pop("name", "ppo-agent"),
**kwargs)
self.sample_episodes = sample_episodes
# TODO: Have to manually set it here for multi-GPU synchronizer to know its number
# TODO: of return values when calling _graph_fn_calculate_update_from_external_batch.
# self.root_component.graph_fn_num_outputs["_graph_fn_update_from_external_batch"] = 4
# Extend input Space definitions to this Agent's specific API-methods.
preprocessed_state_space = self.preprocessed_state_space.with_batch_rank(
)
reward_space = FloatBox(add_batch_rank=True)
terminal_space = BoolBox(add_batch_rank=True)
self.input_spaces.update(
dict(actions=self.action_space.with_batch_rank(),
policy_weights="variables:policy",
value_function_weights="variables:value-function",
deterministic=bool,
preprocessed_states=preprocessed_state_space,
rewards=reward_space,
terminals=terminal_space,
sequence_indices=BoolBox(add_batch_rank=True),
apply_postprocessing=bool))
# The merger to merge inputs into one record Dict going into the memory.
self.merger = ContainerMerger("states", "actions", "rewards",
"terminals")
self.memory = Memory.from_spec(memory_spec)
assert isinstance(
self.memory, RingBuffer
), "ERROR: PPO memory must be ring-buffer for episode-handling!"
# Make sure the python buffer is not larger than our memory capacity.
assert self.observe_spec["buffer_size"] <= self.memory.capacity, \
"ERROR: Buffer's size ({}) in `observe_spec` must be smaller or equal to the memory's capacity ({})!". \
format(self.observe_spec["buffer_size"], self.memory.capacity)
# The splitter for splitting up the records coming from the memory.
self.splitter = ContainerSplitter("states", "actions", "rewards",
"terminals")
self.gae_function = GeneralizedAdvantageEstimation(
gae_lambda=gae_lambda,
discount=self.discount,
clip_rewards=clip_rewards)
self.loss_function = PPOLossFunction(
clip_ratio=clip_ratio,
standardize_advantages=standardize_advantages,
weight_entropy=weight_entropy)
self.iterations = self.update_spec["num_iterations"]
self.sample_size = self.update_spec["sample_size"]
self.batch_size = self.update_spec["batch_size"]
# Add all our sub-components to the core.
self.root_component.add_components(
self.preprocessor, self.merger, self.memory, self.splitter,
self.policy, self.exploration, self.loss_function, self.optimizer,
self.value_function, self.value_function_optimizer,
self.vars_merger, self.vars_splitter, self.gae_function)
# Define the Agent's (root-Component's) API.
self.define_graph_api()
self.build_options = dict(vf_optimizer=self.value_function_optimizer)
if self.auto_build:
self._build_graph(
[self.root_component],
self.input_spaces,
optimizer=self.optimizer,
# Important: Use sample-size, not batch-size as the sub-samples (from a batch) are the ones that get
# multi-gpu-split.
batch_size=self.update_spec["sample_size"],
build_options=self.build_options)
self.graph_built = True
def __init__(
self,
state_space,
action_space,
discount=0.98,
preprocessing_spec=None,
network_spec=None,
internal_states_space=None,
policy_spec=None,
exploration_spec=None,
execution_spec=None,
optimizer_spec=None,
observe_spec=None,
update_spec=None,
summary_spec=None,
saver_spec=None,
auto_build=True,
name="dqn-agent",
double_q=True,
dueling_q=True,
huber_loss=False,
n_step=1,
shared_container_action_target=True,
memory_spec=None,
):
"""
Args:
state_space (Union[dict,Space]): Spec dict for the state Space or a direct Space object.
action_space (Union[dict,Space]): Spec dict for the action Space or a direct Space object.
preprocessing_spec (Optional[list,PreprocessorStack]): The spec list for the different necessary states
preprocessing steps or a PreprocessorStack object itself.
discount (float): The discount factor (gamma).
network_spec (Optional[list,NeuralNetwork]): Spec list for a NeuralNetwork Component or the NeuralNetwork
object itself.
internal_states_space (Optional[Union[dict,Space]]): Spec dict for the internal-states Space or a direct
Space object for the Space(s) of the internal (RNN) states.
policy_spec (Optional[dict]): An optional dict for further kwargs passing into the Policy c'tor.
exploration_spec (Optional[dict]): The spec-dict to create the Exploration Component.
execution_spec (Optional[dict,Execution]): The spec-dict specifying execution settings.
optimizer_spec (Optional[dict,Optimizer]): The spec-dict to create the Optimizer for this Agent.
observe_spec (Optional[dict]): Spec-dict to specify `Agent.observe()` settings.
update_spec (Optional[dict]): Spec-dict to specify `Agent.update()` settings.
summary_spec (Optional[dict]): Spec-dict to specify summary settings.
saver_spec (Optional[dict]): Spec-dict to specify saver settings.
auto_build (Optional[bool]): If True (default), immediately builds the graph using the agent's
graph builder. If false, users must separately call agent.build(). Useful for debugging or analyzing
components before building.
name (str): Some name for this Agent object.
double_q (bool): Whether to use the double DQN loss function (see [2]).
dueling_q (bool): Whether to use a dueling layer in the ActionAdapter (see [3]).
huber_loss (bool) : Whether to apply a Huber loss. (see [4]).
n_step (Optional[int]): n-step adjustment to discounting.
memory_spec (Optional[dict,Memory]): The spec for the Memory to use for the DQN algorithm.
"""
# Fix action-adapter before passing it to the super constructor.
# Use a DuelingPolicy (instead of a basic Policy) if option is set.
if dueling_q is True:
policy_spec["type"] = "dueling-policy"
# Give us some default state-value nodes.
if "units_state_value_stream" not in policy_spec:
policy_spec["units_state_value_stream"] = 128
super(DQNAgent,
self).__init__(state_space=state_space,
action_space=action_space,
discount=discount,
preprocessing_spec=preprocessing_spec,
network_spec=network_spec,
internal_states_space=internal_states_space,
policy_spec=policy_spec,
exploration_spec=exploration_spec,
execution_spec=execution_spec,
optimizer_spec=optimizer_spec,
observe_spec=observe_spec,
update_spec=update_spec,
summary_spec=summary_spec,
saver_spec=saver_spec,
auto_build=auto_build,
name=name)
# TODO: Have to manually set it here for multi-GPU synchronizer to know its number
# TODO: of return values when calling _graph_fn_calculate_update_from_external_batch.
# self.root_component.graph_fn_num_outputs["_graph_fn_update_from_external_batch"] = 4
# Assert that the synch interval is a multiple of the update_interval.
if self.update_spec["sync_interval"] / self.update_spec["update_interval"] != \
self.update_spec["sync_interval"] // self.update_spec["update_interval"]:
raise RLGraphError(
"ERROR: sync_interval ({}) must be multiple of update_interval "
"({})!".format(self.update_spec["sync_interval"],
self.update_spec["update_interval"]))
self.double_q = double_q
self.dueling_q = dueling_q
self.huber_loss = huber_loss
self.shared_container_action_target = shared_container_action_target
# Extend input Space definitions to this Agent's specific API-methods.
preprocessed_state_space = self.preprocessed_state_space.with_batch_rank(
)
reward_space = FloatBox(add_batch_rank=True)
terminal_space = BoolBox(add_batch_rank=True)
weight_space = FloatBox(add_batch_rank=True)
self.input_spaces.update(
dict(
actions=self.action_space.with_batch_rank(),
# Weights will have a Space derived from the vars of policy.
policy_weights="variables:{}".format(self.policy.scope),
use_exploration=bool,
preprocessed_states=preprocessed_state_space,
rewards=reward_space,
terminals=terminal_space,
next_states=preprocessed_state_space,
preprocessed_next_states=preprocessed_state_space,
importance_weights=weight_space,
apply_postprocessing=bool))
if self.value_function is not None:
self.input_spaces[
"value_function_weights"] = "variables:{}".format(
self.value_function.scope),
# The replay memory.
self.memory = Memory.from_spec(memory_spec)
# The splitter for splitting up the records coming from the memory.
self.splitter = ContainerSplitter("states", "actions", "rewards",
"terminals", "next_states")
# Make sure the python buffer is not larger than our memory capacity.
assert self.observe_spec["buffer_size"] <= self.memory.capacity,\
"ERROR: Buffer's size ({}) in `observe_spec` must be smaller or equal to the memory's capacity ({})!".\
format(self.observe_spec["buffer_size"], self.memory.capacity)
# Copy our Policy (target-net), make target-net synchronizable.
self.target_policy = self.policy.copy(scope="target-policy",
trainable=False)
# Number of steps since the last target-net synching from the main policy.
self.steps_since_target_net_sync = 0
use_importance_weights = isinstance(self.memory, PrioritizedReplay)
self.loss_function = DQNLossFunction(
discount=self.discount,
double_q=self.double_q,
huber_loss=self.huber_loss,
shared_container_action_target=shared_container_action_target,
importance_weights=use_importance_weights,
n_step=n_step)
self.root_component.add_components(
self.preprocessor,
self.memory,
self.splitter,
self.policy,
self.target_policy,
self.value_function,
self.value_function_optimizer, # <- should both be None for DQN
self.exploration,
self.loss_function,
self.optimizer,
self.vars_merger,
self.vars_splitter)
# Define the Agent's (root-Component's) API.
self.define_graph_api()
# markup = get_graph_markup(self.graph_builder.root_component)
# print(markup)
if self.auto_build:
self._build_graph([self.root_component],
self.input_spaces,
optimizer=self.optimizer,
batch_size=self.update_spec["batch_size"])
self.graph_built = True