import dataclasses
from ...base import DeviceArg, LearnableConfig, register_learnable
from ...constants import ActionSpace
from ...models.builders import (
create_continuous_q_function,
create_normal_policy,
)
from ...models.encoders import EncoderFactory, make_encoder_field
from ...models.optimizers import OptimizerFactory, make_optimizer_field
from ...models.q_functions import QFunctionFactory, make_q_func_field
from ...types import Shape
from .base import QLearningAlgoBase
from .torch.crr_impl import CRRImpl, CRRModules
__all__ = ["CRRConfig", "CRR"]
[docs]@dataclasses.dataclass()
class CRRConfig(LearnableConfig):
r"""Config of Critic Reguralized Regression algorithm.
CRR is a simple offline RL method similar to AWAC.
The policy is trained as a supervised regression.
.. math::
J(\phi) = \mathbb{E}_{s_t, a_t \sim D}
[\log \pi_\phi(a_t|s_t) f(Q_\theta, \pi_\phi, s_t, a_t)]
where :math:`f` is a filter function which has several options. The first
option is ``binary`` function.
.. math::
f := \mathbb{1} [A_\theta(s, a) > 0]
The other is ``exp`` function.
.. math::
f := \exp(A(s, a) / \beta)
The :math:`A(s, a)` is an average function which also has several options.
The first option is ``mean``.
.. math::
A(s, a) = Q_\theta (s, a) - \frac{1}{m} \sum^m_j Q(s, a_j)
The other one is ``max``.
.. math::
A(s, a) = Q_\theta (s, a) - \max^m_j Q(s, a_j)
where :math:`a_j \sim \pi_\phi(s)`.
In evaluation, the action is determined by Critic Weighted Policy (CWP).
In CWP, the several actions are sampled from the policy function, and the
final action is re-sampled from the estimated action-value distribution.
References:
* `Wang et al., Critic Reguralized Regression.
<https://arxiv.org/abs/2006.15134>`_
Args:
observation_scaler (d3rlpy.preprocessing.ObservationScaler):
Observation preprocessor.
action_scaler (d3rlpy.preprocessing.ActionScaler): Action preprocessor.
reward_scaler (d3rlpy.preprocessing.RewardScaler): Reward preprocessor.
actor_learning_rate (float): Learning rate for policy function.
critic_learning_rate (float): Learning rate for Q functions.
actor_optim_factory (d3rlpy.models.optimizers.OptimizerFactory):
Optimizer factory for the actor.
critic_optim_factory (d3rlpy.models.optimizers.OptimizerFactory):
Optimizer factory for the critic.
actor_encoder_factory (d3rlpy.models.encoders.EncoderFactory):
Encoder factory for the actor.
critic_encoder_factory (d3rlpy.models.encoders.EncoderFactory):
Encoder factory for the critic.
q_func_factory (d3rlpy.models.q_functions.QFunctionFactory):
Q function factory.
batch_size (int): Mini-batch size.
gamma (float): Discount factor.
beta (float): Temperature value defined as :math:`\beta` above.
n_action_samples (int): Number of sampled actions to calculate
:math:`A(s, a)` and for CWP.
advantage_type (str): Advantage function type. The available options
are ``['mean', 'max']``.
weight_type (str): Filter function type. The available options
are ``['binary', 'exp']``.
max_weight (float): Maximum weight for cross-entropy loss.
n_critics (int): Number of Q functions for ensemble.
target_update_type (str): Target update type. The available options are
``['hard', 'soft']``.
tau (float): Target network synchronization coefficiency used with
``soft`` target update.
update_actor_interval (int): Interval to update policy function used
with ``hard`` target update.
"""
actor_learning_rate: float = 3e-4
critic_learning_rate: float = 3e-4
actor_optim_factory: OptimizerFactory = make_optimizer_field()
critic_optim_factory: OptimizerFactory = make_optimizer_field()
actor_encoder_factory: EncoderFactory = make_encoder_field()
critic_encoder_factory: EncoderFactory = make_encoder_field()
q_func_factory: QFunctionFactory = make_q_func_field()
batch_size: int = 100
gamma: float = 0.99
beta: float = 1.0
n_action_samples: int = 4
advantage_type: str = "mean"
weight_type: str = "exp"
max_weight: float = 20.0
n_critics: int = 1
target_update_type: str = "hard"
tau: float = 5e-3
target_update_interval: int = 100
update_actor_interval: int = 1
[docs] def create(self, device: DeviceArg = False) -> "CRR":
return CRR(self, device)
@staticmethod
def get_type() -> str:
return "crr"
[docs]class CRR(QLearningAlgoBase[CRRImpl, CRRConfig]):
def inner_create_impl(
self, observation_shape: Shape, action_size: int
) -> None:
policy = create_normal_policy(
observation_shape,
action_size,
self._config.actor_encoder_factory,
device=self._device,
)
targ_policy = create_normal_policy(
observation_shape,
action_size,
self._config.actor_encoder_factory,
device=self._device,
)
q_funcs, q_func_forwarder = create_continuous_q_function(
observation_shape,
action_size,
self._config.critic_encoder_factory,
self._config.q_func_factory,
n_ensembles=self._config.n_critics,
device=self._device,
)
targ_q_funcs, targ_q_func_forwarder = create_continuous_q_function(
observation_shape,
action_size,
self._config.critic_encoder_factory,
self._config.q_func_factory,
n_ensembles=self._config.n_critics,
device=self._device,
)
actor_optim = self._config.actor_optim_factory.create(
policy.named_modules(), lr=self._config.actor_learning_rate
)
critic_optim = self._config.critic_optim_factory.create(
q_funcs.named_modules(), lr=self._config.critic_learning_rate
)
modules = CRRModules(
policy=policy,
targ_policy=targ_policy,
q_funcs=q_funcs,
targ_q_funcs=targ_q_funcs,
actor_optim=actor_optim,
critic_optim=critic_optim,
)
self._impl = CRRImpl(
observation_shape=observation_shape,
action_size=action_size,
modules=modules,
q_func_forwarder=q_func_forwarder,
targ_q_func_forwarder=targ_q_func_forwarder,
gamma=self._config.gamma,
beta=self._config.beta,
n_action_samples=self._config.n_action_samples,
advantage_type=self._config.advantage_type,
weight_type=self._config.weight_type,
max_weight=self._config.max_weight,
tau=self._config.tau,
target_update_type=self._config.target_update_type,
target_update_interval=self._config.target_update_interval,
device=self._device,
)
[docs] def get_action_type(self) -> ActionSpace:
return ActionSpace.CONTINUOUS
register_learnable(CRRConfig)