from typing import Any, Dict, Optional, Sequence
from ..argument_utility import (
ActionScalerArg,
EncoderArg,
QFuncArg,
RewardScalerArg,
ScalerArg,
UseGPUArg,
check_encoder,
check_q_func,
check_use_gpu,
)
from ..constants import IMPL_NOT_INITIALIZED_ERROR, ActionSpace
from ..dataset import TransitionMiniBatch
from ..gpu import Device
from ..models.encoders import EncoderFactory
from ..models.optimizers import AdamFactory, OptimizerFactory
from ..models.q_functions import QFunctionFactory
from .base import AlgoBase
from .torch.sac_impl import DiscreteSACImpl, SACImpl
[docs]class SAC(AlgoBase):
r"""Soft Actor-Critic algorithm.
SAC is a DDPG-based maximum entropy RL algorithm, which produces
state-of-the-art performance in online RL settings.
SAC leverages twin Q functions proposed in TD3. Additionally,
`delayed policy update` in TD3 is also implemented, which is not done in
the paper.
.. math::
L(\theta_i) = \mathbb{E}_{s_t,\, a_t,\, r_{t+1},\, s_{t+1} \sim D,\,
a_{t+1} \sim \pi_\phi(\cdot|s_{t+1})} \Big[
\big(y - Q_{\theta_i}(s_t, a_t)\big)^2\Big]
.. math::
y = r_{t+1} + \gamma \Big(\min_j Q_{\theta_j}(s_{t+1}, a_{t+1})
- \alpha \log \big(\pi_\phi(a_{t+1}|s_{t+1})\big)\Big)
.. math::
J(\phi) = \mathbb{E}_{s_t \sim D,\, a_t \sim \pi_\phi(\cdot|s_t)}
\Big[\alpha \log (\pi_\phi (a_t|s_t))
- \min_i Q_{\theta_i}\big(s_t, \pi_\phi(a_t|s_t)\big)\Big]
The temperature parameter :math:`\alpha` is also automatically adjustable.
.. math::
J(\alpha) = \mathbb{E}_{s_t \sim D,\, a_t \sim \pi_\phi(\cdot|s_t)}
\bigg[-\alpha \Big(\log \big(\pi_\phi(a_t|s_t)\big) + H\Big)\bigg]
where :math:`H` is a target
entropy, which is defined as :math:`\dim a`.
References:
* `Haarnoja et al., Soft Actor-Critic: Off-Policy Maximum Entropy Deep
Reinforcement Learning with a Stochastic Actor.
<https://arxiv.org/abs/1801.01290>`_
* `Haarnoja et al., Soft Actor-Critic Algorithms and Applications.
<https://arxiv.org/abs/1812.05905>`_
Args:
actor_learning_rate (float): learning rate for policy function.
critic_learning_rate (float): learning rate for Q functions.
temp_learning_rate (float): learning rate for temperature parameter.
actor_optim_factory (d3rlpy.models.optimizers.OptimizerFactory):
optimizer factory for the actor.
critic_optim_factory (d3rlpy.models.optimizers.OptimizerFactory):
optimizer factory for the critic.
temp_optim_factory (d3rlpy.models.optimizers.OptimizerFactory):
optimizer factory for the temperature.
actor_encoder_factory (d3rlpy.models.encoders.EncoderFactory or str):
encoder factory for the actor.
critic_encoder_factory (d3rlpy.models.encoders.EncoderFactory or str):
encoder factory for the critic.
q_func_factory (d3rlpy.models.q_functions.QFunctionFactory or str):
Q function factory.
batch_size (int): mini-batch size.
n_frames (int): the number of frames to stack for image observation.
n_steps (int): N-step TD calculation.
gamma (float): discount factor.
tau (float): target network synchronization coefficiency.
n_critics (int): the number of Q functions for ensemble.
target_reduction_type (str): ensemble reduction method at target value
estimation. The available options are
``['min', 'max', 'mean', 'mix', 'none']``.
initial_temperature (float): initial temperature value.
use_gpu (bool, int or d3rlpy.gpu.Device):
flag to use GPU, device ID or device.
scaler (d3rlpy.preprocessing.Scaler or str): preprocessor.
The available options are `['pixel', 'min_max', 'standard']`.
action_scaler (d3rlpy.preprocessing.ActionScaler or str):
action preprocessor. The available options are ``['min_max']``.
reward_scaler (d3rlpy.preprocessing.RewardScaler or str):
reward preprocessor. The available options are
``['clip', 'min_max', 'standard']``.
impl (d3rlpy.algos.torch.sac_impl.SACImpl): algorithm implementation.
"""
_actor_learning_rate: float
_critic_learning_rate: float
_temp_learning_rate: float
_actor_optim_factory: OptimizerFactory
_critic_optim_factory: OptimizerFactory
_temp_optim_factory: OptimizerFactory
_actor_encoder_factory: EncoderFactory
_critic_encoder_factory: EncoderFactory
_q_func_factory: QFunctionFactory
_tau: float
_n_critics: int
_target_reduction_type: str
_initial_temperature: float
_use_gpu: Optional[Device]
_impl: Optional[SACImpl]
def __init__(
self,
*,
actor_learning_rate: float = 3e-4,
critic_learning_rate: float = 3e-4,
temp_learning_rate: float = 3e-4,
actor_optim_factory: OptimizerFactory = AdamFactory(),
critic_optim_factory: OptimizerFactory = AdamFactory(),
temp_optim_factory: OptimizerFactory = AdamFactory(),
actor_encoder_factory: EncoderArg = "default",
critic_encoder_factory: EncoderArg = "default",
q_func_factory: QFuncArg = "mean",
batch_size: int = 100,
n_frames: int = 1,
n_steps: int = 1,
gamma: float = 0.99,
tau: float = 0.005,
n_critics: int = 2,
target_reduction_type: str = "min",
initial_temperature: float = 1.0,
use_gpu: UseGPUArg = False,
scaler: ScalerArg = None,
action_scaler: ActionScalerArg = None,
reward_scaler: RewardScalerArg = None,
impl: Optional[SACImpl] = None,
**kwargs: Any
):
super().__init__(
batch_size=batch_size,
n_frames=n_frames,
n_steps=n_steps,
gamma=gamma,
scaler=scaler,
action_scaler=action_scaler,
reward_scaler=reward_scaler,
kwargs=kwargs,
)
self._actor_learning_rate = actor_learning_rate
self._critic_learning_rate = critic_learning_rate
self._temp_learning_rate = temp_learning_rate
self._actor_optim_factory = actor_optim_factory
self._critic_optim_factory = critic_optim_factory
self._temp_optim_factory = temp_optim_factory
self._actor_encoder_factory = check_encoder(actor_encoder_factory)
self._critic_encoder_factory = check_encoder(critic_encoder_factory)
self._q_func_factory = check_q_func(q_func_factory)
self._tau = tau
self._n_critics = n_critics
self._target_reduction_type = target_reduction_type
self._initial_temperature = initial_temperature
self._use_gpu = check_use_gpu(use_gpu)
self._impl = impl
def _create_impl(
self, observation_shape: Sequence[int], action_size: int
) -> None:
self._impl = SACImpl(
observation_shape=observation_shape,
action_size=action_size,
actor_learning_rate=self._actor_learning_rate,
critic_learning_rate=self._critic_learning_rate,
temp_learning_rate=self._temp_learning_rate,
actor_optim_factory=self._actor_optim_factory,
critic_optim_factory=self._critic_optim_factory,
temp_optim_factory=self._temp_optim_factory,
actor_encoder_factory=self._actor_encoder_factory,
critic_encoder_factory=self._critic_encoder_factory,
q_func_factory=self._q_func_factory,
gamma=self._gamma,
tau=self._tau,
n_critics=self._n_critics,
target_reduction_type=self._target_reduction_type,
initial_temperature=self._initial_temperature,
use_gpu=self._use_gpu,
scaler=self._scaler,
action_scaler=self._action_scaler,
reward_scaler=self._reward_scaler,
)
self._impl.build()
def _update(self, batch: TransitionMiniBatch) -> Dict[str, float]:
assert self._impl is not None, IMPL_NOT_INITIALIZED_ERROR
metrics = {}
# lagrangian parameter update for SAC temperature
if self._temp_learning_rate > 0:
temp_loss, temp = self._impl.update_temp(batch)
metrics.update({"temp_loss": temp_loss, "temp": temp})
critic_loss = self._impl.update_critic(batch)
metrics.update({"critic_loss": critic_loss})
actor_loss = self._impl.update_actor(batch)
metrics.update({"actor_loss": actor_loss})
self._impl.update_critic_target()
self._impl.update_actor_target()
return metrics
[docs] def get_action_type(self) -> ActionSpace:
return ActionSpace.CONTINUOUS
[docs]class DiscreteSAC(AlgoBase):
r"""Soft Actor-Critic algorithm for discrete action-space.
This discrete version of SAC is built based on continuous version of SAC
with additional modifications.
The target state-value is calculated as expectation of all action-values.
.. math::
V(s_t) = \pi_\phi (s_t)^T [Q_\theta(s_t) - \alpha \log (\pi_\phi (s_t))]
Similarly, the objective function for the temperature parameter is as
follows.
.. math::
J(\alpha) = \pi_\phi (s_t)^T [-\alpha (\log(\pi_\phi (s_t)) + H)]
Finally, the objective function for the policy function is as follows.
.. math::
J(\phi) = \mathbb{E}_{s_t \sim D}
[\pi_\phi(s_t)^T [\alpha \log(\pi_\phi(s_t)) - Q_\theta(s_t)]]
References:
* `Christodoulou, Soft Actor-Critic for Discrete Action Settings.
<https://arxiv.org/abs/1910.07207>`_
Args:
actor_learning_rate (float): learning rate for policy function.
critic_learning_rate (float): learning rate for Q functions.
temp_learning_rate (float): learning rate for temperature parameter.
actor_optim_factory (d3rlpy.models.optimizers.OptimizerFactory):
optimizer factory for the actor.
critic_optim_factory (d3rlpy.models.optimizers.OptimizerFactory):
optimizer factory for the critic.
temp_optim_factory (d3rlpy.models.optimizers.OptimizerFactory):
optimizer factory for the temperature.
actor_encoder_factory (d3rlpy.models.encoders.EncoderFactory or str):
encoder factory for the actor.
critic_encoder_factory (d3rlpy.models.encoders.EncoderFactory or str):
encoder factory for the critic.
q_func_factory (d3rlpy.models.q_functions.QFunctionFactory or str):
Q function factory.
batch_size (int): mini-batch size.
n_frames (int): the number of frames to stack for image observation.
n_steps (int): N-step TD calculation.
gamma (float): discount factor.
n_critics (int): the number of Q functions for ensemble.
initial_temperature (float): initial temperature value.
use_gpu (bool, int or d3rlpy.gpu.Device):
flag to use GPU, device ID or device.
scaler (d3rlpy.preprocessing.Scaler or str): preprocessor.
The available options are `['pixel', 'min_max', 'standard']`
reward_scaler (d3rlpy.preprocessing.RewardScaler or str):
reward preprocessor. The available options are
``['clip', 'min_max', 'standard']``.
impl (d3rlpy.algos.torch.sac_impl.DiscreteSACImpl):
algorithm implementation.
"""
_actor_learning_rate: float
_critic_learning_rate: float
_temp_learning_rate: float
_actor_optim_factory: OptimizerFactory
_critic_optim_factory: OptimizerFactory
_temp_optim_factory: OptimizerFactory
_actor_encoder_factory: EncoderFactory
_critic_encoder_factory: EncoderFactory
_q_func_factory: QFunctionFactory
_n_critics: int
_initial_temperature: float
_target_update_interval: int
_use_gpu: Optional[Device]
_impl: Optional[DiscreteSACImpl]
def __init__(
self,
*,
actor_learning_rate: float = 3e-4,
critic_learning_rate: float = 3e-4,
temp_learning_rate: float = 3e-4,
actor_optim_factory: OptimizerFactory = AdamFactory(eps=1e-4),
critic_optim_factory: OptimizerFactory = AdamFactory(eps=1e-4),
temp_optim_factory: OptimizerFactory = AdamFactory(eps=1e-4),
actor_encoder_factory: EncoderArg = "default",
critic_encoder_factory: EncoderArg = "default",
q_func_factory: QFuncArg = "mean",
batch_size: int = 64,
n_frames: int = 1,
n_steps: int = 1,
gamma: float = 0.99,
n_critics: int = 2,
initial_temperature: float = 1.0,
target_update_interval: int = 8000,
use_gpu: UseGPUArg = False,
scaler: ScalerArg = None,
reward_scaler: RewardScalerArg = None,
impl: Optional[DiscreteSACImpl] = None,
**kwargs: Any
):
super().__init__(
batch_size=batch_size,
n_frames=n_frames,
n_steps=n_steps,
gamma=gamma,
scaler=scaler,
action_scaler=None,
reward_scaler=reward_scaler,
kwargs=kwargs,
)
self._actor_learning_rate = actor_learning_rate
self._critic_learning_rate = critic_learning_rate
self._temp_learning_rate = temp_learning_rate
self._actor_optim_factory = actor_optim_factory
self._critic_optim_factory = critic_optim_factory
self._temp_optim_factory = temp_optim_factory
self._actor_encoder_factory = check_encoder(actor_encoder_factory)
self._critic_encoder_factory = check_encoder(critic_encoder_factory)
self._q_func_factory = check_q_func(q_func_factory)
self._n_critics = n_critics
self._initial_temperature = initial_temperature
self._target_update_interval = target_update_interval
self._use_gpu = check_use_gpu(use_gpu)
self._impl = impl
def _create_impl(
self, observation_shape: Sequence[int], action_size: int
) -> None:
self._impl = DiscreteSACImpl(
observation_shape=observation_shape,
action_size=action_size,
actor_learning_rate=self._actor_learning_rate,
critic_learning_rate=self._critic_learning_rate,
temp_learning_rate=self._temp_learning_rate,
actor_optim_factory=self._actor_optim_factory,
critic_optim_factory=self._critic_optim_factory,
temp_optim_factory=self._temp_optim_factory,
actor_encoder_factory=self._actor_encoder_factory,
critic_encoder_factory=self._critic_encoder_factory,
q_func_factory=self._q_func_factory,
gamma=self._gamma,
n_critics=self._n_critics,
initial_temperature=self._initial_temperature,
use_gpu=self._use_gpu,
scaler=self._scaler,
reward_scaler=self._reward_scaler,
)
self._impl.build()
def _update(self, batch: TransitionMiniBatch) -> Dict[str, float]:
assert self._impl is not None, IMPL_NOT_INITIALIZED_ERROR
metrics = {}
# lagrangian parameter update for SAC temeprature
if self._temp_learning_rate > 0:
temp_loss, temp = self._impl.update_temp(batch)
metrics.update({"temp_loss": temp_loss, "temp": temp})
critic_loss = self._impl.update_critic(batch)
metrics.update({"critic_loss": critic_loss})
actor_loss = self._impl.update_actor(batch)
metrics.update({"actor_loss": actor_loss})
if self._grad_step % self._target_update_interval == 0:
self._impl.update_target()
return metrics
[docs] def get_action_type(self) -> ActionSpace:
return ActionSpace.DISCRETE