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.plas_impl import PLASImpl, PLASWithPerturbationImpl
[docs]class PLAS(AlgoBase):
r"""Policy in Latent Action Space algorithm.
PLAS is an offline deep reinforcement learning algorithm whose policy
function is trained in latent space of Conditional VAE.
Unlike other algorithms, PLAS can achieve good performance by using
its less constrained policy function.
.. math::
a \sim p_\beta (a|s, z=\pi_\phi(s))
where :math:`\beta` is a parameter of the decoder in Conditional VAE.
References:
* `Zhou et al., PLAS: latent action space for offline reinforcement
learning. <https://arxiv.org/abs/2011.07213>`_
Args:
actor_learning_rate (float): learning rate for policy function.
critic_learning_rate (float): learning rate for Q functions.
imitator_learning_rate (float): learning rate for Conditional VAE.
actor_optim_factory (d3rlpy.models.optimizers.OptimizerFactory):
optimizer factory for the actor.
critic_optim_factory (d3rlpy.models.optimizers.OptimizerFactory):
optimizer factory for the critic.
imitator_optim_factory (d3rlpy.models.optimizers.OptimizerFactory):
optimizer factory for the conditional VAE.
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.
imitator_encoder_factory (d3rlpy.models.encoders.EncoderFactory or str):
encoder factory for the conditional VAE.
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']``.
update_actor_interval (int): interval to update policy function.
lam (float): weight factor for critic ensemble.
warmup_steps (int): the number of steps to warmup the VAE.
beta (float): KL reguralization term for Conditional VAE.
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.bcq_impl.BCQImpl): algorithm implementation.
"""
_actor_learning_rate: float
_critic_learning_rate: float
_imitator_learning_rate: float
_actor_optim_factory: OptimizerFactory
_critic_optim_factory: OptimizerFactory
_imitator_optim_factory: OptimizerFactory
_actor_encoder_factory: EncoderFactory
_critic_encoder_factory: EncoderFactory
_imitator_encoder_factory: EncoderFactory
_q_func_factory: QFunctionFactory
_tau: float
_n_critics: int
_target_reduction_type: str
_update_actor_interval: int
_lam: float
_warmup_steps: int
_beta: float
_use_gpu: Optional[Device]
_impl: Optional[PLASImpl]
def __init__(
self,
*,
actor_learning_rate: float = 1e-4,
critic_learning_rate: float = 1e-3,
imitator_learning_rate: float = 1e-4,
actor_optim_factory: OptimizerFactory = AdamFactory(),
critic_optim_factory: OptimizerFactory = AdamFactory(),
imitator_optim_factory: OptimizerFactory = AdamFactory(),
actor_encoder_factory: EncoderArg = "default",
critic_encoder_factory: EncoderArg = "default",
imitator_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 = "mix",
update_actor_interval: int = 1,
lam: float = 0.75,
warmup_steps: int = 500000,
beta: float = 0.5,
use_gpu: UseGPUArg = False,
scaler: ScalerArg = None,
action_scaler: ActionScalerArg = None,
reward_scaler: RewardScalerArg = None,
impl: Optional[PLASImpl] = 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._imitator_learning_rate = imitator_learning_rate
self._actor_optim_factory = actor_optim_factory
self._critic_optim_factory = critic_optim_factory
self._imitator_optim_factory = imitator_optim_factory
self._actor_encoder_factory = check_encoder(actor_encoder_factory)
self._critic_encoder_factory = check_encoder(critic_encoder_factory)
self._imitator_encoder_factory = check_encoder(imitator_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._update_actor_interval = update_actor_interval
self._lam = lam
self._warmup_steps = warmup_steps
self._beta = beta
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 = PLASImpl(
observation_shape=observation_shape,
action_size=action_size,
actor_learning_rate=self._actor_learning_rate,
critic_learning_rate=self._critic_learning_rate,
imitator_learning_rate=self._imitator_learning_rate,
actor_optim_factory=self._actor_optim_factory,
critic_optim_factory=self._critic_optim_factory,
imitator_optim_factory=self._imitator_optim_factory,
actor_encoder_factory=self._actor_encoder_factory,
critic_encoder_factory=self._critic_encoder_factory,
imitator_encoder_factory=self._imitator_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,
lam=self._lam,
beta=self._beta,
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 = {}
if self._grad_step < self._warmup_steps:
imitator_loss = self._impl.update_imitator(batch)
metrics.update({"imitator_loss": imitator_loss})
else:
critic_loss = self._impl.update_critic(batch)
metrics.update({"critic_loss": critic_loss})
if self._grad_step % self._update_actor_interval == 0:
actor_loss = self._impl.update_actor(batch)
metrics.update({"actor_loss": actor_loss})
self._impl.update_actor_target()
self._impl.update_critic_target()
return metrics
[docs] def get_action_type(self) -> ActionSpace:
return ActionSpace.CONTINUOUS
[docs]class PLASWithPerturbation(PLAS):
r"""Policy in Latent Action Space algorithm with perturbation layer.
PLAS with perturbation layer enables PLAS to output out-of-distribution
action.
References:
* `Zhou et al., PLAS: latent action space for offline reinforcement
learning. <https://arxiv.org/abs/2011.07213>`_
Args:
actor_learning_rate (float): learning rate for policy function.
critic_learning_rate (float): learning rate for Q functions.
imitator_learning_rate (float): learning rate for Conditional VAE.
actor_optim_factory (d3rlpy.models.optimizers.OptimizerFactory):
optimizer factory for the actor.
critic_optim_factory (d3rlpy.models.optimizers.OptimizerFactory):
optimizer factory for the critic.
imitator_optim_factory (d3rlpy.models.optimizers.OptimizerFactory):
optimizer factory for the conditional VAE.
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.
imitator_encoder_factory (d3rlpy.models.encoders.EncoderFactory or str):
encoder factory for the conditional VAE.
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']``.
update_actor_interval (int): interval to update policy function.
lam (float): weight factor for critic ensemble.
action_flexibility (float): output scale of perturbation layer.
warmup_steps (int): the number of steps to warmup the VAE.
beta (float): KL reguralization term for Conditional VAE.
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.bcq_impl.BCQImpl): algorithm implementation.
"""
_action_flexibility: float
_impl: Optional[PLASWithPerturbationImpl]
def __init__(
self,
*,
actor_learning_rate: float = 1e-4,
critic_learning_rate: float = 1e-3,
imitator_learning_rate: float = 1e-4,
actor_optim_factory: OptimizerFactory = AdamFactory(),
critic_optim_factory: OptimizerFactory = AdamFactory(),
imitator_optim_factory: OptimizerFactory = AdamFactory(),
actor_encoder_factory: EncoderArg = "default",
critic_encoder_factory: EncoderArg = "default",
imitator_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 = "mix",
update_actor_interval: int = 1,
lam: float = 0.75,
action_flexibility: float = 0.05,
warmup_steps: int = 500000,
beta: float = 0.5,
use_gpu: UseGPUArg = False,
scaler: ScalerArg = None,
action_scaler: ActionScalerArg = None,
reward_scaler: RewardScalerArg = None,
impl: Optional[PLASWithPerturbationImpl] = None,
**kwargs: Any
):
super().__init__(
actor_learning_rate=actor_learning_rate,
critic_learning_rate=critic_learning_rate,
imitator_learning_rate=imitator_learning_rate,
actor_optim_factory=actor_optim_factory,
critic_optim_factory=critic_optim_factory,
imitator_optim_factory=imitator_optim_factory,
actor_encoder_factory=actor_encoder_factory,
critic_encoder_factory=critic_encoder_factory,
imitator_encoder_factory=imitator_encoder_factory,
q_func_factory=q_func_factory,
batch_size=batch_size,
n_frames=n_frames,
n_steps=n_steps,
gamma=gamma,
tau=tau,
n_critics=n_critics,
target_reduction_type=target_reduction_type,
update_actor_interval=update_actor_interval,
lam=lam,
warmup_steps=warmup_steps,
beta=beta,
use_gpu=use_gpu,
scaler=scaler,
action_scaler=action_scaler,
reward_scaler=reward_scaler,
impl=impl,
**kwargs,
)
self._action_flexibility = action_flexibility
def _create_impl(
self, observation_shape: Sequence[int], action_size: int
) -> None:
self._impl = PLASWithPerturbationImpl(
observation_shape=observation_shape,
action_size=action_size,
actor_learning_rate=self._actor_learning_rate,
critic_learning_rate=self._critic_learning_rate,
imitator_learning_rate=self._imitator_learning_rate,
actor_optim_factory=self._actor_optim_factory,
critic_optim_factory=self._critic_optim_factory,
imitator_optim_factory=self._imitator_optim_factory,
actor_encoder_factory=self._actor_encoder_factory,
critic_encoder_factory=self._critic_encoder_factory,
imitator_encoder_factory=self._imitator_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,
lam=self._lam,
beta=self._beta,
action_flexibility=self._action_flexibility,
use_gpu=self._use_gpu,
scaler=self._scaler,
action_scaler=self._action_scaler,
reward_scaler=self._reward_scaler,
)
self._impl.build()