d3rlpy/algos/qlearning/dqn.py
import dataclasses
from ...base import DeviceArg, LearnableConfig, register_learnable
from ...constants import ActionSpace
from ...models.builders import create_discrete_q_function
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.dqn_impl import DoubleDQNImpl, DQNImpl, DQNModules
__all__ = ["DQNConfig", "DQN", "DoubleDQNConfig", "DoubleDQN"]
@dataclasses.dataclass()
class DQNConfig(LearnableConfig):
r"""Config of Deep Q-Network algorithm.
.. math::
L(\theta) = \mathbb{E}_{s_t, a_t, r_{t+1}, s_{t+1} \sim D} [(r_{t+1}
+ \gamma \max_a Q_{\theta'}(s_{t+1}, a) - Q_\theta(s_t, a_t))^2]
where :math:`\theta'` is the target network parameter. The target network
parameter is synchronized every `target_update_interval` iterations.
References:
* `Mnih et al., Human-level control through deep reinforcement
learning. <https://www.nature.com/articles/nature14236>`_
Args:
observation_scaler (d3rlpy.preprocessing.ObservationScaler):
Observation preprocessor.
reward_scaler (d3rlpy.preprocessing.RewardScaler): Reward preprocessor.
learning_rate (float): Learning rate.
optim_factory (d3rlpy.models.optimizers.OptimizerFactory):
Optimizer factory.
encoder_factory (d3rlpy.models.encoders.EncoderFactory):
Encoder factory.
q_func_factory (d3rlpy.models.q_functions.QFunctionFactory):
Q function factory.
batch_size (int): Mini-batch size.
gamma (float): Discount factor.
n_critics (int): Number of Q functions for ensemble.
target_update_interval (int): Interval to update the target network.
"""
batch_size: int = 32
learning_rate: float = 6.25e-5
optim_factory: OptimizerFactory = make_optimizer_field()
encoder_factory: EncoderFactory = make_encoder_field()
q_func_factory: QFunctionFactory = make_q_func_field()
gamma: float = 0.99
n_critics: int = 1
target_update_interval: int = 8000
def create(self, device: DeviceArg = False) -> "DQN":
return DQN(self, device)
@staticmethod
def get_type() -> str:
return "dqn"
class DQN(QLearningAlgoBase[DQNImpl, DQNConfig]):
def inner_create_impl(
self, observation_shape: Shape, action_size: int
) -> None:
q_funcs, forwarder = create_discrete_q_function(
observation_shape,
action_size,
self._config.encoder_factory,
self._config.q_func_factory,
n_ensembles=self._config.n_critics,
device=self._device,
)
targ_q_funcs, targ_forwarder = create_discrete_q_function(
observation_shape,
action_size,
self._config.encoder_factory,
self._config.q_func_factory,
n_ensembles=self._config.n_critics,
device=self._device,
)
optim = self._config.optim_factory.create(
q_funcs.named_modules(), lr=self._config.learning_rate
)
modules = DQNModules(
q_funcs=q_funcs,
targ_q_funcs=targ_q_funcs,
optim=optim,
)
self._impl = DQNImpl(
observation_shape=observation_shape,
action_size=action_size,
q_func_forwarder=forwarder,
targ_q_func_forwarder=targ_forwarder,
target_update_interval=self._config.target_update_interval,
modules=modules,
gamma=self._config.gamma,
device=self._device,
)
def get_action_type(self) -> ActionSpace:
return ActionSpace.DISCRETE
@dataclasses.dataclass()
class DoubleDQNConfig(DQNConfig):
r"""Config of Double Deep Q-Network algorithm.
The difference from DQN is that the action is taken from the current Q
function instead of the target Q function.
This modification significantly decreases overestimation bias of TD
learning.
.. math::
L(\theta) = \mathbb{E}_{s_t, a_t, r_{t+1}, s_{t+1} \sim D} [(r_{t+1}
+ \gamma Q_{\theta'}(s_{t+1}, \text{argmax}_a
Q_\theta(s_{t+1}, a)) - Q_\theta(s_t, a_t))^2]
where :math:`\theta'` is the target network parameter. The target network
parameter is synchronized every `target_update_interval` iterations.
References:
* `Hasselt et al., Deep reinforcement learning with double Q-learning.
<https://arxiv.org/abs/1509.06461>`_
Args:
observation_scaler (d3rlpy.preprocessing.ObservationScaler):
Observation preprocessor.
reward_scaler (d3rlpy.preprocessing.RewardScaler): Reward preprocessor.
learning_rate (float): Learning rate.
optim_factory (d3rlpy.models.optimizers.OptimizerFactory):
Optimizer factory.
encoder_factory (d3rlpy.models.encoders.EncoderFactory):
Encoder factory.
q_func_factory (d3rlpy.models.q_functions.QFunctionFactory):
Q function factory.
batch_size (int): Mini-batch size.
gamma (float): Discount factor.
n_critics (int): Number of Q functions.
target_update_interval (int): Interval to synchronize the target
network.
"""
batch_size: int = 32
learning_rate: float = 6.25e-5
optim_factory: OptimizerFactory = make_optimizer_field()
encoder_factory: EncoderFactory = make_encoder_field()
q_func_factory: QFunctionFactory = make_q_func_field()
gamma: float = 0.99
n_critics: int = 1
target_update_interval: int = 8000
def create(self, device: DeviceArg = False) -> "DoubleDQN":
return DoubleDQN(self, device)
@staticmethod
def get_type() -> str:
return "double_dqn"
class DoubleDQN(DQN):
def inner_create_impl(
self, observation_shape: Shape, action_size: int
) -> None:
q_funcs, forwarder = create_discrete_q_function(
observation_shape,
action_size,
self._config.encoder_factory,
self._config.q_func_factory,
n_ensembles=self._config.n_critics,
device=self._device,
)
targ_q_funcs, targ_forwarder = create_discrete_q_function(
observation_shape,
action_size,
self._config.encoder_factory,
self._config.q_func_factory,
n_ensembles=self._config.n_critics,
device=self._device,
)
optim = self._config.optim_factory.create(
q_funcs.named_modules(), lr=self._config.learning_rate
)
modules = DQNModules(
q_funcs=q_funcs,
targ_q_funcs=targ_q_funcs,
optim=optim,
)
self._impl = DoubleDQNImpl(
observation_shape=observation_shape,
action_size=action_size,
modules=modules,
q_func_forwarder=forwarder,
targ_q_func_forwarder=targ_forwarder,
target_update_interval=self._config.target_update_interval,
gamma=self._config.gamma,
device=self._device,
)
register_learnable(DQNConfig)
register_learnable(DoubleDQNConfig)