from .base import DynamicsBase
from ..optimizers import AdamFactory
from ..argument_utils import check_encoder, check_use_gpu
[docs]class MOPO(DynamicsBase):
""" Model-based Offline Policy Optimization.
MOPO is a model-based RL approach for offline policy optimization.
MOPO leverages the probablistic ensemble dynamics model to generate
new dynamics data with uncertainty penalties.
The ensemble dynamics model consists of :math:`N` probablistic models
:math:`\\{T_{\\theta_i}\\}_{i=1}^N`.
At each epoch, new transitions are generated via randomly picked dynamics
model :math:`T_\\theta`.
.. math::
s_{t+1}, r_{t+1} \\sim T_\\theta(s_t, a_t)
where :math:`s_t \\sim D` for the first step, otherwise :math:`s_t` is the
previous generated observation, and :math:`a_t \\sim \\pi(\\cdot|s_t)`.
The generated :math:`r_{t+1}` would be far from the ground truth if the
actions sampled from the policy function is out-of-distribution.
Thus, the uncertainty penalty reguralizes this bias.
.. math::
\\tilde{r_{t+1}} = r_{t+1} - \\lambda \\max_{i=1}^N
|| \\Sigma_i (s_t, a_t) ||
where :math:`\\Sigma(s_t, a_t)` is the estimated variance.
Finally, the generated transitions
:math:`(s_t, a_t, \\tilde{r_{t+1}}, s_{t+1})` are appended to dataset
:math:`D`.
This generation process starts with randomly sampled `n_transitions`
transitions till `horizon` steps.
Note:
Currently, MOPO only supports vector observations.
References:
* `Yu et al., MOPO: Model-based Offline Policy Optimization.
<https://arxiv.org/abs/2005.13239>`_
Args:
learning_rate (float): learning rate for dynamics model.
optim_factory (d3rlpy.optimizers.OptimizerFactory): optimizer factory.
encoder_factory (d3rlpy.encoders.EncoderFactory or str):
encoder factory.
batch_size (int): mini-batch size.
n_frames (int): the number of frames to stack for image observation.
n_ensembles (int): the number of dynamics model for ensemble.
n_transitions (int): the number of parallel trajectories to generate.
horizon (int): the number of steps to generate.
lam (float): :math:`\\lambda` for uncertainty penalties.
discrete_action (bool): flag to take discrete actions.
scaler (d3rlpy.preprocessing.scalers.Scaler or str): preprocessor.
The available options are `['pixel', 'min_max', 'standard']`.
use_gpu (bool or d3rlpy.gpu.Device): flag to use GPU or device.
impl (d3rlpy.dynamics.base.DynamicsImplBase): dynamics implementation.
Attributes:
learning_rate (float): learning rate for dynamics model.
optim_factory (d3rlpy.optimizers.OptimizerFactory): optimizer factory.
encoder_factory (d3rlpy.encoders.EncoderFactory): encoder factory.
batch_size (int): mini-batch size.
n_frames (int): the number of frames to stack for image observation.
n_ensembles (int): the number of dynamics model for ensemble.
n_transitions (int): the number of parallel trajectories to generate.
horizon (int): the number of steps to generate.
lam (float): :math:`\\lambda` for uncertainty penalties.
discrete_action (bool): flag to take discrete actions.
scaler (d3rlpy.preprocessing.scalers.Scaler): preprocessor.
use_gpu (d3rlpy.gpu.Device): flag to use GPU or device.
impl (d3rlpy.dynamics.base.DynamicsImplBase): dynamics implementation.
eval_results_ (dict): evaluation results.
"""
def __init__(self,
*,
learning_rate=1e-3,
optim_factory=AdamFactory(weight_decay=1e-4),
encoder_factory='default',
batch_size=100,
n_frames=1,
n_ensembles=5,
n_transitions=400,
horizon=5,
lam=1.0,
discrete_action=False,
scaler=None,
use_gpu=False,
impl=None,
**kwargs):
super().__init__(batch_size=batch_size,
n_frames=n_frames,
n_transitions=n_transitions,
horizon=horizon,
scaler=scaler)
self.learning_rate = learning_rate
self.optim_factory = optim_factory
self.encoder_factory = check_encoder(encoder_factory)
self.n_ensembles = n_ensembles
self.lam = lam
self.discrete_action = discrete_action
self.use_gpu = check_use_gpu(use_gpu)
self.impl = impl
[docs] def create_impl(self, observation_shape, action_size):
from .torch.mopo_impl import MOPOImpl
self.impl = MOPOImpl(observation_shape=observation_shape,
action_size=action_size,
learning_rate=self.learning_rate,
optim_factory=self.optim_factory,
encoder_factory=self.encoder_factory,
n_ensembles=self.n_ensembles,
lam=self.lam,
discrete_action=self.discrete_action,
scaler=self.scaler,
use_gpu=self.use_gpu)
[docs] def update(self, epoch, total_step, batch):
loss = self.impl.update(batch.observations, batch.actions,
batch.next_rewards, batch.next_observations)
return [loss]
def _get_loss_labels(self):
return ['loss']