Source code for d3rlpy.algos.ddpg

from .base import AlgoBase
from .torch.ddpg_impl import DDPGImpl


[docs]class DDPG(AlgoBase): """ Deep Deterministic Policy Gradients algorithm. DDPG is an actor-critic algorithm that trains a Q function parametrized with :math:`\\theta` and a policy function parametrized with :math:`\\phi`. .. 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}, \\pi_{\\phi'}(s_{t+1})) - Q_\\theta(s_t, a_t))^2] .. math:: J(\\phi) = \\mathbb{E}_{s_t \\sim D} [Q_\\theta(s_t, \\pi_\\phi(s_t))] where :math:`\\theta'` and :math:`\\phi` are the target network parameters. There target network parameters are updated every iteration. .. math:: \\theta' \\gets \\tau \\theta + (1 - \\tau) \\theta' \\phi' \\gets \\tau \\phi + (1 - \\tau) \\phi' References: * `Silver et al., Deterministic policy gradient algorithms. <http://proceedings.mlr.press/v32/silver14.html>`_ * `Lillicrap et al., Continuous control with deep reinforcement learning. <https://arxiv.org/abs/1509.02971>`_ Args: actor_learning_rate (float): learning rate for policy function. critic_learning_rate (float): learning rate for Q function. batch_size (int): mini-batch size. n_frames (int): the number of frames to stack for image observation. gamma (float): discount factor. tau (float): target network synchronization coefficiency. n_critics (int): the number of Q functions for ensemble. bootstrap (bool): flag to bootstrap Q functions. share_encoder (bool): flag to share encoder network. reguralizing_rate (float): reguralizing term for policy function. eps (float): :math:`\\epsilon` for Adam optimizer. use_batch_norm (bool): flag to insert batch normalization layers. q_func_type (str): type of Q function. Available options are `['mean', 'qr', 'iqn', 'fqf']`. 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']` augmentation (d3rlpy.augmentation.AugmentationPipeline or list(str)): augmentation pipeline. n_augmentations (int): the number of data augmentations to update. encoder_params (dict): optional arguments for encoder setup. If the observation is pixel, you can pass ``filters`` with list of tuples consisting with ``(filter_size, kernel_size, stride)`` and ``feature_size`` with an integer scaler for the last linear layer size. If the observation is vector, you can pass ``hidden_units`` with list of hidden unit sizes. dynamics (d3rlpy.dynamics.base.DynamicsBase): dynamics model for data augmentation. impl (d3rlpy.algos.torch.ddpg_impl.DDPGImpl): algorithm implementation. Attributes: actor_learning_rate (float): learning rate for policy function. critic_learning_rate (float): learning rate for Q function. batch_size (int): mini-batch size. n_frames (int): the number of frames to stack for image observation. gamma (float): discount factor. tau (float): target network synchronization coefficiency. n_critics (int): the number of Q functions for ensemble. bootstrap (bool): flag to bootstraep Q functions. share_encoder (bool): flag to share encoder network. reguralizing_rate (float): reguralizing term for policy function. eps (float): :math:`\\epsilon` for Adam optimizer. use_batch_norm (bool): flag to insert batch normalization layers. q_func_type (str): type of Q function. use_gpu (d3rlpy.gpu.Device): GPU device. scaler (d3rlpy.preprocessing.Scaler): preprocessor. augmentation (d3rlpy.augmentation.AugmentationPipeline): augmentation pipeline. n_augmentations (int): the number of data augmentations to update. encoder_params (dict): optional arguments for encoder setup. dynamics (d3rlpy.dynamics.base.DynamicsBase): dynamics model. impl (d3rlpy.algos.torch.ddpg_impl.DDPGImpl): algorithm implementation. eval_results_ (dict): evaluation results. """ def __init__(self, *, actor_learning_rate=3e-4, critic_learning_rate=3e-4, batch_size=100, n_frames=1, gamma=0.99, tau=0.005, n_critics=1, bootstrap=False, share_encoder=False, reguralizing_rate=1e-10, eps=1e-8, use_batch_norm=False, q_func_type='mean', use_gpu=False, scaler=None, augmentation=[], n_augmentations=1, encoder_params={}, dynamics=None, impl=None, **kwargs): super().__init__(batch_size=batch_size, n_frames=n_frames, scaler=scaler, augmentation=augmentation, dynamics=dynamics, use_gpu=use_gpu) self.actor_learning_rate = actor_learning_rate self.critic_learning_rate = critic_learning_rate self.gamma = gamma self.tau = tau self.n_critics = n_critics self.bootstrap = bootstrap self.share_encoder = share_encoder self.reguralizing_rate = reguralizing_rate self.eps = eps self.use_batch_norm = use_batch_norm self.q_func_type = q_func_type self.n_augmentations = n_augmentations self.encoder_params = encoder_params self.impl = impl
[docs] def create_impl(self, observation_shape, action_size): self.impl = DDPGImpl(observation_shape=observation_shape, action_size=action_size, actor_learning_rate=self.actor_learning_rate, critic_learning_rate=self.critic_learning_rate, gamma=self.gamma, tau=self.tau, n_critics=self.n_critics, bootstrap=self.bootstrap, share_encoder=self.share_encoder, reguralizing_rate=self.reguralizing_rate, eps=self.eps, use_batch_norm=self.use_batch_norm, q_func_type=self.q_func_type, use_gpu=self.use_gpu, scaler=self.scaler, augmentation=self.augmentation, n_augmentations=self.n_augmentations, encoder_params=self.encoder_params) self.impl.build()
[docs] def update(self, epoch, itr, batch): critic_loss = self.impl.update_critic(batch.observations, batch.actions, batch.next_rewards, batch.next_observations, batch.terminals) actor_loss = self.impl.update_actor(batch.observations) self.impl.update_critic_target() self.impl.update_actor_target() return critic_loss, actor_loss
def _get_loss_labels(self): return ['critic_loss', 'actor_loss']