Source code for d3rlpy.algos.td3

from .base import AlgoBase
from .torch.td3_impl import TD3Impl


[docs]class TD3(AlgoBase): """ Twin Delayed Deep Deterministic Policy Gradients algorithm. TD3 is an improved DDPG-based algorithm. Major differences from DDPG are as follows. * TD3 has twin Q functions to reduce overestimation bias at TD learning. The number of Q functions can be designated by `n_critics`. * TD3 adds noise to target value estimation to avoid overfitting with the deterministic policy. * TD3 updates the policy function after several Q function updates in order to reduce variance of action-value estimation. The interval of the policy function update can be designated by `update_actor_interval`. .. math:: L(\\theta_i) = \mathbb{E}_{s_t, a_t, r_{t+1}, s_{t+1} \sim D} [(r_{t+1} + \gamma \min_j Q_{\\theta_j'}(s_{t+1}, \pi_{\phi'}(s_{t+1}) + \epsilon) - Q_{\\theta_i}(s_t, a_t))^2] .. math:: J(\\phi) = \mathbb{E}_{s_t \sim D} [\min_i Q_{\\theta_i}(s_t, \pi_\phi(s_t))] where :math:`\\epsilon \sim clip (N(0, \\sigma), -c, c)` References: * `Fujimoto et al., Addressing Function Approximation Error in Actor-Critic Methods. <https://arxiv.org/abs/1802.09477>`_ Args: actor_learning_rate (float): learning rate for a policy function. critic_learning_rate (float): learning rate for Q functions. 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. reguralizing_rate (float): reguralizing term for policy function. 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. target_smoothing_sigma (float): standard deviation for target noise. target_smoothing_clip (float): clipping range for target noise. update_actor_interval (int): interval to update policy function described as `delayed policy update` in the paper. 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.td3_impl.TD3Impl): algorithm implementation. Attributes: actor_learning_rate (float): learning rate for a policy function. critic_learning_rate (float): learning rate for Q functions. 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. reguralizing_rate (float): reguralizing term for policy function. 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. target_smoothing_sigma (float): standard deviation for target noise. target_smoothing_clip (float): clipping range for target noise. update_actor_interval (int): interval to update policy function described as `delayed policy update` in the paper. 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.td3_impl.TD3Impl): 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, reguralizing_rate=0.0, n_critics=2, bootstrap=False, share_encoder=False, target_smoothing_sigma=0.2, target_smoothing_clip=0.5, update_actor_interval=2, 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.reguralizing_rate = reguralizing_rate self.n_critics = n_critics self.bootstrap = bootstrap self.share_encoder = share_encoder self.target_smoothing_sigma = target_smoothing_sigma self.target_smoothing_clip = target_smoothing_clip self.update_actor_interval = update_actor_interval 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 = TD3Impl(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, reguralizing_rate=self.reguralizing_rate, n_critics=self.n_critics, bootstrap=self.bootstrap, share_encoder=self.share_encoder, target_smoothing_sigma=self.target_smoothing_sigma, target_smoothing_clip=self.target_smoothing_clip, 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, total_step, batch): critic_loss = self.impl.update_critic(batch.observations, batch.actions, batch.next_rewards, batch.next_observations, batch.terminals) # delayed policy update if total_step % self.update_actor_interval == 0: actor_loss = self.impl.update_actor(batch.observations) self.impl.update_critic_target() self.impl.update_actor_target() else: actor_loss = None return critic_loss, actor_loss
def _get_loss_labels(self): return ['critic_loss', 'actor_loss']