Source code for d3rlpy.algos.bcq

from .base import AlgoBase
from .torch.bcq_impl import BCQImpl, DiscreteBCQImpl


[docs]class BCQ(AlgoBase): """ Batch-Constrained Q-learning algorithm. BCQ is the very first practical data-driven deep reinforcement learning lgorithm. The major difference from DDPG is that the policy function is represented as combination of conditional VAE and perturbation function in order to remedy extrapolation error emerging from target value estimation. The encoder and the decoder of the conditional VAE is represented as :math:`E_\omega` and :math:`D_\omega` respectively. .. math:: L(\omega) = E_{s_t, a_t \sim D} [(a - \\tilde{a})^2 + D_{KL}(N(\mu, \sigma)|N(0, 1))] where :math:`\mu, \sigma = E_\omega(s_t, a_t)`, :math:`\\tilde{a} = D_\omega(s_t, z)` and :math:`z \sim N(\mu, \sigma)`. The policy function is represented as a residual function with the VAE and the perturbation function represented as :math:`\\xi_\phi (s, a)`. .. math:: \pi(s, a) = a + \Phi \\xi_\phi (s, a) where :math:`a = D_\omega (s, z)`, :math:`z \sim N(0, 0.5)` and :math:`\Phi` is a perturbation scale designated by `action_flexibility`. Although the policy is learned closely to data distribution, the perturbation function can lead to more rewarded states. BCQ also leverages twin Q functions and computes weighted average over maximum values and minimum values. .. math:: L(\\theta_i) = \mathbb{E}_{s_t, a_t, r_{t+1}, s_{t+1} \sim D} [(y - Q_{\\theta_i}(s_t, a_t))^2] .. math:: y = r_{t+1} + \gamma \max_{a_i} [ \lambda \min_j Q_{\\theta_j'}(s_{t+1}, a_i) + (1 - \lambda) \max_j Q_{\\theta_j'}(s_{t+1}, a_i)] where :math:`\{a_i \sim D(s_{t+1}, z), z \sim N(0, 0.5)\}_{i=1}^n`. The number of sampled actions is designated with `n_action_samples`. Finally, the perturbation function is trained just like DDPG's policy function. .. math:: J(\phi) = \mathbb{E}_{s_t \sim D, a_t \sim D_\omega(s_t, z), z \sim N(0, 0.5)} [Q_{\\theta_1} (s_t, \pi(s_t, a_t))] At inference time, action candidates are sampled as many as `n_action_samples`, and the action with highest value estimation is taken. .. math:: \pi'(s) = \\text{argmax}_{\pi(s, a_i)} Q_{\\theta_1} (s, \pi(s, a_i)) Note: The greedy action is not deterministic because the action candidates are always randomly sampled. This might affect `save_policy` method and the performance at production. References: * `Fujimoto et al., Off-Policy Deep Reinforcement Learning without Exploration. <https://arxiv.org/abs/1812.02900>`_ 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. 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. update_actor_interval (int): interval to update policy function. lam (float): weight factor for critic ensemble. n_action_samples (int): the number of action samples to estimate action-values. action_flexibility (float): output scale of perturbation function represented as :math:`\Phi`. rl_start_epoch (int): epoch to start to update policy function and Q functions. If this is large, RL training would be more stabilized. latent_size (int): size of latent vector for Conditional VAE. beta (float): KL reguralization term for Conditional VAE. 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.bcq_impl.BCQImpl): algorithm implementation. Attributes: 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. 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. update_actor_interval (int): interval to update policy function. lam (float): weight factor for critic ensemble. n_action_samples (int): the number of action samples to estimate action-values. action_flexibility (float): output scale of perturbation function. rl_start_epoch (int): epoch to start to update policy function and Q functions. latent_size (int): size of latent vector for Conditional VAE. beta (float): KL reguralization term for Conditional VAE. 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.bcq_impl.BCQImpl): algorithm implementation. eval_results_ (dict): evaluation results. """ def __init__(self, *, actor_learning_rate=1e-3, critic_learning_rate=1e-3, imitator_learning_rate=1e-3, batch_size=100, n_frames=1, gamma=0.99, tau=0.005, n_critics=2, bootstrap=False, share_encoder=False, update_actor_interval=1, lam=0.75, n_action_samples=100, action_flexibility=0.05, rl_start_epoch=0, latent_size=32, beta=0.5, 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.imitator_learning_rate = imitator_learning_rate self.gamma = gamma self.tau = tau self.n_critics = n_critics self.bootstrap = bootstrap self.share_encoder = share_encoder self.update_actor_interval = update_actor_interval self.lam = lam self.n_action_samples = n_action_samples self.action_flexibility = action_flexibility self.rl_start_epoch = rl_start_epoch self.latent_size = latent_size self.beta = beta 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 = BCQImpl(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, gamma=self.gamma, tau=self.tau, n_critics=self.n_critics, bootstrap=self.bootstrap, share_encoder=self.share_encoder, lam=self.lam, n_action_samples=self.n_action_samples, action_flexibility=self.action_flexibility, latent_size=self.latent_size, beta=self.beta, 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): imitator_loss = self.impl.update_imitator(batch.observations, batch.actions) if epoch >= self.rl_start_epoch: critic_loss = self.impl.update_critic(batch.observations, batch.actions, batch.next_rewards, batch.next_observations, batch.terminals) if total_step % self.update_actor_interval == 0: actor_loss = self.impl.update_actor(batch.observations) self.impl.update_actor_target() self.impl.update_critic_target() else: actor_loss = None else: critic_loss = None actor_loss = None return critic_loss, actor_loss, imitator_loss
[docs] def sample_action(self, x): """ BCQ does not support sampling action. """ raise NotImplementedError('BCQ does not support sampling action.')
def _get_loss_labels(self): return ['critic_loss', 'actor_loss', 'imitator_loss']
[docs]class DiscreteBCQ(AlgoBase): """ Discrete version of Batch-Constrained Q-learning algorithm. Discrete version takes theories from the continuous version, but the algorithm is much simpler than that. The imitation function :math:`G_\omega(a|s)` is trained as supervised learning just like Behavior Cloning. .. math:: L(\omega) = \mathbb{E}_{a_t, s_t \sim D} [-\sum_a p(a|s_t) \log G_\omega(a|s_t)] With this imitation function, the greedy policy is defined as follows. .. math:: \pi(s_t) = \\text{argmax}_{a|G_\omega(a|s_t) / \max_{\\tilde{a}} G_\omega(\\tilde{a}|s_t) > \\tau} Q_\\theta (s_t, a) which eliminates actions with probabilities :math:`\\tau` times smaller than the maximum one. Finally, the loss function is computed in Double DQN style with the above constrained policy. .. 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(s_{t+1})) - Q_\\theta(s_t, a_t))^2] References: * `Fujimoto et al., Off-Policy Deep Reinforcement Learning without Exploration. <https://arxiv.org/abs/1812.02900>`_ * `Fujimoto et al., Benchmarking Batch Deep Reinforcement Learning Algorithms. <https://arxiv.org/abs/1910.01708>`_ Args: learning_rate (float): learning rate. batch_size (int): mini-batch size. n_frames (int): the number of frames to stack for image observation. gamma (float): discount factor. 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. action_flexibility (float): probability threshold represented as :math:`\tau`. beta (float): reguralization term for imitation function. eps (float): :math:`\\epsilon` for Adam optimizer. target_update_interval (int): interval to update the target network. 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.bcq_impl.DiscreteBCQImpl): algorithm implementation. Attributes: learning_rate (float): learning rate. batch_size (int): mini-batch size. n_frames (int): the number of frames to stack for image observation. gamma (float): discount factor. 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. action_flexibility (float): probability threshold represented as :math:`\tau`. beta (float): reguralization term for imitation function. eps (float): :math:`\\epsilon` for Adam optimizer. target_update_interval (int): interval to update the target network. 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.bcq_impl.DiscreteBCQImpl): algorithm implementation. eval_results_ (dict): evaluation results. """ def __init__(self, *, learning_rate=6.25e-5, batch_size=32, n_frames=1, gamma=0.99, n_critics=1, bootstrap=False, share_encoder=False, action_flexibility=0.3, beta=0.5, eps=1.5e-4, target_update_interval=8e3, 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.learning_rate = learning_rate self.gamma = gamma self.n_critics = n_critics self.bootstrap = bootstrap self.share_encoder = share_encoder self.action_flexibility = action_flexibility self.beta = beta self.eps = eps self.target_update_interval = target_update_interval 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 = DiscreteBCQImpl(observation_shape=observation_shape, action_size=action_size, learning_rate=self.learning_rate, gamma=self.gamma, n_critics=self.n_critics, bootstrap=self.bootstrap, share_encoder=self.share_encoder, action_flexibility=self.action_flexibility, beta=self.beta, 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): loss = self.impl.update(batch.observations, batch.actions, batch.next_rewards, batch.next_observations, batch.terminals) if total_step % self.target_update_interval == 0: self.impl.update_target() return [loss]
def _get_loss_labels(self): return ['loss']