Method __init__

acme/agents/tf/d4pg/agent.py:252–365

Initialize the agent. Args: environment_spec: description of the actions, observations, etc. policy_network: the online (optimized) policy. critic_network: the online critic. observation_network: optional network to transform the observations before they are fed

(
      self,
      environment_spec: specs.EnvironmentSpec,
      policy_network: snt.Module,
      critic_network: snt.Module,
      observation_network: types.TensorTransformation = tf.identity,
      discount: float = 0.99,
      batch_size: int = 256,
      prefetch_size: int = 4,
      target_update_period: int = 100,
      policy_optimizer: Optional[snt.Optimizer] = None,
      critic_optimizer: Optional[snt.Optimizer] = None,
      min_replay_size: int = 1000,
      max_replay_size: int = 1000000,
      samples_per_insert: float = 32.0,
      n_step: int = 5,
      sigma: float = 0.3,
      clipping: bool = True,
      replay_table_name: str = reverb_adders.DEFAULT_PRIORITY_TABLE,
      counter: Optional[counting.Counter] = None,
      logger: Optional[loggers.Logger] = None,
      checkpoint: bool = True,
  )

  """

  def __init__(
      self,
      environment_spec: specs.EnvironmentSpec,
      policy_network: snt.Module,
      critic_network: snt.Module,
      observation_network: types.TensorTransformation = tf.identity,
      discount: float = 0.99,
      batch_size: int = 256,
      prefetch_size: int = 4,
      target_update_period: int = 100,
      policy_optimizer: Optional[snt.Optimizer] = None,
      critic_optimizer: Optional[snt.Optimizer] = None,
      min_replay_size: int = 1000,
      max_replay_size: int = 1000000,
      samples_per_insert: float = 32.0,
      n_step: int = 5,
      sigma: float = 0.3,
      clipping: bool = True,
      replay_table_name: str = reverb_adders.DEFAULT_PRIORITY_TABLE,
      counter: Optional[counting.Counter] = None,
      logger: Optional[loggers.Logger] = None,
      checkpoint: bool = True,
  ):
    """Initialize the agent.

    Args:
      environment_spec: description of the actions, observations, etc.
      policy_network: the online (optimized) policy.
      critic_network: the online critic.
      observation_network: optional network to transform the observations before
        they are fed into any network.
      discount: discount to use for TD updates.
      batch_size: batch size for updates.
      prefetch_size: size to prefetch from replay.
      target_update_period: number of learner steps to perform before updating
        the target networks.
      policy_optimizer: optimizer for the policy network updates.
      critic_optimizer: optimizer for the critic network updates.
      min_replay_size: minimum replay size before updating.
      max_replay_size: maximum replay size.
      samples_per_insert: number of samples to take from replay for every insert
        that is made.
      n_step: number of steps to squash into a single transition.
      sigma: standard deviation of zero-mean, Gaussian exploration noise.
      clipping: whether to clip gradients by global norm.
      replay_table_name: string indicating what name to give the replay table.
      counter: counter object used to keep track of steps.
      logger: logger object to be used by learner.
      checkpoint: boolean indicating whether to checkpoint the learner.
    """
    # Create the Builder object which will internally create agent components.
    builder = D4PGBuilder(
        # TODO(mwhoffman): pass the config dataclass in directly.
        # TODO(mwhoffman): use the limiter rather than the workaround below.
        # Right now this modifies min_replay_size and samples_per_insert so that
        # they are not controlled by a limiter and are instead handled by the
        # Agent base class (the above TODO directly references this behavior).
        D4PGConfig(