tf_agents.bandits.policies.constraints.RelativeQuantileConstraint

Class for representing a trainable relative quantile constraint.

Inherits From: NeuralConstraint, BaseConstraint

tf_agents.bandits.policies.constraints.RelativeQuantileConstraint(
    time_step_spec: tf_agents.typing.types.TimeStep,
    action_spec: tf_agents.typing.types.BoundedTensorSpec,
    constraint_network: tf_agents.typing.types.Network,
    quantile: float = 0.5,
    comparator_fn: tf_agents.typing.types.ComparatorFn = tf.greater,
    baseline_action_fn: Optional[Callable[[types.Tensor], types.Tensor]] = None,
    name: Text = 'RelativeQuantileConstraint'
)

This constraint class implements a relative quantile constraint such as

Q_tau(action) >= Q_tau(baseline_action)

or

Q_tau(action) <= Q_tau(baseline_action)

time_step_spec A TimeStep spec of the expected time_steps.
action_spec A nest of BoundedTensorSpec representing the actions.
constraint_network An instance of tf_agents.network.Network used to provide estimates of action feasibility. The input structure should be consistent with the observation_spec.
quantile A float between 0. and 1., the quantile we want to regress.
comparator_fn a comparator function, such as tf.greater or tf.less.
baseline_action_fn a callable that given the observation returns the baseline action. If None, the baseline action is set to 0.
name Python str name of this agent. All variables in this module will fall under that name. Defaults to the class name.

constraint_network

observation_spec

Methods

compute_loss

View source

compute_loss(
    observations: tf_agents.typing.types.NestedTensor,
    actions: tf_agents.typing.types.NestedTensor,
    rewards: tf_agents.typing.types.Tensor,
    weights: Optional[types.Float] = None,
    training: bool = False
) -> tf_agents.typing.types.Tensor

Computes loss for training the constraint network.

Args
observations A batch of observations.
actions A batch of actions.
rewards A batch of rewards.
weights Optional scalar or elementwise (per-batch-entry) importance weights. The output batch loss will be scaled by these weights, and the final scalar loss is the mean of these values.
training Whether the loss is being used for training.

Returns
loss A Tensor containing the loss for the training step.

initialize

View source

initialize()

Returns an op to initialize the constraint.

__call__

View source

__call__(
    observation, actions=None
)

Returns the probability of input actions being feasible.