tf_agents.bandits.agents.neural_falcon_agent.NeuralFalconAgent

A neural network based agent implementing the Falcon sampling strategy.

Inherits From: GreedyRewardPredictionAgent, TFAgent

This agent receives a neural network that it trains to predict rewards. The action is chosen by a stochastic policy that uses the action distribution in: David Simchi-Levi and Yunzong Xu, "Bypassing the Monster: A Faster and Simpler Optimal Algorithm for Contextual Bandits under Realizability", Mathematics of Operations Research, 2021. https://arxiv.org/pdf/2003.12699.pdf

time_step_spec A TimeStep spec of the expected time_steps.
action_spec A nest of BoundedTensorSpec representing the actions.
reward_network A tf_agents.network.Network to be used by the agent. The network will be called with call(observation, step_type) and it is expected to provide a reward prediction for all actions. Note: when using observation_and_action_constraint_splitter, make sure the reward_network is compatible with the network-specific half of the output of the observation_and_action_constraint_splitter. In particular, observation_and_action_constraint_splitter will be called on the observation before passing to the network.
optimizer The optimizer to use for training.
num_samples_list list or tuple of tf.Variable's tracking the number of training examples for every action.
exploitation_coefficient float or callable that returns a float. Its value will be internally lower-bounded at 0. It controls how exploitative the policy behaves with respect to the predicted rewards: A larger value makes the policy sample the greedy action (one with the best predicted reward) with a higher probability.
max_exploration_probability_hint An optional float, representing a hint on the maximum exploration probability, internally clipped to [0, 1]. When this argument is set, exploitation_coefficient is ignored and the policy attempts to choose non-greedy actions with at most this probability. When such an upper bound cannot be achieved, e.g. due to insufficient training data, the policy attempts to minimize the probability of choosing non-greedy actions on a best-effort basis. For a demonstration of how it affects the policy behavior, see the unit test testTrainedPolicyWithMaxExplorationProbabilityHint in neural_falcon_agent_test.
observation_and_action_constraint_splitter A function used for masking valid/invalid actions with each state of the environment. The function takes in a full observation and returns a tuple consisting of 1) the part of the observation intended as input to the bandit agent and policy, and 2) the boolean mask. This function should also work with a TensorSpec as input, and should output TensorSpec objects for the observation and mask.
accepts_per_arm_features (bool) Whether the policy accepts per-arm features.
constraints iterable of constraints objects that are instances of tf_agents.bandits.agents.NeuralConstraint.
error_loss_fn A function for computing the error loss, taking parameters labels, predictions, and weights (any function from tf.losses would work). The default is tf.losses.mean_squared_error.
gradient_clipping A float representing the norm length to clip gradients (or None for no clipping.)
debug_summaries A Python bool, default False. When True, debug summaries are gathered.
summarize_grads_and_vars A Python bool, default False. When True, gradients and network variable summaries are written during training.
enable_summaries A Python bool, default True. When False, all summaries (debug or otherwise) should not be written.
emit_policy_info (tuple of strings) what side information we want to get as part of the policy info. Allowed values can be found in policy_utilities.PolicyInfo.
train_step_counter An optional tf.Variable to increment every time the train op is run. Defaults to the global_step.
laplacian_matrix A float Tensor or a numpy array shaped [num_actions, num_actions]. This holds the Laplacian matrix used to regularize the smoothness of the estimated expected reward function. This only applies to problems where the actions have a graph structure. If None, the regularization is not applied.
laplacian_smoothing_weight A float that determines the weight of the regularization term. Note that this has no effect if laplacian_matrix above is None.
name Python str name of this agent. All variables in this module will fall under that name. Defaults to the class name.

ValueError If the action spec contains more than one action or or it is not a bounded scalar int32 spec with minimum 0.

action_spec TensorSpec describing the action produced by the agent.
collect_data_context

collect_data_spec Returns a Trajectory spec, as expected by the collect_policy.
collect_policy Return a policy that can be used to collect data from the environment.
data_context

debug_summaries

num_samples

policy Return the current policy held by the agent.
summaries_enabled

summarize_grads_and_vars

time_step_spec Describes the TimeStep tensors expected by the agent.
train_sequence_length The number of time steps needed in experience tensors passed to train.

Train requires experience to be a Trajectory containing tensors shaped [B, T, ...]. This argument describes the value of T required.

For example, for non-RNN DQN training, T=2 because DQN requires single transitions.

If this value is None, then train can handle an unknown T (it can be determined at runtime from the data). Most RNN-based agents fall into this category.

train_step_counter

training_data_spec Returns a trajectory spec, as expected by the train() function.

Methods

compute_summaries

View source

initialize

View source

Initializes the agent.

Returns
An operation that can be used to initialize the agent.

Raises
RuntimeError If the class was not initialized properly (super.__init__ was not called).

loss

View source

Gets loss from the agent.

If the user calls this from _train, it must be in a tf.GradientTape scope in order to apply gradients to trainable variables. If intermediate gradient steps are needed, _loss and _train will return different values since _loss only supports updating all gradients at once after all losses have been calculated.

Args
experience A batch of experience data in the form of a Trajectory. The structure of experience must match that of self.training_data_spec. All tensors in experience must be shaped [batch, time, ...] where time must be equal to self.train_step_length if that property is not None.
weights (optional). A Tensor, either 0-D or shaped [batch], containing weights to be used when calculating the total train loss. Weights are typically multiplied elementwise against the per-batch loss, but the implementation is up to the Agent.
training Explicit argument to pass to loss. This typically affects network computation paths like dropout and batch normalization.
**kwargs Any additional data as args to loss.

Returns
A LossInfo loss tuple containing loss and info tensors.

Raises
RuntimeError If the class was not initialized properly (super.__init__ was not called).

post_process_policy

View source

Post process policies after training.

The policies of some agents require expensive post processing after training before they can be used. e.g. A Recommender agent might require rebuilding an index of actions. For such agents, this method will return a post processed version of the policy. The post processing may either update the existing policies in place or create a new policy, depnding on the agent. The default implementation for agents that do not want to override this method is to return agent.policy.

Returns
The post processed policy.

preprocess_sequence

View source

Defines preprocess_sequence function to be fed into replay buffers.

This defines how we preprocess the collected data before training. Defaults to pass through for most agents. Structure of experience must match that of self.collect_data_spec.

Args
experience a Trajectory shaped [batch, time, ...] or [time, ...] which represents the collected experience data.

Returns
A post processed Trajectory with the same shape as the input.

reward_loss

View source

Computes loss for reward prediction training.

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.

Raises
ValueError if the number of actions is greater than 1.

train

View source

Trains the agent.

Args
experience A batch of experience data in the form of a Trajectory. The structure of experience must match that of self.training_data_spec. All tensors in experience must be shaped [batch, time, ...] where time must be equal to self.train_step_length if that property is not None.
weights (optional). A Tensor, either 0-D or shaped [batch], containing weights to be used when calculating the total train loss. Weights are typically multiplied elementwise against the per-batch loss, but the implementation is up to the Agent.
**kwargs Any additional data to pass to the subclass.

Returns
A LossInfo loss tuple containing loss and info tensors.

  • In eager mode, the loss values are first calculated, then a train step is performed before they are returned.
  • In graph mode, executing any or all of the loss tensors will first calculate the loss value(s), then perform a train step, and return the pre-train-step LossInfo.

Raises
RuntimeError If the class was not initialized properly (super.__init__ was not called).