Module: tf_agents.networks.dynamic_unroll_layer

View source on GitHub

Tensorflow RL Agent RNN utilities.

This module provides helper functions that Agents can use to train RNN-based policies.


The layer DynamicUnroll allows an Agent to train an RNN-based policy by running an RNN over a batch of episode chunks from a replay buffer.

The agent creates a subclass of tf.contrib.rnn.LayerRNNCell or a Keras RNN cell, such as tf.keras.layers.LSTMCell, instances of which which can themselves be wrappers of RNNCell. Training this instance involes passing it to DynamicUnroll constructor; and then pass a set of episode tensors in the form of inputs.

See the unit tests in for more details.


class DynamicUnroll: Process a history of sequences that are concatenated without padding.