The following diagram provides a conceptual overview of data flow between different components in RLlib. We start with an ``Environment``, which given an action produces an observation. The observation is preprocessed by a ``Preprocessor`` and ``Filter`` (e.g. for running mean normalization) before being sent to a neural network ``Model``. The model output is in turn interpreted by an ``ActionDistribution`` to determine the next action.
..image:: rllib-components.svg
The components highlighted in green can be replaced with custom user-defined implementations, as described in the next sections. The purple components are RLlib internal, which means they can only be modified by changing the algorithm source code.
RLlib picks default models based on a simple heuristic: a `vision network <https://github.com/ray-project/ray/blob/master/rllib/models/tf/visionnet_v1.py>`__ for image observations, and a `fully connected network <https://github.com/ray-project/ray/blob/master/rllib/models/tf/fcnet_v1.py>`__ for everything else. These models can be configured via the ``model`` config key, documented in the model `catalog <https://github.com/ray-project/ray/blob/master/rllib/models/catalog.py>`__. Note that you'll probably have to configure ``conv_filters`` if your environment observations have custom sizes, e.g., ``"model": {"dim": 42, "conv_filters": [[16, [4, 4], 2], [32, [4, 4], 2], [512, [11, 11], 1]]}`` for 42x42 observations.
In addition, if you set ``"model": {"use_lstm": true}``, then the model output will be further processed by a `LSTM cell <https://github.com/ray-project/ray/blob/master/rllib/models/tf/lstm.py>`__. More generally, RLlib supports the use of recurrent models for its policy gradient algorithms (A3C, PPO, PG, IMPALA), and RNN support is built into its policy evaluation utilities.
For preprocessors, RLlib tries to pick one of its built-in preprocessor based on the environment's observation space. Discrete observations are one-hot encoded, Atari observations downscaled, and Tuple and Dict observations flattened (these are unflattened and accessible via the ``input_dict`` parameter in custom models). Note that for Atari, RLlib defaults to using the `DeepMind preprocessors <https://github.com/ray-project/ray/blob/master/rllib/env/atari_wrappers.py>`__, which are also used by the OpenAI baselines library.
TFModelV2 replaces the previous ``rllib.models.Model`` class, which did not support Keras-style reuse of variables. The ``rllib.models.Model`` class is deprecated and should not be used.
Custom TF models should subclass `TFModelV2 <https://github.com/ray-project/ray/blob/master/rllib/models/tf/tf_modelv2.py>`__ to implement the ``__init__()`` and ``forward()`` methods. Forward takes in a dict of tensor inputs (the observation ``obs``, ``prev_action``, and ``prev_reward``, ``is_training``), optional RNN state, and returns the model output of size ``num_outputs`` and the new state. You can also override extra methods of the model such as ``value_function`` to implement a custom value branch. Additional supervised / self-supervised losses can be added via the ``custom_loss`` method:
For a full example of a custom model in code, see the `keras model example <https://github.com/ray-project/ray/blob/master/rllib/examples/custom_keras_model.py>`__. You can also reference the `unit tests <https://github.com/ray-project/ray/blob/master/rllib/tests/test_nested_spaces.py>`__ for Tuple and Dict spaces, which show how to access nested observation fields.
Instead of using the ``use_lstm: True`` option, it can be preferable use a custom recurrent model. This provides more control over postprocessing of the LSTM output and can also allow the use of multiple LSTM cells to process different portions of the input. For a RNN model it is preferred to subclass ``RecurrentTFModelV2`` to implement ``__init__()``, ``get_initial_state()``, and ``forward_rnn()``. You can check out the `custom_keras_rnn_model.py <https://github.com/ray-project/ray/blob/master/rllib/examples/custom_keras_rnn_model.py>`__ model as an example to implement your own model:
You can use ``tf.layers.batch_normalization(x, training=input_dict["is_training"])`` to add batch norm layers to your custom model: `code example <https://github.com/ray-project/ray/blob/master/rllib/examples/batch_norm_model.py>`__. RLlib will automatically run the update ops for the batch norm layers during optimization (see `tf_policy.py <https://github.com/ray-project/ray/blob/master/rllib/policy/tf_policy.py>`__ and `multi_gpu_impl.py <https://github.com/ray-project/ray/blob/master/rllib/optimizers/multi_gpu_impl.py>`__ for the exact handling of these updates).
In case RLlib does not properly detect the update ops for your custom model, you can override the ``update_ops()`` method to return the list of ops to run for updates.
Similarly, you can create and register custom PyTorch models for use with PyTorch-based algorithms (e.g., A2C, PG, QMIX). See these examples of `fully connected <https://github.com/ray-project/ray/blob/master/rllib/models/torch/fcnet.py>`__, `convolutional <https://github.com/ray-project/ray/blob/master/rllib/models/torch/visionnet.py>`__, and `recurrent <https://github.com/ray-project/ray/blob/master/rllib/agents/qmix/model.py>`__ torch models.
Custom preprocessors should subclass the RLlib `preprocessor class <https://github.com/ray-project/ray/blob/master/rllib/models/preprocessors.py>`__ and be registered in the model catalog. Note that you can alternatively use `gym wrapper classes <https://github.com/openai/gym/tree/master/gym/wrappers>`__ around your environment instead of preprocessors.
Similar to custom models and preprocessors, you can also specify a custom action distribution class as follows. The action dist class is passed a reference to the ``model``, which you can use to access ``model.model_config`` or other attributes of the model. This is commonly used to implement `autoregressive action outputs <#autoregressive-action-distributions>`__.
..code-block:: python
import ray
import ray.rllib.agents.ppo as ppo
from ray.rllib.models import ModelCatalog
from ray.rllib.models.preprocessors import Preprocessor
You can mix supervised losses into any RLlib algorithm through custom models. For example, you can add an imitation learning loss on expert experiences, or a self-supervised autoencoder loss within the model. These losses can be defined over either policy evaluation inputs, or data read from `offline storage <rllib-offline.html#input-pipeline-for-supervised-losses>`__.
**TensorFlow**: To add a supervised loss to a custom TF model, you need to override the ``custom_loss()`` method. This method takes in the existing policy loss for the algorithm, which you can add your own supervised loss to before returning. For debugging, you can also return a dictionary of scalar tensors in the ``metrics()`` method. Here is a `runnable example <https://github.com/ray-project/ray/blob/master/rllib/examples/custom_loss.py>`__ of adding an imitation loss to CartPole training that is defined over a `offline dataset <rllib-offline.html#input-pipeline-for-supervised-losses>`__.
**PyTorch**: There is no explicit API for adding losses to custom torch models. However, you can modify the loss in the policy definition directly. Like for TF models, offline datasets can be incorporated by creating an input reader and calling ``reader.next()`` in the loss forward pass.
Custom models can be used to work with environments where (1) the set of valid actions `varies per step <https://neuro.cs.ut.ee/the-use-of-embeddings-in-openai-five>`__, and/or (2) the number of valid actions is `very large <https://arxiv.org/abs/1811.00260>`__. The general idea is that the meaning of actions can be completely conditioned on the observation, i.e., the ``a`` in ``Q(s, a)`` becomes just a token in ``[0, MAX_AVAIL_ACTIONS)`` that only has meaning in the context of ``s``. This works with algorithms in the `DQN and policy-gradient families <rllib-env.html>`__ and can be implemented as follows:
1. The environment should return a mask and/or list of valid action embeddings as part of the observation for each step. To enable batching, the number of actions can be allowed to vary from 1 to some max number:
2. A custom model can be defined that can interpret the ``action_mask`` and ``avail_actions`` portions of the observation. Here the model computes the action logits via the dot product of some network output and each action embedding. Invalid actions can be masked out of the softmax by scaling the probability to zero:
Depending on your use case it may make sense to use just the masking, just action embeddings, or both. For a runnable example of this in code, check out `parametric_action_cartpole.py <https://github.com/ray-project/ray/blob/master/rllib/examples/parametric_action_cartpole.py>`__. Note that since masking introduces ``tf.float32.min`` values into the model output, this technique might not work with all algorithm options. For example, algorithms might crash if they incorrectly process the ``tf.float32.min`` values. The cartpole example has working configurations for DQN (must set ``hiddens=[]``), PPO (must disable running mean and set ``vf_share_layers=True``), and several other algorithms. Not all algorithms support parametric actions; see the `feature compatibility matrix <rllib-env.html#feature-compatibility-matrix>`__.
In an action space with multiple components (e.g., ``Tuple(a1, a2)``), you might want ``a2`` to be conditioned on the sampled value of ``a1``, i.e., ``a2_sampled ~ P(a2 | a1_sampled, obs)``. Normally, ``a1`` and ``a2`` would be sampled independently, reducing the expressivity of the policy.
To do this, you need both a custom model that implements the autoregressive pattern, and a custom action distribution class that leverages that model. The `autoregressive_action_dist.py <https://github.com/ray-project/ray/blob/master/rllib/examples/autoregressive_action_dist.py>`__ example shows how this can be implemented for a simple binary action space. For a more complex space, a more efficient architecture such as a `MADE <https://arxiv.org/abs/1502.03509>`__ is recommended. Note that sampling a `N-part` action requires `N` forward passes through the model, however computing the log probability of an action can be done in one pass:
Not all algorithms support autoregressive action distributions; see the `feature compatibility matrix <rllib-env.html#feature-compatibility-matrix>`__.
