This page describes the internal concepts used to implement algorithms in RLlib. You might find this useful if modifying or adding new algorithms to RLlib.
Policy classes encapsulate the core numerical components of RL algorithms. This typically includes the policy model that determines actions to take, a trajectory postprocessor for experiences, and a loss function to improve the policy given postprocessed experiences. For a simple example, see the policy gradients `policy definition <https://github.com/ray-project/ray/blob/master/python/ray/rllib/agents/pg/pg_policy.py>`__.
Most interaction with deep learning frameworks is isolated to the `Policy interface <https://github.com/ray-project/ray/blob/master/python/ray/rllib/policy/policy.py>`__, allowing RLlib to support multiple frameworks. To simplify the definition of policies, RLlib includes `Tensorflow <#building-policies-in-tensorflow>`__ and `PyTorch-specific <#building-policies-in-pytorch>`__ templates. You can also write your own from scratch. Here is an example:
The above basic policy, when run, will produce batches of observations with the basic ``obs``, ``new_obs``, ``actions``, ``rewards``, ``dones``, and ``infos`` columns. There are two more mechanisms to pass along and emit extra information:
**Policy recurrent state**: Suppose you want to compute actions based on the current timestep of the episode. While it is possible to have the environment provide this as part of the observation, we can instead compute and store it as part of the Policy recurrent state:
..code-block:: python
def get_initial_state(self):
"""Returns initial RNN state for the current policy."""
return [0] # list of single state element (t=0)
# you could also return multiple values, e.g., [0, "foo"]
# can access array of the state elements at each timestep
# or state_in_1, 2, etc. if there are multiple state elements
assert "state_in_0" in samples.keys()
assert "state_out_0" in samples.keys()
**Extra action info output**: You can also emit extra outputs at each step which will be available for learning on. For example, you might want to output the behaviour policy logits as extra action info, which can be used for importance weighting, but in general arbitrary values can be stored here (as long as they are convertible to numpy arrays):
..code-block:: python
def compute_actions(self,
obs_batch,
state_batches,
prev_action_batch=None,
prev_reward_batch=None,
info_batch=None,
episodes=None,
**kwargs):
action_info_batch = {
"some_value": ["foo" for _ in obs_batch],
"other_value": [12345 for _ in obs_batch],
}
return ..., [], action_info_batch
def learn_on_batch(self, samples):
# can access array of the extra values at each timestep
assert "some_value" in samples.keys()
assert "other_value" in samples.keys()
Building Policies in TensorFlow
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
This section covers how to build a TensorFlow RLlib policy using ``tf_policy_template.build_tf_policy()``.
To start, you first have to define a loss function. In RLlib, loss functions are defined over batches of trajectory data produced by policy evaluation. A basic policy gradient loss that only tries to maximize the 1-step reward can be defined as follows:
..code-block:: python
import tensorflow as tf
from ray.rllib.policy.sample_batch import SampleBatch
In the above snippet, ``actions`` is a Tensor placeholder of shape ``[batch_size, action_dim...]``, and ``rewards`` is a placeholder of shape ``[batch_size]``. The ``policy.action_dist`` object is an `ActionDistribution <rllib-package-ref.html#ray.rllib.models.ActionDistribution>`__ that represents the output of the neural network policy model. Passing this loss function to ``build_tf_policy`` is enough to produce a very basic TF policy:
..code-block:: python
from ray.rllib.policy.tf_policy_template import build_tf_policy
If you run the above snippet `(runnable file here) <https://github.com/ray-project/ray/blob/master/python/ray/rllib/examples/custom_tf_policy.py>`__, you'll probably notice that CartPole doesn't learn so well:
Let's modify our policy loss to include rewards summed over time. To enable this advantage calculation, we need to define a *trajectory postprocessor* for the policy. This can be done by defining ``postprocess_fn``:
..code-block:: python
from ray.rllib.evaluation.postprocessing import compute_advantages, \
The ``postprocess_advantages()`` function above uses calls RLlib's ``compute_advantages`` function to compute advantages for each timestep. If you re-run the trainer with this improved policy, you'll find that it quickly achieves the max reward of 200.
You might be wondering how RLlib makes the advantages placeholder automatically available as ``batch_tensors[Postprocessing.ADVANTAGES]``. When building your policy, RLlib will create a "dummy" trajectory batch where all observations, actions, rewards, etc. are zeros. It then calls your ``postprocess_fn``, and generates TF placeholders based on the numpy shapes of the postprocessed batch. RLlib tracks which placeholders that ``loss_fn`` and ``stats_fn`` access, and then feeds the corresponding sample data into those placeholders during loss optimization. You can also access these placeholders via ``policy.get_placeholder(<name>)`` after loss initialization.
**Example 1: Proximal Policy Optimization**
In the above section you saw how to compose a simple policy gradient algorithm with RLlib. In this example, we'll dive into how PPO was built with RLlib and how you can modify it. First, check out the `PPO trainer definition <https://github.com/ray-project/ray/blob/master/python/ray/rllib/agents/ppo/ppo.py>`__:
..code-block:: python
PPOTrainer = build_trainer(
name="PPOTrainer",
default_config=DEFAULT_CONFIG,
default_policy=PPOTFPolicy,
make_policy_optimizer=choose_policy_optimizer,
validate_config=validate_config,
after_optimizer_step=update_kl,
before_train_step=warn_about_obs_filter,
after_train_result=warn_about_bad_reward_scales)
Besides some boilerplate for defining the PPO configuration and some warnings, there are two important arguments to take note of here: ``make_policy_optimizer=choose_policy_optimizer``, and ``after_optimizer_step=update_kl``.
The ``choose_policy_optimizer`` function chooses which `Policy Optimizer <#policy-optimization>`__ to use for distributed training. You can think of these policy optimizers as coordinating the distributed workflow needed to improve the policy. Depending on the trainer config, PPO can switch between a simple synchronous optimizer, or a multi-GPU optimizer that implements minibatch SGD (the default):
Suppose we want to customize PPO to use an asynchronous-gradient optimization strategy similar to A3C. To do that, we could define a new function that returns ``AsyncGradientsOptimizer`` and pass in ``make_policy_optimizer=make_async_optimizer`` when building the trainer:
..code-block:: python
from ray.rllib.agents.ppo.ppo_policy import *
from ray.rllib.optimizers import AsyncGradientsOptimizer
from ray.rllib.policy.tf_policy_template import build_tf_policy
Now let's take a look at the ``update_kl`` function. This is used to adaptively adjust the KL penalty coefficient on the PPO loss, which bounds the policy change per training step. You'll notice the code handles both single and multi-agent cases (where there are be multiple policies each with different KL coeffs):
..code-block:: python
def update_kl(trainer, fetches):
if "kl" in fetches:
# single-agent
trainer.workers.local_worker().for_policy(
lambda pi: pi.update_kl(fetches["kl"]))
else:
def update(pi, pi_id):
if pi_id in fetches:
pi.update_kl(fetches[pi_id]["kl"])
else:
logger.debug("No data for {}, not updating kl".format(pi_id))
The ``update_kl`` method on the policy is defined in `PPOTFPolicy <https://github.com/ray-project/ray/blob/master/python/ray/rllib/agents/ppo/ppo_policy.py>`__ via the ``KLCoeffMixin``, along with several other advanced features. Let's look at each new feature used by the policy:
``stats_fn``: The stats function returns a dictionary of Tensors that will be reported with the training results. This also includes the ``kl`` metric which is used by the trainer to adjust the KL penalty. Note that many of the values below reference ``policy.loss_obj``, which is assigned by ``loss_fn`` (not shown here since the PPO loss is quite complex). RLlib will always call ``stats_fn`` after ``loss_fn``, so you can rely on using values saved by ``loss_fn`` as part of your statistics:
``extra_actions_fetches_fn``: This function defines extra outputs that will be recorded when generating actions with the policy. For example, this enables saving the raw policy logits in the experience batch, which e.g. means it can be referenced in the PPO loss function via ``batch_tensors[BEHAVIOUR_LOGITS]``. Other values such as the current value prediction can also be emitted for debugging or optimization purposes:
..code-block:: python
def vf_preds_and_logits_fetches(policy):
return {
SampleBatch.VF_PREDS: policy.value_function,
BEHAVIOUR_LOGITS: policy.model.outputs,
}
``gradients_fn``: If defined, this function returns TF gradients for the loss function. You'd typically only want to override this to apply transformations such as gradient clipping:
``mixins``: To add arbitrary stateful components, you can add mixin classes to the policy. Methods defined by these mixins will have higher priority than the base policy class, so you can use these to override methods (as in the case of ``LearningRateSchedule``), or define extra methods and attributes (e.g., ``KLCoeffMixin``, ``ValueNetworkMixin``). Like any other Python superclass, these should be initialized at some point, which is what the ``setup_mixins`` function does:
In PPO we run ``setup_mixins`` before the loss function is called (i.e., ``before_loss_init``), but other callbacks you can use include ``before_init`` and ``after_init``.
**Example 2: Deep Q Networks**
(todo)
Finally, note that you do not have to use ``build_tf_policy`` to define a TensorFlow policy. You can alternatively subclass ``Policy``, ``TFPolicy``, or ``DynamicTFPolicy`` as convenient.
While RLlib runs all TF operations in graph mode, you can still leverage TensorFlow eager using `tf.py_function <https://www.tensorflow.org/api_docs/python/tf/py_function>`__. However, note that eager and non-eager tensors cannot be mixed within the ``py_function``. Here's an example of embedding eager execution within a policy loss function:
..code-block:: python
def eager_loss(policy, batch_tensors):
"""Example of using embedded eager execution in a custom loss.
Here `compute_penalty` prints the actions and rewards for debugging, and
also computes a (dummy) penalty term to add to the loss.
You can find a runnable file for the above eager execution example `here <https://github.com/ray-project/ray/blob/master/python/ray/rllib/examples/eager_execution.py>`__.
There is also experimental support for running the entire loss function in eager mode. This can be enabled with ``use_eager: True``, e.g., ``rllib train --env=CartPole-v0 --run=PPO --config='{"use_eager": true, "simple_optimizer": true}'``. However this currently only works for a couple algorithms.
Defining a policy in PyTorch is quite similar to that for TensorFlow (and the process of defining a trainer given a Torch policy is exactly the same). Here's a simple example of a trivial torch policy `(runnable file here) <https://github.com/ray-project/ray/blob/master/python/ray/rllib/examples/custom_torch_policy.py>`__:
..code-block:: python
from ray.rllib.policy.sample_batch import SampleBatch
from ray.rllib.policy.torch_policy_template import build_torch_policy
Now, building on the TF examples above, let's look at how the `A3C torch policy <https://github.com/ray-project/ray/blob/master/python/ray/rllib/agents/a3c/a3c_torch_policy.py>`__ is defined:
``loss_fn``: Similar to the TF example, the actor critic loss is defined over ``batch_tensors``. We imperatively execute the forward pass by calling ``policy.model()`` on the observations followed by ``policy.dist_class()`` on the output logits. The output Tensors are saved as attributes of the policy object (e.g., ``policy.entropy = dist.entropy.mean()``), and we return the scalar loss:
``stats_fn``: The stats function references ``entropy``, ``pi_err``, and ``value_err`` saved from the call to the loss function, similar in the PPO TF example:
``extra_action_out_fn``: We save value function predictions given model outputs. This makes the value function predictions of the model available in the trajectory as ``batch_tensors[SampleBatch.VF_PREDS]``:
``postprocess_fn`` and ``mixins``: Similar to the PPO example, we need access to the value function during postprocessing (i.e., ``add_advantages`` below calls ``policy._value()``. The value function is exposed through a mixin class that defines the method:
You can find the full policy definition in `a3c_torch_policy.py <https://github.com/ray-project/ray/blob/master/python/ray/rllib/agents/a3c/a3c_torch_policy.py>`__.
In summary, the main differences between the PyTorch and TensorFlow policy builder functions is that the TF loss and stats functions are built symbolically when the policy is initialized, whereas for PyTorch these functions are called imperatively each time they are used.
Given an environment and policy, policy evaluation produces `batches <https://github.com/ray-project/ray/blob/master/python/ray/rllib/policy/sample_batch.py>`__ of experiences. This is your classic "environment interaction loop". Efficient policy evaluation can be burdensome to get right, especially when leveraging vectorization, RNNs, or when operating in a multi-agent environment. RLlib provides a `RolloutWorker <https://github.com/ray-project/ray/blob/master/python/ray/rllib/evaluation/rollout_worker.py>`__ class that manages all of this, and this class is used in most RLlib algorithms.
You can use rollout workers standalone to produce batches of experiences. This can be done by calling ``worker.sample()`` on a worker instance, or ``worker.sample.remote()`` in parallel on worker instances created as Ray actors (see `WorkerSet <https://github.com/ray-project/ray/blob/master/python/ray/rllib/evaluation/worker_set.py>`__).
Here is an example of creating a set of rollout workers and using them gather experiences in parallel. The trajectories are concatenated, the policy learns on the trajectory batch, and then we broadcast the policy weights to the workers for the next round of rollouts:
Similar to how a `gradient-descent optimizer <https://www.tensorflow.org/api_docs/python/tf/train/GradientDescentOptimizer>`__ can be used to improve a model, RLlib's `policy optimizers <https://github.com/ray-project/ray/tree/master/python/ray/rllib/optimizers>`__ implement different strategies for improving a policy.
For example, in A3C you'd want to compute gradients asynchronously on different workers, and apply them to a central policy replica. This strategy is implemented by the `AsyncGradientsOptimizer <https://github.com/ray-project/ray/blob/master/python/ray/rllib/optimizers/async_gradients_optimizer.py>`__. Another alternative is to gather experiences synchronously in parallel and optimize the model centrally, as in `SyncSamplesOptimizer <https://github.com/ray-project/ray/blob/master/python/ray/rllib/optimizers/sync_samples_optimizer.py>`__. Policy optimizers abstract these strategies away into reusable modules.
Trainers are the boilerplate classes that put the above components together, making algorithms accessible via Python API and the command line. They manage algorithm configuration, setup of the rollout workers and optimizer, and collection of training metrics. Trainers also implement the `Trainable API <https://ray.readthedocs.io/en/latest/tune-usage.html#training-api>`__ for easy experiment management.
