RLlib Concepts and Custom Algorithms

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.

Most interaction with deep learning frameworks is isolated to the Policy interface, allowing RLlib to support multiple frameworks. To simplify the definition of policies, RLlib includes Tensorflow and PyTorch-specific templates. You can also write your own from scratch. Here is an example:

class CustomPolicy(Policy):
    """Example of a custom policy written from scratch.

    You might find it more convenient to use the `build_tf_policy` and
    `build_torch_policy` helpers instead for a real policy, which are
    described in the next sections.

    def __init__(self, observation_space, action_space, config):
        Policy.__init__(self, observation_space, action_space, config)
        # example parameter
        self.w = 1.0

    def compute_actions(self,
        # return action batch, RNN states, extra values to include in batch
        return [self.action_space.sample() for _ in obs_batch], [], {}

    def learn_on_batch(self, samples):
        # implement your learning code here
        return {}  # return stats

    def get_weights(self):
        return {"w": self.w}

    def set_weights(self, weights):
        self.w = weights["w"]

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:

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"]

def compute_actions(self,
    assert len(state_batches) == len(self.get_initial_state())
    new_state_batches = [[
        t + 1 for t in state_batches[0]
    return ..., new_state_batches, {}

def learn_on_batch(self, samples):
    # 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):

def compute_actions(self,
    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()

Policies in Multi-Agent

Beyond being agnostic of framework implementation, one of the main reasons to have a Policy abstraction is for use in multi-agent environments. For example, the rock-paper-scissors example shows how you can leverage the Policy abstraction to evaluate heuristic policies against learned policies.

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:

import tensorflow as tf
from ray.rllib.policy.sample_batch import SampleBatch

def policy_gradient_loss(policy, model, dist_class, train_batch):
    actions = train_batch[SampleBatch.ACTIONS]
    rewards = train_batch[SampleBatch.REWARDS]
    logits, _ = model.from_batch(train_batch)
    action_dist = dist_class(logits, model)
    return -tf.reduce_mean(action_dist.logp(actions) * rewards)

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 action_dist object is an ActionDistribution that is parameterized by 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:

from ray.rllib.policy.tf_policy_template import build_tf_policy

# <class 'ray.rllib.policy.tf_policy_template.MyTFPolicy'>
MyTFPolicy = build_tf_policy(

We can create a Trainer and try running this policy on a toy env with two parallel rollout workers:

import ray
from ray import tune
from ray.rllib.agents.trainer_template import build_trainer

# <class 'ray.rllib.agents.trainer_template.MyCustomTrainer'>
MyTrainer = build_trainer(

tune.run(MyTrainer, config={"env": "CartPole-v0", "num_workers": 2})

If you run the above snippet (runnable file here), you’ll probably notice that CartPole doesn’t learn so well:

== Status ==
Using FIFO scheduling algorithm.
Resources requested: 3/4 CPUs, 0/0 GPUs
Memory usage on this node: 4.6/12.3 GB
Result logdir: /home/ubuntu/ray_results/MyAlgTrainer
Number of trials: 1 ({'RUNNING': 1})
RUNNING trials:
 - MyAlgTrainer_CartPole-v0_0:      RUNNING, [3 CPUs, 0 GPUs], [pid=26784],
                                    32 s, 156 iter, 62400 ts, 23.1 rew

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:

from ray.rllib.evaluation.postprocessing import compute_advantages, \

def postprocess_advantages(policy,
    return compute_advantages(
        sample_batch, 0.0, policy.config["gamma"], use_gae=False)

def policy_gradient_loss(policy, model, dist_class, train_batch):
    logits, _ = model.from_batch(train_batch)
    action_dist = dist_class(logits, model)
    return -tf.reduce_mean(
        action_dist.logp(train_batch[SampleBatch.ACTIONS]) *

MyTFPolicy = build_tf_policy(

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 train_batch[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:

PPOTrainer = build_trainer(

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 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):

def choose_policy_optimizer(workers, config):
    if config["simple_optimizer"]:
        return SyncSamplesOptimizer(

    return LocalMultiGPUOptimizer(

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 override the make_policy_optimizer component of PPOTrainer.

from ray.rllib.agents.ppo import PPOTrainer
from ray.rllib.optimizers import AsyncGradientsOptimizer

def make_async_optimizer(workers, config):
    return AsyncGradientsOptimizer(workers, grads_per_step=100)

CustomTrainer = PPOTrainer.with_updates(

The with_updates method that we use here is also available for Torch and TF policies built from templates.

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):

def update_kl(trainer, fetches):
    if "kl" in fetches:
        # single-agent
            lambda pi: pi.update_kl(fetches["kl"]))

        def update(pi, pi_id):
            if pi_id in fetches:
                logger.debug("No data for {}, not updating kl".format(pi_id))

        # multi-agent

The update_kl method on the policy is defined in PPOTFPolicy via the KLCoeffMixin, along with several other advanced features. Let’s look at each new feature used by the policy:

PPOTFPolicy = build_tf_policy(
    get_default_config=lambda: ray.rllib.agents.ppo.ppo.DEFAULT_CONFIG,
    mixins=[LearningRateSchedule, KLCoeffMixin, ValueNetworkMixin])

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:

def kl_and_loss_stats(policy, train_batch):
    policy.explained_variance = explained_variance(
        train_batch[Postprocessing.VALUE_TARGETS], policy.model.value_function())

    stats_fetches = {
        "cur_kl_coeff": policy.kl_coeff,
        "cur_lr": tf.cast(policy.cur_lr, tf.float64),
        "total_loss": policy.loss_obj.loss,
        "policy_loss": policy.loss_obj.mean_policy_loss,
        "vf_loss": policy.loss_obj.mean_vf_loss,
        "vf_explained_var": policy.explained_variance,
        "kl": policy.loss_obj.mean_kl,
        "entropy": policy.loss_obj.mean_entropy,

    return stats_fetches

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[BEHAVIOUR_LOGITS]. Other values such as the current value prediction can also be emitted for debugging or optimization purposes:

def vf_preds_and_logits_fetches(policy):
    return {
        SampleBatch.VF_PREDS: policy.model.value_function(),
        BEHAVIOUR_LOGITS: policy.model.last_output(),

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:

def clip_gradients(policy, optimizer, loss):
    if policy.config["grad_clip"] is not None:
        grads = tf.gradients(loss, policy.model.trainable_variables())
        policy.grads, _ = tf.clip_by_global_norm(grads,
        clipped_grads = list(zip(policy.grads, policy.model.trainable_variables()))
        return clipped_grads
        return optimizer.compute_gradients(
            loss, colocate_gradients_with_ops=True)

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:

def setup_mixins(policy, obs_space, action_space, config):
    ValueNetworkMixin.__init__(policy, obs_space, action_space, config)
    KLCoeffMixin.__init__(policy, config)
    LearningRateSchedule.__init__(policy, config["lr"], config["lr_schedule"])

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

Let’s look at how to implement a different family of policies, by looking at the SimpleQ policy definition:

SimpleQPolicy = build_tf_policy(
    get_default_config=lambda: ray.rllib.agents.dqn.dqn.DEFAULT_CONFIG,
    extra_action_fetches_fn=lambda policy: {"q_values": policy.q_values},
    extra_learn_fetches_fn=lambda policy: {"td_error": policy.td_error},

The biggest difference from the policy gradient policies you saw previously is that SimpleQPolicy defines its own make_model and action_sampler_fn. This means that the policy builder will not internally create a model and action distribution, rather it will call build_q_models and build_action_sampler to get the output action tensors.

The model creation function actually creates two different models for DQN: the base Q network, and also a target network. It requires each model to be of type SimpleQModel, which implements a get_q_values() method. The model catalog will raise an error if you try to use a custom ModelV2 model that isn’t a subclass of SimpleQModel. Similarly, the full DQN policy requires models to subclass DistributionalQModel, which implements get_q_value_distributions() and get_state_value():

def build_q_models(policy, obs_space, action_space, config):

    policy.q_model = ModelCatalog.get_model_v2(

    policy.target_q_model = ModelCatalog.get_model_v2(

    return policy.q_model

The action sampler is straightforward, it just takes the q_model, runs a forward pass, and returns the argmax over the actions:

def build_action_sampler(policy, q_model, input_dict, obs_space, action_space,
    # do max over Q values...
    return action, action_logp

The remainder of DQN is similar to other algorithms. Target updates are handled by a after_optimizer_step callback that periodically copies the weights of the Q network to the target.

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.

Building Policies in TensorFlow Eager

Policies built with build_tf_policy (most of the reference algorithms are) can be run in eager mode by setting the "eager": True / "eager_tracing": True config options or using rllib train --eager [--trace]. This will tell RLlib to execute the model forward pass, action distribution, loss, and stats functions in eager mode.

Eager mode makes debugging much easier, since you can now use line-by-line debugging with breakpoints or Python print() to inspect intermediate tensor values. However, eager can be slower than graph mode unless tracing is enabled.

You can also selectively leverage eager operations within graph mode execution with tf.py_function. Here’s an example of using eager ops embedded within a loss function.

Building Policies in PyTorch

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):

from ray.rllib.policy.sample_batch import SampleBatch
from ray.rllib.policy.torch_policy_template import build_torch_policy

def policy_gradient_loss(policy, model, dist_class, train_batch):
    logits, _ = model.from_batch(train_batch)
    action_dist = dist_class(logits)
    log_probs = action_dist.logp(train_batch[SampleBatch.ACTIONS])
    return -train_batch[SampleBatch.REWARDS].dot(log_probs)

# <class 'ray.rllib.policy.torch_policy_template.MyTorchPolicy'>
MyTorchPolicy = build_torch_policy(

Now, building on the TF examples above, let’s look at how the A3C torch policy is defined:

A3CTorchPolicy = build_torch_policy(
    get_default_config=lambda: ray.rllib.agents.a3c.a3c.DEFAULT_CONFIG,

loss_fn: Similar to the TF example, the actor critic loss is defined over batch. We imperatively execute the forward pass by calling model() on the observations followed by 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:

def actor_critic_loss(policy, model, dist_class, train_batch):
    logits, _ = model.from_batch(train_batch)
    values = model.value_function()
    action_dist = dist_class(logits)
    log_probs = action_dist.logp(train_batch[SampleBatch.ACTIONS])
    policy.entropy = action_dist.entropy().mean()
    return overall_err

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:

def loss_and_entropy_stats(policy, train_batch):
    return {
        "policy_entropy": policy.entropy.item(),
        "policy_loss": policy.pi_err.item(),
        "vf_loss": policy.value_err.item(),

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[SampleBatch.VF_PREDS]:

def model_value_predictions(policy, input_dict, state_batches, model):
    return {SampleBatch.VF_PREDS: model.value_function().cpu().numpy()}

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:

def add_advantages(policy,
    completed = sample_batch[SampleBatch.DONES][-1]
    if completed:
        last_r = 0.0
        last_r = policy._value(sample_batch[SampleBatch.NEXT_OBS][-1])
    return compute_advantages(sample_batch, last_r, policy.config["gamma"],

class ValueNetworkMixin(object):
    def _value(self, obs):
        with self.lock:
            obs = torch.from_numpy(obs).float().unsqueeze(0).to(self.device)
            _, _, vf, _ = self.model({"obs": obs}, [])
            return vf.detach().cpu().numpy().squeeze()

You can find the full policy definition in 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 (or TensorFlow Eager) these functions are called imperatively each time they are used.

Extending Existing Policies

You can use the with_updates method on Trainers and Policy objects built with make_* to create a copy of the object with some changes, for example:

from ray.rllib.agents.ppo import PPOTrainer
from ray.rllib.agents.ppo.ppo_tf_policy import PPOTFPolicy

CustomPolicy = PPOTFPolicy.with_updates(

CustomTrainer = PPOTrainer.with_updates(

Policy Evaluation

Given an environment and policy, policy evaluation produces batches 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 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).

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:

# Setup policy and rollout workers
env = gym.make("CartPole-v0")
policy = CustomPolicy(env.observation_space, env.action_space, {})
workers = WorkerSet(
    env_creator=lambda c: gym.make("CartPole-v0"),

while True:
    # Gather a batch of samples
    T1 = SampleBatch.concat_samples(
        ray.get([w.sample.remote() for w in workers.remote_workers()]))

    # Improve the policy using the T1 batch

    # Broadcast weights to the policy evaluation workers
    weights = ray.put({"default_policy": policy.get_weights()})
    for w in workers.remote_workers():

Policy Optimization

Similar to how a gradient-descent optimizer can be used to improve a model, RLlib’s policy 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. Another alternative is to gather experiences synchronously in parallel and optimize the model centrally, as in SyncSamplesOptimizer. Policy optimizers abstract these strategies away into reusable modules.

This is how the example in the previous section looks when written using a policy optimizer:

# Same setup as before
workers = WorkerSet(
    env_creator=lambda c: gym.make("CartPole-v0"),

# this optimizer implements the IMPALA architecture
optimizer = AsyncSamplesOptimizer(workers, train_batch_size=500)

while True:


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 for easy experiment management.

Example of three equivalent ways of interacting with the PPO trainer, all of which log results in ~/ray_results:

trainer = PPOTrainer(env="CartPole-v0", config={"train_batch_size": 4000})
while True:
rllib train --run=PPO --env=CartPole-v0 --config='{"train_batch_size": 4000}'
from ray import tune
tune.run(PPOTrainer, config={"env": "CartPole-v0", "train_batch_size": 4000})