Working With Offline Data

Getting started

RLlib’s offline dataset APIs enable working with experiences read from offline storage (e.g., disk, cloud storage, streaming systems, HDFS). For example, you might want to read experiences saved from previous training runs, or gathered from policies deployed in web applications. You can also log new agent experiences produced during online training for future use.

RLlib represents trajectory sequences (i.e., (s, a, r, s', ...) tuples) with SampleBatch objects. Using a batch format enables efficient encoding and compression of experiences. During online training, RLlib uses policy evaluation actors to generate batches of experiences in parallel using the current policy. RLlib also uses this same batch format for reading and writing experiences to offline storage.

Example: Training on previously saved experiences


For custom models and enviroments, you’ll need to use the Python API.

In this example, we will save batches of experiences generated during online training to disk, and then leverage this saved data to train a policy offline using DQN. First, we run a simple policy gradient algorithm for 100k steps with "output": "/tmp/cartpole-out" to tell RLlib to write simulation outputs to the /tmp/cartpole-out directory.

$ rllib train
    --run=PG \
    --env=CartPole-v0 \
    --config='{"output": "/tmp/cartpole-out", "output_max_file_size": 5000000}' \
    --stop='{"timesteps_total": 100000}'

The experiences will be saved in compressed JSON batch format:

$ ls -l /tmp/cartpole-out
total 11636
-rw-rw-r-- 1 eric eric 5022257 output-2019-01-01_15-58-57_worker-0_0.json
-rw-rw-r-- 1 eric eric 5002416 output-2019-01-01_15-59-22_worker-0_1.json
-rw-rw-r-- 1 eric eric 1881666 output-2019-01-01_15-59-47_worker-0_2.json

Then, we can tell DQN to train using these previously generated experiences with "input": "/tmp/cartpole-out". We disable exploration since it has no effect on the input:

$ rllib train \
    --run=DQN \
    --env=CartPole-v0 \
        "input": "/tmp/cartpole-out",
        "input_evaluation": [],
        "explore": false}'

Off-policy estimation: Since the input experiences are not from running simulations, RLlib cannot report the true policy performance during training. However, you can use tensorboard --logdir=~/ray_results to monitor training progress via other metrics such as estimated Q-value. Alternatively, off-policy estimation can be used, which requires both the source and target action probabilities to be available (i.e., the action_prob batch key). For DQN, this means enabling soft Q learning so that actions are sampled from a probability distribution:

$ rllib train \
    --run=DQN \
    --env=CartPole-v0 \
        "input": "/tmp/cartpole-out",
        "input_evaluation": ["is", "wis"],
        "exploration_config": {
            "type": "SoftQ",
            "temperature": 1.0,

This example plot shows the Q-value metric in addition to importance sampling (IS) and weighted importance sampling (WIS) gain estimates (>1.0 means there is an estimated improvement over the original policy):


Estimator Python API: For greater control over the evaluation process, you can create off-policy estimators in your Python code and call estimator.estimate(episode_batch) to perform counterfactual estimation as needed. The estimators take in a policy object and gamma value for the environment:

trainer = DQNTrainer(...)
...  # train policy offline

from ray.rllib.offline.json_reader import JsonReader
from ray.rllib.offline.wis_estimator import WeightedImportanceSamplingEstimator

estimator = WeightedImportanceSamplingEstimator(trainer.get_policy(), gamma=0.99)
reader = JsonReader("/path/to/data")
for _ in range(1000):
    batch =
    for episode in batch.split_by_episode():

Simulation-based estimation: If true simulation is also possible (i.e., your env supports step()), you can also set "input_evaluation": ["simulation"] to tell RLlib to run background simulations to estimate current policy performance. The output of these simulations will not be used for learning. Note that in all cases you still need to specify an environment object to define the action and observation spaces. However, you don’t need to implement functions like reset() and step().

Example: Converting external experiences to batch format

When the env does not support simulation (e.g., it is a web application), it is necessary to generate the *.json experience batch files outside of RLlib. This can be done by using the JsonWriter class to write out batches. This runnable example shows how to generate and save experience batches for CartPole-v0 to disk:

import gym
import numpy as np
import os

import ray._private.utils

from ray.rllib.models.preprocessors import get_preprocessor
from ray.rllib.evaluation.sample_batch_builder import SampleBatchBuilder
from ray.rllib.offline.json_writer import JsonWriter

if __name__ == "__main__":
    batch_builder = SampleBatchBuilder()  # or MultiAgentSampleBatchBuilder
    writer = JsonWriter(
        os.path.join(ray._private.utils.get_user_temp_dir(), "demo-out")

    # You normally wouldn't want to manually create sample batches if a
    # simulator is available, but let's do it anyways for example purposes:
    env = gym.make("CartPole-v0")

    # RLlib uses preprocessors to implement transforms such as one-hot encoding
    # and flattening of tuple and dict observations. For CartPole a no-op
    # preprocessor is used, but this may be relevant for more complex envs.
    prep = get_preprocessor(env.observation_space)(env.observation_space)
    print("The preprocessor is", prep)

    for eps_id in range(100):
        obs = env.reset()
        prev_action = np.zeros_like(env.action_space.sample())
        prev_reward = 0
        done = False
        t = 0
        while not done:
            action = env.action_space.sample()
            new_obs, rew, done, info = env.step(action)
                action_prob=1.0,  # put the true action probability here
            obs = new_obs
            prev_action = action
            prev_reward = rew
            t += 1

On-policy algorithms and experience postprocessing

RLlib assumes that input batches are of postprocessed experiences. This isn’t typically critical for off-policy algorithms (e.g., DQN’s post-processing is only needed if n_step > 1 or worker_side_prioritization: True). For off-policy algorithms, you can also safely set the postprocess_inputs: True config to auto-postprocess data.

However, for on-policy algorithms like PPO, you’ll need to pass in the extra values added during policy evaluation and postprocessing to batch_builder.add_values(), e.g., logits, vf_preds, value_target, and advantages for PPO. This is needed since the calculation of these values depends on the parameters of the behaviour policy, which RLlib does not have access to in the offline setting (in online training, these values are automatically added during policy evaluation).

Note that for on-policy algorithms, you’ll also have to throw away experiences generated by prior versions of the policy. This greatly reduces sample efficiency, which is typically undesirable for offline training, but can make sense for certain applications.

Mixing simulation and offline data

RLlib supports multiplexing inputs from multiple input sources, including simulation. For example, in the following example we read 40% of our experiences from /tmp/cartpole-out, 30% from hdfs:/archive/cartpole, and the last 30% is produced via policy evaluation. Input sources are multiplexed using np.random.choice:

$ rllib train \
    --run=DQN \
    --env=CartPole-v0 \
        "input": {
            "/tmp/cartpole-out": 0.4,
            "hdfs:/archive/cartpole": 0.3,
            "sampler": 0.3,
        "explore": false}'

Scaling I/O throughput

Similar to scaling online training, you can scale offline I/O throughput by increasing the number of RLlib workers via the num_workers config. Each worker accesses offline storage independently in parallel, for linear scaling of I/O throughput. Within each read worker, files are chosen in random order for reads, but file contents are read sequentially.

Ray Dataset Integration

RLlib has experimental support for reading/writing training samples from/to large offline datasets using Ray Dataset. We support JSON and Parquet files today. Other file formats supported by Dataset can also be easily added.

Unlike JSON input, a single dataset can be automatically sharded and replayed by multiple rollout workers by simply specifying the desired num_workers config.

To load sample data using Dataset, specify input and input_config keys like the following:

config = {
        "format": "json",  # json or parquet
        # Path to data file or directory.
        "path": "/path/to/json_dir/",
        # Num of tasks reading dataset in parallel, default is num_workers.
        "parallelism": 3,
        # Dataset allocates 0.5 CPU for each reader by default.
        # Adjust this value based on the size of your offline dataset.
        "num_cpus_per_read_task": 0.5,

To write sample data to JSON or Parquet files using Dataset, specify output and output_config keys like the following:

config = {
    "output": "dataset",
    "output_config": {
        "format": "json",  # json or parquet
        # Directory to write data files.
        "path": "/tmp/test_samples/",
        # Break samples into multiple files, each containing about this many records.
        "max_num_samples_per_file": 100000,

Writing Environment Data

To include environment data in the training sample datasets you can use the optional store_infos parameter that is part of the output_config dictionary. This parameter ensures that the infos dictionary, as returned by the RL environment, is included in the output files.

Note 1: It is the responsibility of the user to ensure that the content of infos can be serialized to file. Note 2: This setting is only relevant for the TensorFlow based agents, for PyTorch agents the infos data is always stored.

To write the infos data to JSON or Parquet files using Dataset, specify output and output_config keys like the following:

config = {
    "output": "dataset",
    "output_config": {
        "format": "json",  # json or parquet
        # Directory to write data files.
        "path": "/tmp/test_samples/",
        # Write the infos dict data
        "store_infos" : True,

Input Pipeline for Supervised Losses

You can also define supervised model losses over offline data. This requires defining a custom model loss. We provide a convenience function, InputReader.tf_input_ops(), that can be used to convert any input reader to a TF input pipeline. For example:

def custom_loss(self, policy_loss):
    input_reader = JsonReader("/tmp/cartpole-out")
    # print(  # if you want to access imperatively

    input_ops = input_reader.tf_input_ops()
    print(input_ops["obs"])  # -> output Tensor shape=[None, 4]
    print(input_ops["actions"])  # -> output Tensor shape=[None]

    supervised_loss = some_function_of(input_ops)
    return policy_loss + supervised_loss

See for a runnable example of using these TF input ops in a custom loss.

Input API

You can configure experience input for an agent using the following options:

    # Specify how to generate experiences:
    #  - "sampler": Generate experiences via online (env) simulation (default).
    #  - A local directory or file glob expression (e.g., "/tmp/*.json").
    #  - A list of individual file paths/URIs (e.g., ["/tmp/1.json",
    #    "s3://bucket/2.json"]).
    #  - A dict with string keys and sampling probabilities as values (e.g.,
    #    {"sampler": 0.4, "/tmp/*.json": 0.4, "s3://bucket/expert.json": 0.2}).
    #  - A callable that takes an `IOContext` object as only arg and returns a
    #    ray.rllib.offline.InputReader.
    #  - A string key that indexes a callable with tune.registry.register_input
    "input": "sampler",
    # Arguments accessible from the IOContext for configuring custom input
    "input_config": {},
    # True, if the actions in a given offline "input" are already normalized
    # (between -1.0 and 1.0). This is usually the case when the offline
    # file has been generated by another RLlib algorithm (e.g. PPO or SAC),
    # while "normalize_actions" was set to True.
    "actions_in_input_normalized": False,
    # Specify how to evaluate the current policy. This only has an effect when
    # reading offline experiences ("input" is not "sampler").
    # Available options:
    #  - "wis": the weighted step-wise importance sampling estimator.
    #  - "is": the step-wise importance sampling estimator.
    #  - "simulation": run the environment in the background, but use
    #    this data for evaluation only and not for learning.
    "input_evaluation": ["is", "wis"],
    # Whether to run postprocess_trajectory() on the trajectory fragments from
    # offline inputs. Note that postprocessing will be done using the *current*
    # policy, not the *behavior* policy, which is typically undesirable for
    # on-policy algorithms.
    "postprocess_inputs": False,
    # If positive, input batches will be shuffled via a sliding window buffer
    # of this number of batches. Use this if the input data is not in random
    # enough order. Input is delayed until the shuffle buffer is filled.
    "shuffle_buffer_size": 0,

The interface for a custom input reader is as follows:

class ray.rllib.offline.InputReader[source]

API for collecting and returning experiences during policy evaluation.

abstract next() Union[SampleBatch, MultiAgentBatch][source]

Returns the next batch of read experiences.


The experience read (SampleBatch or MultiAgentBatch).

tf_input_ops(queue_size: int = 1) Dict[str, Any][source]

Returns TensorFlow queue ops for reading inputs from this reader.

The main use of these ops is for integration into custom model losses. For example, you can use tf_input_ops() to read from files of external experiences to add an imitation learning loss to your model.

This method creates a queue runner thread that will call next() on this reader repeatedly to feed the TensorFlow queue.


queue_size – Max elements to allow in the TF queue.


>>> from ray.rllib.models.modelv2 import ModelV2
>>> from ray.rllib.offline.json_reader import JsonReader
>>> imitation_loss = ... # doctest +SKIP
>>> class MyModel(ModelV2): # doctest +SKIP
...     def custom_loss(self, policy_loss, loss_inputs):
...         reader = JsonReader(...)
...         input_ops = reader.tf_input_ops()
...         logits, _ = self._build_layers_v2(
...             {"obs": input_ops["obs"]},
...             self.num_outputs, self.options)
...         il_loss = imitation_loss(logits, input_ops["action"])
...         return policy_loss + il_loss

You can find a runnable version of this in examples/


Dict of Tensors, one for each column of the read SampleBatch.

Example Custom Input API

You can create a custom input reader like the following:

from ray.rllib.offline import InputReader, IOContext, ShuffledInput
from ray.tune.registry import register_input

class CustomInputReader(InputReader):
    def __init__(self, ioctx: IOContext): ...
    def next(self): ...

def input_creator(ioctx: IOContext) -> InputReader:
    return ShuffledInput(CustomInputReader(ioctx))

register_input("custom_input", input_creator)

config = {
    "input": "custom_input",
    "input_config": {},

You can pass arguments from the config to the custom input api through the input_config option which can be accessed with the IOContext. The interface for the IOContext is the following:

class ray.rllib.offline.IOContext(log_dir: Optional[str] = None, config: Optional[dict] = None, worker_index: int = 0, worker: Optional[RolloutWorker] = None)[source]

Class containing attributes to pass to input/output class constructors.

RLlib auto-sets these attributes when constructing input/output classes, such as InputReaders and OutputWriters.

default_sampler_input() Optional[SamplerInput][source]

Returns the RolloutWorker’s SamplerInput object, if any.

Returns None if the RolloutWorker has no SamplerInput. Note that local workers in case there are also one or more remote workers by default do not create a SamplerInput object.


The RolloutWorkers’ SamplerInput object or None if none exists.

See for a runnable example.

Output API

You can configure experience output for an agent using the following options:

    # Specify where experiences should be saved:
    #  - None: don't save any experiences
    #  - "logdir" to save to the agent log dir
    #  - a path/URI to save to a custom output directory (e.g., "s3://bucket/")
    #  - a function that returns a rllib.offline.OutputWriter
    "output": None,
    # Arguments accessible from the IOContext for configuring custom output
    "output_config": {},
    # What sample batch columns to LZ4 compress in the output data.
    "output_compress_columns": ["obs", "new_obs"],
    # Max output file size (in bytes) before rolling over to a new file.
    "output_max_file_size": 64 * 1024 * 1024,

The interface for a custom output writer is as follows:

class ray.rllib.offline.OutputWriter[source]

Writer API for saving experiences from policy evaluation.

write(sample_batch: Union[SampleBatch, MultiAgentBatch])[source]

Saves a batch of experiences.


sample_batch – SampleBatch or MultiAgentBatch to save.

Production RL Summit - March 29

Join us at Production RL Summit — a free virtual reinforcement learning event showcasing how companies like JP Morgan, Riot Games, Dow, and Siemens are apply RL to real business problems. Connect with peers and experts from the RL community, and sharpen your RL skills with hands-on workshop.

Register Now