RLgraph: Flexible computation graphs for deep reinforcement learning
Clone or download


PyPI version Python 3.5 License Documentation Status Build Status


Flexible computation graphs for deep reinforcement learning.

RLgraph is a framework to quickly prototype, define and execute reinforcement learning algorithms both in research and practice. RLgraph is different from most other libraries as it can support TensorFlow (or static graphs in general) or eager/define-by run execution (PyTorch) through a single component interface.

RLgraph exposes a well defined API for using agents, and offers a novel component concept for testing and assembly of machine learning models. By separating graph definition, compilation and execution, multiple distributed backends and device execution strategies can be accessed without modifying agent definitions. This means it is especially suited for a smooth transition from applied use case prototypes to large scale distributed training.

The current state of RLgraph in version 0.2.3 is alpha. The core engine is substantially complete and works for TensorFlow and PyTorch (1.0). Distributed execution on Ray is exemplified via Distributed Prioritized Experience Replay (Ape-X), which also supports multi-gpu mode and solves e.g. Atari-Pong in ~1 hour on a single-node. Algorithms like Ape-X or PPO can be used both with PyTorch and TensorFlow. Distributed TensorFlow can be tested via the IMPALA agent. Please create an issue to discuss improvements or contributions.

RLgraph currently implements the following algorithms:

  • DQN - dqn_agent - paper
  • Double-DQN - dqn_agent - via double_dqn flag - paper
  • Dueling-DQN - dqn_agent - via dueling_dqn flag - paper
  • Prioritized experience replay - via memory_spec option prioritized_replay - paper
  • Deep-Q learning from demonstration dqfd_agent - paper
  • Distributed prioritized experience replay (Ape-X) on Ray - via apex_executor - paper
  • Importance-weighted actor-learner architecture (IMPALA) on distributed TF/Multi-threaded single-node - impala_agents - paper
  • Proximal policy optimization with generalized advantage estimation - ppo_agent - paper
  • Simple actor-critic for REINFORCE/A2C/A3C actor_critic_agent - paper

The SingleThreadedWorker implements high-performance environment vectorisation, and a RayWorker can execute ray actor tasks in conjunction with a RayExecutor. The examples folder contains simple scripts to test these agents. There is also a very extensive test package including tests for virtually every component. Note that we run tests on TensorFlow and have not reached full coverage/test compatibility with PyTorch.

For more detailed documentation on RLgraph and its API-reference, please visit our readthedocs page here.


The simplest way to install RLgraph is from pip:

pip install rlgraph

Note that some backends (e.g. ray) need additional dependencies (see setup.py). For example, to install dependencies for the distributed backend ray, enter:

pip install rlgraph[ray]

To successfully run tests, please also install OpenAI gym, e.g.

pip install gym[all]

Upon calling RLgraph, a config JSON is created under ~.rlgraph/rlgraph.json which can be used to change backend settings. The current default stable backend is TensorFlow ("tf"). The PyTorch backend ("pytorch") does not support all utilities available in TF yet. Namely, device handling for PyTorch is incomplete, and we will likely wait until a stable PyTorch 1.0 release in the coming weeks.

Quickstart / example usage

We provide an example script for training the Ape-X algorithm on ALE using Ray in the examples folder.

First, you'll have to ensure, that Ray is used as the distributed backend. RLgraph checks the file ~/.rlgraph/rlgraph.json for this configuration. You can use this command to configure RLgraph to use TensorFlow as the backend and Ray as the distributed backend:

echo '{"BACKEND":"tf","DISTRIBUTED_BACKEND":"ray"}' > $HOME/.rlgraph/rlgraph.json

Then you can run our Ape-X example:

# Start ray on the head machine
ray start --head --redis-port 6379
# Optionally join to this cluster from other machines with ray start --redis-address=...

# Run script
python apex_pong.py

You can also train a simple DQN agent locally on OpenAI gym environments such as CartPole (this doesn't require Ray):

python dqn_cartpole.py

Import and use agents

Agents can be imported and used as follows:

from rlgraph.agents import DQNAgent
from rlgraph.environments import OpenAIGymEnv

environment = OpenAIGymEnv('CartPole-v0')

# Create from .json file or dict, see agent API for all
# possible configuration parameters.
agent = DQNAgent.from_file(

# Get an action, take a step, observe reward.
state = environment.reset()
preprocessed_state, action = agent.get_action(

# Execute step in environment.
next_state, reward, terminal, info =  environment.step(action)

# Observe result.

# Call update when desired:
loss = agent.update()

Full examples can be found in the examples folder.


If you use RLgraph in your research, please cite the following paper: link

    author = {{Schaarschmidt}, M. and {Mika}, S. and {Fricke}, K. and {Yoneki}, E.},
    title = "{RLgraph: Flexible Computation Graphs for Deep Reinforcement Learning}",
    journal = {ArXiv e-prints},
    archivePrefix = "arXiv",
    eprint = {1810.09028},
    year = 2018,
    month = oct