Tensorforce: a TensorFlow library for applied reinforcement learning

Introduction

Tensorforce is an open-source deep reinforcement learning framework, with an emphasis on modularized flexible library design and straightforward usability for applications in research and practice. Tensorforce is built on top of Google's TensorFlow framework and compatible with Python 3 (Python 2 support was dropped with version 0.5).

Tensorforce follows a set of high-level design choices which differentiate it from other similar libraries:

Modular component-based design: Feature implementations, above all, strive to be as generally applicable and configurable as possible, potentially at some cost of faithfully resembling details of the introducing paper.
Separation of RL algorithm and application: Algorithms are agnostic to the type and structure of inputs (states/observations) and outputs (actions/decisions), as well as the interaction with the application environment.
Full-on TensorFlow models: The entire reinforcement learning logic, including control flow, is implemented in TensorFlow, to enable portable computation graphs independent of application programming language, and to facilitate the deployment of models.

Quicklinks

Documentation and update notes
[Benchmarks]((https://github.com/tensorforce/tensorforce/blob/master/benchmarks) and projects using Tensorforce
Contact and Gitter channel
Roadmap and contribution guidelines

Installation

A stable version of Tensorforce is periodically updated on PyPI and installed as follows:

pip install tensorforce

To always use the latest version of Tensorforce, install the GitHub version instead:

git clone https://github.com/tensorforce/tensorforce.git
cd tensorforce
pip install -e .

Tensorforce is built on top of Google's TensorFlow and requires that either tensorflow or tensorflow-gpu is installed, currently as version 1.13.1. To include the correct version of TensorFlow with the installation of Tensorforce, simply add the flag tf for the normal CPU version or tf_gpu for the GPU version:

# PyPI version plus TensorFlow CPU version
pip install tensorforce[tf]

# GitHub version plus TensorFlow GPU version
pip install -e .[tf_gpu]

Some environments require additional packages, for which there are also options available (mazeexp, gym, retro, vizdoom; or envs for all environments), however, some require other tools to be installed (see environments documentation).

Quickstart example code

from tensorforce.agents import Agent

# Instantiate a Tensorforce agent
agent = Agent.create(
    agent='tensorforce',
    states=dict(type='float', shape=(10,)),
    actions=dict(type='int', num_values=5),
    memory=10000,
    update=dict(unit='timesteps', batch_size=64),
    optimizer=dict(type='adam', learning_rate=3e-4),
    policy=dict(network='auto'),
    objective='policy_gradient',
    reward_estimation=dict(horizon=20)
)

# Initialize the agent
agent.initialize()

# Retrieve the latest (observable) environment state
state = get_current_state()  # (float array of shape [10])

# Query the agent for its action decision
action = agent.act(states=state)  # (scalar between 0 and 4)

# Execute the decision and retrieve the current performance score
reward = execute_decision(action)  # (any scalar float)

# Pass feedback about performance (and termination) to the agent
agent.observe(reward=reward, terminal=False)

Command line usage

Tensorforce comes with a range of example configurations for different popular reinforcement learning environments. For instance, to run Tensorforce's implementation of the popular Proximal Policy Optimization (PPO) algorithm on the OpenAI Gym CartPole environment, execute the following line:

python run.py benchmarks/configs/ppo.json gym --level CartPole-v1 -e 300

For more information check out the documentation.

Features

Neural network layers: Dense fully-connected layer, embedding layer, 1- and 2-dimensional convolution, pooling, LSTM, activation, dropout, normalization, and more; plus support of Keras layers.
Memory types: Simple batch buffer memory, random replay memory.
Policy distributions: Bernoulli distribution for boolean actions, categorical distribution for (finite) integer actions, Gaussian distribution for continuous actions, Beta distribution for range-constrained continuous actions.
Optimization algorithms: Various gradient-based optimizers provided by TensorFlow like Adam/AdaDelta/Momentum/RMSProp/etc, evolutionary optimizer, natural-gradient-based optimizer, plus a range of meta-optimizers.
Execution modes: Parallel execution, distributed execution.
Other features: state/reward preprocessing, exploration, variable noise, regularization losses.
TensorBoard support.

By combining these modular components in different ways, a variety of popular deep reinforcement learning models/features can be replicated: Deep Q-learning (DQN) and variations like Double-DQN or Deep Q-learning from Demonstrations (DQfD), vanilla policy-gradient algorithm / REINFORCE, Proximal Policy Optimization (PPO), Actor-critic and A3C, Trust Region Policy Optimization (TRPO), Normalised Advantage Function (NAF), Generalized Advantage estimation (GAE), etc.

Note that in general the replication is not 100% faithful, since the models as described in the corresponding paper often involve additional minor tweaks and modifications which are hard to support with a modular design (and, arguably, also questionable whether it is important/desirable to support them). On the upside, these models are just a few examples from the multitude of module combinations supported by Tensorforce.

Environment adapters

Arcade Learning Environment, a simple object-oriented framework that allows researchers and hobbyists to develop AI agents for Atari 2600 games.
MazeExplorer, a maze exploration game for AI agents.
OpenAI Gym, a toolkit for developing and comparing reinforcement learning algorithms which supports teaching agents everything from walking to playing games like Pong or Pinball.
OpenAI Retro, lets you turn classic video games into Gym environments for reinforcement learning and comes with integrations for ~1000 games.
OpenSim, reinforcement learning with musculoskeletal models.
PyGame Learning Environment, learning environment which allows a quick start to Reinforcement Learning in Python.
ViZDoom, allows developing AI bots that play Doom using only the visual information.

Contact for support and feedback

Please get in touch via mail or on Gitter if you have questions, feedback, ideas for features/collaboration, or if you seek support for applying Tensorforce to your problem.

Core team and contributors

Tensorforce is currently developed and maintained by Alexander Kuhnle. Earlier versions of Tensorforce (<= 0.4.2) were developed by Michael Schaarschmidt, Alexander Kuhnle and Kai Fricke.

We are very grateful for our open-source contributors (listed according to Github, updated periodically):

Islandman93, sven1977, Mazecreator, wassname, lefnire, daggertye, trickmeyer, mkempers, mryellow, ImpulseAdventure, janislavjankov, andrewekhalel, HassamSheikh, skervim, beflix, coord-e, benelot, tms1337, vwxyzjn, erniejunior, Deathn0t, petrbel, nrhodes, batu, yellowbee686, tgianko, AdamStelmaszczyk, BorisSchaeling, christianhidber, Davidnet, ekerazha, gitter-badger, kborozdin, Kismuz, mannsi, milesmcc, nagachika, neitzal, ngoodger, perara, sohakes, tomhennigan.

Cite Tensorforce

Please cite the framework as follows:

@misc{tensorforce,
  author       = {Kuhnle, Alexander and Schaarschmidt, Michael and Fricke, Kai},
  title        = {Tensorforce: a TensorFlow library for applied reinforcement learning},
  howpublished = {Web page},
  url          = {https://github.com/tensorforce/tensorforce},
  year         = {2017}
}

If you use Tensorforce in your research, you may additionally consider citing the following paper:

@article{lift-tensorforce,
  author       = {Schaarschmidt, Michael and Kuhnle, Alexander and Ellis, Ben and Fricke, Kai and Gessert, Felix and Yoneki, Eiko},
  title        = {{LIFT}: Reinforcement Learning in Computer Systems by Learning From Demonstrations},
  journal      = {CoRR},
  volume       = {abs/1808.07903},
  year         = {2018},
  url          = {http://arxiv.org/abs/1808.07903},
  archivePrefix = {arXiv},
  eprint       = {1808.07903}
}

rshtirmer/tensorforce