Github url


by google

google /dopamine

Dopamine is a research framework for fast prototyping of reinforcement learning algorithms.

9.1K Stars 1.2K Forks Last release: 11 months ago (v2) Apache License 2.0 197 Commits 2 Releases

Available items

No Items, yet!

The developer of this repository has not created any items for sale yet. Need a bug fixed? Help with integration? A different license? Create a request here:


Dopamine is a research framework for fast prototyping of reinforcement learning algorithms. It aims to fill the need for a small, easily grokked codebase in which users can freely experiment with wild ideas (speculative research).

Our design principles are:

  • Easy experimentation: Make it easy for new users to run benchmark experiments.
  • Flexible development: Make it easy for new users to try out research ideas.
  • Compact and reliable: Provide implementations for a few, battle-tested algorithms.
  • Reproducible: Facilitate reproducibility in results. In particular, our setup follows the recommendations given by Machado et al. (2018).

In the spirit of these principles, this first version focuses on supporting the state-of-the-art, single-GPU Rainbow agent (Hessel et al., 2018) applied to Atari 2600 game-playing (Bellemare et al., 2013). Specifically, our Rainbow agent implements the three components identified as most important by Hessel et al.:

For completeness, we also provide an implementation of DQN (Mnih et al., 2015). For additional details, please see our documentation.

We provide a set of Colaboratory notebooks which demonstrate how to use Dopamine.

This is not an official Google product.

What's new

  • 03/08/2020: Dopamine now supports JAX agents! This includes an implementation of the Quantile Regression agent (QR-DQN) which has been a common request. Find out more in our jax subdirectory, which includes trained agent checkpoints.

  • 27/07/2020: Dopamine now runs on TensorFlow 2. However, Dopamine is still written as TensorFlow 1.X code. This means your project may need to explicity disable TensorFlow 2 behaviours with:

  import tensorflow.compat.v1 as tf

if you are using custom entry-point for training your agent. The migration to TensorFlow 2 also means that Dopamine no longer supports Python 2.

  • 02/09/2019: Dopamine has switched its network definitions to use tf.keras.Model. The previous

    based networks are removed. If your agents inherit from dopamine agents you need to update your code.
    • ._get_network_type()
      functions are replaced with
      definitions are moved inside the model definition.
      # The following two functions are replaced with `_create_network()`.
      # def _get_network_type(self):
      #   return collections.namedtuple('DQN_network', ['q_values'])
      # def _network_template(self, state):
      #   return, self._get_network_type(), state)

    def _create_network(self, name): """Builds the convolutional network used to compute the agent's Q-values.

      name: str, this name is passed to the tf.keras.Model and used to create
        variable scope under the hood by the tf.keras.Model.
      network: tf.keras.Model, the network instantiated by the Keras model.
    # `` is set to `atari_lib.NatureDQNNetwork`.
    network =, name=name)
    return network

    def _build_networks(self): # The following two lines are replaced. # self.online_convnet = tf.make_template('Online', self._network_template) # self.target_convnet = tf.make_template('Target', self._network_template) self.online_convnet = self._create_network(name='Online') self.target_convnet = self._create_network(name='Target')

    • If your code overwrites
      make sure you update your code to fit with the new API. If your code overwrites
      you need to replace
      tf.make_template('Online', self._network_template)
    • The variables of each network can be obtained from the networks as follows:
      vars = self.online_convnet.variables
    • Baselines and older checkpoints can be loaded by adding the following line to your gin file.
     atari_lib.maybe_transform_variable_names.legacy_checkpoint_load = True
  • 11/06/2019: Visualization utilities added to generate videos and still images of a trained agent interacting with its environment. See an example colaboratory here.

  • 30/01/2019: Dopamine 2.0 now supports general discrete-domain gym environments.

  • 01/11/2018: Download links for each individual checkpoint, to avoid having to download all of the checkpoints.

  • 29/10/2018: Graph definitions now show up in Tensorboard.

  • 16/10/2018: Fixed a subtle bug in the IQN implementation and upated the colab tools, the JSON files, and all the downloadable data.

  • 18/09/2018: Added support for double-DQN style updates for the

    • Can be enabled via the
      constructor parameter.
  • 18/09/2018: Added support for reporting in-iteration losses directly from the agent to Tensorboard.

    • Set the
      run_experiment.create_agent.debug_mode = True
      via the configuration file or using the
      flag to enable it.
    • Control frequency of writes with the
      agent constructor parameter (defaults to
  • 27/08/2018: Dopamine launched!


Install via source

Installing from source allows you to modify the agents and experiments as you please, and is likely to be the pathway of choice for long-term use. The instructions below assume that you will be running Dopamine in a virtual environment. A virtual environment lets you control which dependencies are installed for which program.

Dopamine is a Tensorflow-based framework, and we recommend you also consult the Tensorflow documentation for additional details. Finally, these instructions are for Python 3.6 and above.

First download the Dopamine source.

git clone

Then create a virtual environment and activate it.

python3 -m venv ./dopamine-venv
source dopamine-venv/bin/activate

Finally setup the environment and install Dopamine's dependencies

pip install -U pip
pip install -r dopamine/requirements.txt

Running tests

You can test whether the installation was successful by running the following:

cd dopamine
python -m tests.dopamine.atari_init_test

Training agents

Atari games

The entry point to the standard Atari 2600 experiment is

. To run the basic DQN agent,

python -um dopamine.discrete_domains.train \
  --base_dir /tmp/dopamine_runs \
  --gin_files dopamine/agents/dqn/configs/dqn.gin

By default, this will kick off an experiment lasting 200 million frames. The command-line interface will output statistics about the latest training episode:

I0824 17:13:33.078342 140196395337472] gamma: 0.990000
I0824 17:13:33.795608 140196395337472] Beginning training...
Steps executed: 5903 Episode length: 1203 Return: -19.

To get finer-grained information about the process, you can adjust the experiment parameters in

, in particular by reducing

, which together determine the total number of steps needed to complete an iteration. This is useful if you want to inspect log files or checkpoints, which are generated at the end of each iteration.

More generally, the whole of Dopamine is easily configured using the gin configuration framework.

Non-Atari discrete environments

We provide sample configuration files for training an agent on Cartpole and Acrobot. For example, to train C51 on Cartpole with default settings, run the following command:

python -um dopamine.discrete_domains.train \
  --base_dir /tmp/dopamine_runs \
  --gin_files dopamine/agents/rainbow/configs/c51_cartpole.gin

You can train Rainbow on Acrobot with the following command:

python -um dopamine.discrete_domains.train \
  --base_dir /tmp/dopamine_runs \
  --gin_files dopamine/agents/rainbow/configs/rainbow_acrobot.gin

Install as a library

An easy, alternative way to install Dopamine is as a Python library:

pip install dopamine-rl


Bellemare et al., The Arcade Learning Environment: An evaluation platform for general agents. Journal of Artificial Intelligence Research, 2013.

Machado et al., Revisiting the Arcade Learning Environment: Evaluation Protocols and Open Problems for General Agents, Journal of Artificial Intelligence Research, 2018.

Hessel et al., Rainbow: Combining Improvements in Deep Reinforcement Learning. Proceedings of the AAAI Conference on Artificial Intelligence, 2018.

Mnih et al., Human-level Control through Deep Reinforcement Learning. Nature, 2015.

Mnih et al., Asynchronous Methods for Deep Reinforcement Learning. Proceedings of the International Conference on Machine Learning, 2016.

Schaul et al., Prioritized Experience Replay. Proceedings of the International Conference on Learning Representations, 2016.

Giving credit

If you use Dopamine in your work, we ask that you cite our white paper. Here is an example BibTeX entry:

  author    = {Pablo Samuel Castro and
               Subhodeep Moitra and
               Carles Gelada and
               Saurabh Kumar and
               Marc G. Bellemare},
  title     = {Dopamine: {A} {R}esearch {F}ramework for {D}eep {R}einforcement {L}earning},
  year      = {2018},
  url       = {},
  archivePrefix = {arXiv}

We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.