This document gives examples and pointers on how to experiment with and extend Dopamine.

You can find the documentation for each module in our codebase in our API documentation.

File organization

Dopamine is organized as follows:

Training agents

Atari games

The entry point to the standard Atari 2600 experiment is dopamine/discrete_domains/ 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 dopamine/agents/dqn/configs/dqn.gin, in particular by reducing Runner.training_steps and Runner.evaluation_steps, 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

Continuous control environments

The entry point for continuous control agents is dopamine/continuous_domains/ You will need a Mujoco key to run the following example. To run SAC on the HalfCheetah environment of Mujoco, run:

python -um dopamine.continuous_domains.train \
  --base_dir /tmp/dopamine_runs \
  --gin_files dopamine/jax/agents/sac/configs/sac.gin

By default, this will kick off an experiment lasting 3200 episodes, with 1000 environment steps per episode. The command-line interface will output statistics about the latest training episode:

I0908 17:19:39.618797 1803949] Starting iteration 0
I0908 17:19:40.592262 1803949] Average undiscounted return per training episode: -168.19
I0908 17:19:40.592391 1803949] Average training steps per second: 1027.80
I0908 17:19:45.699378 1803949] Average undiscounted return per evaluation episode: -279.07

To run with different environments/hyperparemeters, adjust the gin config file found here: dopamine/jax/agents/sac/configs/sac.gin. For your experiments, you may choose to supply a new gin config file, or override the existing config file with command line gin_bindings args.

For more information on using gin, see the gin github repo.

Configuring agents

The whole of Dopamine is easily configured using the gin configuration framework.

We provide a number of configuration files for each of the agents. The main configuration file for each agent corresponds to an “apples to apples” comparison, where hyperparameters have been selected to give a standardized performance comparison between agents. These are

More details on the exact choices behind these parameters are given in our baselines page.

We also provide configuration files corresponding to settings previously used in the literature. These are

All of these use the deterministic version of the Arcade Learning Environment (ALE), and slightly different hyperparameters.

Checkpointing and logging

Dopamine provides basic functionality for performing experiments. This functionality can be broken down into two main components: checkpointing and logging. Both components depend on the command-line parameter base_dir, which informs Dopamine of where it should store experimental data.


By default, Dopamine will save an experiment checkpoint every iteration: one training and one evaluation phase, following a standard set by Mnih et al. Checkpoints are saved in the checkpoints subdirectory under base_dir. At a high-level, the following are checkpointed:

If you’re curious, the checkpointing code itself is in dopamine/common/


At the end of each iteration, Dopamine also records the agent’s performance, both during training and (if enabled) during an optional evaluation phase. The log files are generated in dopamine/atari/ and more specifically in dopamine/common/, and are pickle files containing a dictionary mapping iteration keys (e.g., "iteration_47") to dictionaries containing data.

A simple way to read log data from multiple experiments is to use the provided read_experiment method in colab/

We provide a colab to illustrate how you can load the statistics from an experiment and plot them against our provided baseline runs.

Modifying and extending agents

Dopamine is designed to make algorithmic research simple. With this in mind, we decided to keep a relatively flat class hierarchy, with no abstract base class; we’ve found this sufficient for our research purposes, with the added benefits of simplicity and ease of use. To begin, we recommend modifying the agent code directly to suit your research purposes.

We provide a colab where we illustrate how one can extend the DQN agent, or create a new agent from scratch, and then plot the experimental results against our provided baselines.


The DQN agent is contained in two files:

The agent class defines the DQN network, the update rule, and also the basic operations of a RL agent (epsilon-greedy action selection, storing transitions, episode bookkeeping, etc.). For example, the Q-Learning update rule used in DQN is defined in two methods, _build_target_q_op and _build_train_op.

Rainbow and C51

The Rainbow agent is contained in two files:

The C51 agent is a specific parametrization of the Rainbow agent, where update_horizon (the n in n-step update) is set to 1 and a uniform replay scheme is used.

Implicit quantile networks (IQN)

The IQN agent is defined by one additional file:


We provide a series of files for all 4 agents on all 60 games. These are all *.tar.gz files which you will need to uncompress:

*  You can also view these with Tensorboard on your machine. For instance, after
   uncompressing the files you can run:

   tensorboard --logdir c51/Asterix/

   to display the training runs for C51 on Asterix: