DQN Zoo

DQN Zoo is a collection of reference implementations of reinforcement learning agents developed at DeepMind based on the Deep Q-Network (DQN) agent.

It aims to be research-friendly, self-contained and readable. Each agent is implemented using JAX, Haiku and RLax, and is a best-effort replication of the corresponding paper implementation. Each agent reproduces results on the standard set of 57 Atari games, on average.

Plot of median human-normalized score over all 57 Atari games for each agent:

Quick start

NOTE: Only Python 3.9 and above and Linux is supported.

Follow these steps to quickly clone the DQN Zoo repository, install all required dependencies and start running DQN. Prerequisites for these steps are a NVIDIA GPU with recent CUDA drivers.

1. Install Docker version 19.03 or later (for the --gpus flag).

2. Install NVIDIA Container Toolkit.

3. Enable sudoless docker.

4. Verify the previous steps were successful by running:
   docker run --gpus all --rm nvidia/cuda:11.1-base nvidia-smi

5. Download the script run.sh. This automatically downloads the Atari ROMs from http://www.atarimania.com. The ROMs are available here for free but make sure the respective license covers your particular use case.

Running this script will:

1.  Clone the DQN Zoo repository.
1.  Build a Docker image with all necessary dependencies and run unit tests.
1.  Start a short run of DQN on Pong in a GPU-accelerated container.

NOTE: run.sh, Dockerfile and docker_requirements.txt together provide a self-contained example of the dependencies and commands needed to run an agent in DQN Zoo. Using Docker is not a requirement and if Dockerfile is not used then the list of dependencies to install may have to be adapted depending on your environment. Also it is not a hard requirement to run on the GPU. Agents can be run on the CPU by specifying the flag --jax_platform_name=cpu.

Goals

• Serve as a collection of reference implementations of DQN-based agents developed at DeepMind.
• Reproduce results reported in papers, on average.
• Implement agents purely in Python, using JAX, Haiku and RLax.
• Have minimal dependencies.
• Be easy to read.
• Be easy to modify and customize after forking.

Non-goals

• Be a library or framework (these agents are intended to be forked for research).
• Be flexible, general and support multiple use cases (at odds with understandability).
• Support many environments (users can easily add new ones).
• Include every DQN variant that exists.
• Incorporate many cool libraries (harder to read, easy for the user to do this after forking, different users prefer different libraries, less self-contained).
• Optimize speed and efficiency at the cost of readability or matching algorithmic details in the papers (no C++, keep to a single stream of experience).

Code structure

• Each directory contains a published DQN variant configured to run on Atari.
• agent.py in each agent directory contains an agent class that includes reset(), step(), get_state(), set_state() methods.
• parts.py contains functions and classes used by many of the agents including classes for accumulating statistics and the main training and evaluation loop run_loop().
• replay.py contains functions and classes relating to experience replay.
• networks.py contains Haiku networks used by the agents.
• processors.py contains components for standard Atari preprocessing.

Implementation notes

Generally we went with a flatter approach for easier code comprehension. Excessive nesting, indirection and generalization have been avoided, but not to the extreme of having a single file per agent. This has resulted in some degree of code duplication, but this is less of a maintenance issue as the code base is intended to be relatively static.

Some implementation details:

• The main training and evaluation loop parts.run_loop() is implemented as a generator to decouple it from other concerns like logging statistics and checkpointing.
• We adopted the pattern of returning a new JAX PRNG key from jitted functions. This allows for splitting keys inside jitted functions which is currently more efficient than splitting outside and passing a key in.
• Agent functions to be jitted are defined inline in the agent class __init__() instead of as decorated class methods. This emphasizes such functions should be free of side-effects; class methods are generally not pure as they often alter the class instance.
• parts.NullCheckpoint is a placeholder for users to optionally plug in a checkpointing library appropriate for the file system they are using. This would allow resuming an interrupted training run.
• The preprocessing and action repeat logic lives inside each agent. Doing this instead of taking the common approach of environment wrappers allows the run loop to see the "true" timesteps. This makes things like recording performance statistics and videos easier since the unmodified rewards and observations are readily available. It also allows us to express all relevant flag values in terms of environment frames, instead of a more confusing mix of environment frames and learning steps.

Learning curves

Learning curve data is included in results.tar.gz. The archive contains a CSV file for each agent, with statistics logged during training runs. These training runs span the standard set of 57 Atari games, 5 seeds each, using default agent settings. Note Gym was used instead of Xitari.

These CSV files can be theoretically equivalently generated by the following pseudocode:

for agent in "${AGENTS[@]}"; do
  for game in "${ATARI_GAMES[@]}"; do
    for seed in {1..5}; do
      python -m "dqn_zoo.${agent}.run_atari" \
          --environment_name="${game}" \
          --seed="${seed}" \
          --results_csv_path="/tmp/dqn_zoo/${agent}/${game}/${seed}/results.csv"
    done
  done
done

Each agent CSV file in results.tar.gz is then a concatenation of all associated results.csv files, with additional environment_name and seed fields. Note the learning curve data is missing state_value since logging for this quantity was added after the data was generated.

Plots show the average score at periodic evaluation phases during training. Each episode during evaluation starts with up to 30 random no-op actions and lasts a maximum of 30 minutes. To make the plots more readable, scores have been smoothed using a moving average with window size 10.

Plot of average score on each individual Atari game for each agent:

FAQ

Q: Do these agents replicate results from their respective papers?

We aim to replicate the mean and median human normalized score over all 57 Atari games and to implement the algorithm described in each paper as closely as possible.

However there are potential sources of differences at the level of an individual game. These include:

• Differences between Gym + Arcade Learning Environment (ALE) and Xitari.
• Changes in underlying libraries such as the exact image resizing algorithm used in the observation preprocessing.
• Atari ROM version.

Q: Is the execution of these agents deterministic?

We try to allow for it on CPU. However it is easily broken and note that convolutions on GPU are not deterministic. To allow for determinism we:

• Build a new environment at the start of every iteration.
• Include in the training state:
  • Random number generator state.
  • Target network parameters (in addition to online network parameters).
  • Evaluation agent.

Q: Why is DQN-based agent X not included?

There was a bias towards implementing the variants the authors are most familiar with. Also one or more of the following reasons may apply:

• Did not get round to implementing X.
• Have yet to replicate the algorithmic details and learning performance of X.
• It is easy to create X from components in DQN Zoo.

Q: Why not incorporate library / environment X?

X is probably very useful, but every additional library or feature is another thing new users need to read and understand. Also everyone differs in the auxiliary libraries they like to use. So the recommendation is to fork the agent you want and incorporate the features you wish in the copy. This also gives us the usual benefits of keeping dependencies to a minimum.

Q: Can I generalize X, then I can do Y with minimal modifications?

Code generalization often makes code harder to read. This is not intended to be a library in the sense that you import an agent and inject customized components to do research. Instead it is designed to be easy to customize after forking. So rather than be everything for everyone, we aimed to keep things minimal. Then users can fork and generalize in the directions they specifically care about.

Q: Why Gym instead of Xitari?

Most DeepMind papers with experiments on Atari published results on Xitari, a fork of the Arcade Learning Environment (ALE). The learning performance of agents in DQN Zoo were also verified on Xitari. However since Gym and the ALE are more widely used we have chosen to open source DQN Zoo using Gym. This does introduce another source of differences, though the settings for the Gym Atari environments have been chosen so they behave as similar as possible to Xitari.

Contributing

Note we are currently not accepting contributions. See CONTRIBUTING.md for details.

Citing DQN Zoo

If you use DQN Zoo in your research, please cite the papers corresponding to the agents used and this repository:

@software{dqnzoo2020github,
  title = {{DQN} {Zoo}: Reference implementations of {DQN}-based agents},
  author = {John Quan and Georg Ostrovski},
  url = {http://github.com/deepmind/dqn_zoo},
  version = {1.2.0},
  year = {2020},
}