Skip to content
Branch: master
Go to file

Latest commit


Failed to load latest commit information.
Latest commit message
Commit time

Simple DQN

Unfortunately this repo is outdated and there are much better codebases out there. I would suggest to take a look at this to learn the basics or this for full-blown DQN implementation for Atari.

Deep Q-learning agent for replicating DeepMind's results in paper "Human-level control through deep reinforcement learning". It is designed to be simple, fast and easy to extend. In particular:

See the example gameplay videos for Breakout, Pong, Seaquest and Space Invaders:

Breakout Pong Seaquest Space Invaders


Currently only instructions for Ubuntu are provided. For OS X refer to ALE and Neon documentation.


Install prerequisites:

sudo apt-get install libhdf5-dev libyaml-dev libopencv-dev pkg-config
sudo apt-get install python python-dev python-pip python-virtualenv
sudo apt-get install libcurl4-openssl-dev
sudo apt-get install libsox-fmt-all libsox-dev sox

Check out and compile the code:

git clone
cd neon

If you want to try out the filter visualization, use latest Neon and run make -e VIS=true instead. If you’ve already installed Neon without enabling visualization dependencies you’ll need to touch vis_requirements.txt prior to the make -e VIS=true call to ensure virtualenv Python dependencies get triggered.

Neon installs itself into virtual environment in .venv. You need to activate that to import Neon in Python:

source .venv/bin/activate

Arcade Learning Environment

You can skip this, if you only plan to use OpenAI Gym.

Install prerequisites:

sudo apt-get install cmake libsdl1.2-dev

Check out and compile the code:

git clone
cd Arcade-Learning-Environment
make -j 4

Install Python library (assuming you have activated Neon virtual environment):

pip install .

OpenAI Gym

You can skip this, if you only plan to use Arcade Learning Environment directly.

To install OpenAI Gym:

pip install gym
pip install gym[atari]

Simple DQN


pip install numpy argparse logging

Neon virtual environment already contains those libraries, but they are listed here, just in case.

Also you need OpenCV, which is pain to install to virtual environment. I ended up with this hack:

sudo apt-get install python-opencv
ln -s /usr/lib/python2.7/dist-packages/ NEON_HOME/.venv/lib/python2.7/site-packages/

NEON_HOME here means the folder where you installed (cloned) Neon.

Then just check out the code:

git clone
cd simple_dqn


For plotting install matplotlib:

pip install matplotlib

For producing game videos install avconv:

sudo apt-get install libav-tools

Running the code


To run training for Breakout:

./ roms/breakout.bin

If using OpenAI Gym:

./ Breakout-v0 --environment gym

There are plethora of options, just run ./ --help to see them. While training, the network weights are saved to snapshots folder after each epoch. Name of the file is <game>_<epoch_nr>.pkl. Training statistics are saved to results/<game>.csv, see below how to produce plots from it.

Resuming training

You can resume training by running

./ snapshots/breakout_10.pkl

Pay attention that the replay memory is empty.

Only testing

To run only testing on a pre-trained model:

./ snapshots/breakout_77.pkl

To test using OpenAI Gym:

./ snapshots/Breakout-v0_77.pkl

This saves testing results in folder results/Breakout-v0. Now you can then upload your results to OpenAI Gym:

./ results/Breakout-v0 --api_key <your_key>

Note that the OpenAI Gym environment differs from the default environment so testing using OpenAI Gym should use a model trained using OpenAI Gym.

Play one game with visualization

To play one game and show game screen while playing:

./ snapshots/breakout_77.pkl

You can do this even without GPU, by adding --backend cpu to command line. During gameplay you can use following keys:

  • a - slow down,
  • s - speed up,
  • m - manual control mode,
  • [ - volume down,
  • ] - volume up.

Visualization works even in text terminal!

Record game video

To play one game and record a video:

./ snapshots/breakout_77.pkl

First game frames are extracted to videos/<game> folder as PNG files. Then avconv is used to convert these into video, which is saved to videos/<game>_<epoch_nr>.mov.

Plotting results

To plot results:

./ results/breakout.csv

This produces results/breakout.png, which includes four main figures: average reward per game, number of games per phase (training, test or random), average Q-value of validation set and average network loss. You can customize the plotting result with --fields option - list comma separated CSV field names (the first row). For example default results are achieved with --fields average_reward,meanq,nr_games,meancost. Order of figures is left to right, top to bottom.

Visualizing filters

To produce filter visualizations with guided backpropagation:

./ snapshots/breakout_77.pkl

What the filter visualization does:

  1. first it plays one game to produce a set of states (one state is 4 frames),
  2. then it finds the states which activate each filter the most,
  3. finally it carries out guided backpropagation to show which parts of the screen affect the "activeness" of each filter the most.

The result is written to file results/<game>.html. By default only 4 filters from each convolutional layer are visualized. To see more filters add --visualization_filters <nr_filters> to the command line.

NB! Because it is not very clear how to visualize the state consisting of 4 frames, I made a simplification - I'm using only the last 3 frames and putting them to different color channels. So everything that is gray hasn't changed, blue is the most recent change, then green and then red. It is easier to understand if you look at the trace of a ball - it is marked by red-green-blue.

Nervana Cloud

To train a model with Nervana Cloud, first install and configure Nervana Cloud.

Assuming the necessary dependencies are installed, run

ncloud train src/ --args "roms/breakout.bin --save_weights_prefix snapshopts/breakout --csv_file results/breakout.csv" --custom_code_url

This will download the repo and run the training script.

To test a model using Nervana Cloud run:

ncloud train src/ --args "roms/breakout.bin --random_steps 0 --train_steps 0 --epochs 1 --load_weights snapshops/breakout_77.pkl" --custom_code_url


There are three additional scripts for profiling:

  • - runs Pong game 1000 steps in training mode. This is for figuring out bottlenecks in minibatch sampling and network training code. Prediction is disabled by setting exploration rate to 1.
  • - runs Pong game 1000 steps in testing mode. This is for figuring out bottlenecks in prediction code. Exploration is disabled by setting exploration rate to 0.
  • - runs Pong game 1000 steps with random actions. This is for measuring performance of ALE interface, network is not used at all.

Known differences

  • Simple DQN uses Neon's default RMSProp implementation, DeepMind uses different formulation from Alex Graves' paper (see page 23, eq 40).
  • Simple DQN uses averaged frame among skipped frame (which is ALE's built-in functionality), instead of max values from successive two frames as in the DeepMind paper.
  • Simple DQN uses Neon's Xavier initializer, DeepMind uses a fan_in parameter initializer.


This wouldn't have happened without inspiration and preceding work from my fellow PhD students Kristjan Korjus, Ardi Tampuu, Ilya Kuzovkin and Taivo Pungas from Computational Neuroscience lab run by Raul Vicente in University of Tartu, Estonia. Also I would like to thank Nathan Sprague and other nice folks at Deep Q-Learning list.


Simple deep Q-learning agent.




No releases published

Contributors 4



You can’t perform that action at this time.