Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
..
Failed to load latest commit information.
README.md
atari_game.py
base.py
dqn_demo.py
dqn_run_test.py
game.py
operators.py
replay_memory.py
setup.sh
utils.py

README.md

Deep Q-Network Example in MXNet

Prerequisites:

  • Install the development version of MXNet first.
  • Install the Arcade Learning Environment (ALE)
    • You might need to copy the ale.cfg into the dqn folder
  • Install SciPy: python -m pip install --user scipy
  • Make sure the Roms are available for example download as: git clone https://github.com/npow/atari/tree/master/roms

You can also run the script setup.sh to do this automatically for you, and run the required steps. You might want to consider running it in a dedicated virtualenv with mxnet installed.

Training

  • Run ./dqn_demo.py to train AI for Breakout.
  • In case breakout.bin is not found run as: ./dqn_demo.py -r <Path_To_Roms>

Testing

Run ./dqn_run_test.py -d YOUR_SAVE_DIR -m QNet -e EPOCH_RANGE -v where EPOCH_RANGE is a pair of numbers delimited by comma

Sample Play Video

https://www.dropbox.com/s/cglqxaipyap1pe7/dqn-breakout-197.avi?dl=0