libfun/deephack3

Public repo with Will Press Lever For Food solution at DeepHack.RL

rl.deephack.me

Gym submissions:

• Skiing-v0 -3492.77 ± 1.94
• MsPacman-v0 5512.30 ± 57.51
• Centipede-v0 7511.87 ± 445.89

Team Will Press Lever For Food:

Alexander Guschin, Sergey Korolev, Sergey Ovcharenko, Sergey Sviridov, Max Kharchenko

How to reproduce results

1. requirements.txt
2. python tensorpack/setup.py install to install modified tensorpack version.
3. Train modified universe A3C agent (https://github.com/libfun/deephack3/blob/master/universe-starter-agent/README.md)
4. Train modified tensorpack A3C agent (https://github.com/libfun/deephack3/blob/master/tensorpack/examples/A3C-Gym/README.md)
5. player.py contains code for agent combination. Examples in Untitled.ipynb.