sonic2_reinforcement_learning

Final project for the course,
Deep Learning (RAF117F fall 2018)
University of Iceland

Example videos

Previous version after 6 days of training:

Couple of modifications were made to the environment and hyper-parameters were re-tuned:

• The agent is limited to seeing 3 frames out of 60 per second and would play the same chosen action over the next 19 unseen frames.

• The environment terminated after 90 seconds instead of the default 10 minutes.

Modified version after 48h of training:

Prerequisites

Python 3
numpy
tensorflow
Gym Retro
Game ROM of Sonic the Hedghog 2

Running the code

Install gym retro, https://github.com/openai/gym

sudo apt install -y python3-dev zlib1g-dev libjpeg-dev cmake swig python-pyglet python3-opengl libboost-all-dev libsdl2-dev libosmesa6-dev patchelf ffmpeg xvfb
pip3 install gym
pip3 install gym retro
# pygame needed for some functions
pip3 install pygame

Install game rom from steam (if owned, else acquire the rom somehow else), https://store.steampowered.com/app/71163/Sonic_The_Hedgehog_2/

ls ~/.steam/steam/steamapps/common/'Sega Classics'/'uncompressed ROMs'/

Copy roms to some location with extention for system e.g. .md for mega drive, then import to gym retro

cp ~/.steam/steam/steamapps/common/'Sega Classics'/'uncompressed ROMs'/SONIC2_W.68K ~/roms/SONIC2_W.md
python3 -m retro.import ~/roms/

Run:

Python DQN_tensorflow.py

Author

Eyþór Einarsson

Links

• RAF117F, University of Iceland
• UCL Course on RL
• Reinforcement Learning, book by Richard S. Sutton and Andrew G. Barto
• Andrew Ng's Deep Learning Specialization on coursera.com