Reinforcement Learning for video games

Our experiments with Q-learning, DQN, RNN and more

Directories:

• atari: DQN for atari video games
• catch: simple catch game for beginners in RL
• torch-flappy: flappy bird in python with Torch7 hooks
• doom: Doom simulator from: https://github.com/Marqt/ViZDoom

Status:

Atari:

• DQN CNN script working
• RNN script some good results, needs more testing

Catch:

• DQN CNN script working
• RNN script working

Flappy:

• DQN CNN script functional, all code tested, slow, never fully tested, training takes too long

Doom:

• DQN CNN script working
• RNN script functional, all code tested, but no good results

Notes:

Note1: RNN in Catch works on grid of 10 x 10 because we know exactly how long the sequence is going to be it is always gridSize-2! in other games this is not possible

FIXED: just needed to shift sequences so reward is at the end!