Deep Q-learning with Flappy bird

Mini project: Implementation of Deep Q-learning algorithm to create an AI agent capable of scoring high on Flappy Bird game

The algorithm implemented is fully described in this paper: Playing Atari with Deep Reinforcement Learning (Deep Q-learning with experience replay, page 5)

A demo video can be found here:

Package dependencies

• gym
• gym_ple
• Pygame-Development-Environment
• Pygame
• keras

Experience

To train from scratch, just run the command python flappy_bird.py -r

You can also download pretrained weights to ./model folder in the current dir and run the command python flappy_bird.py -l -r to see how the trained AI plays.

The training is undertaken on a machine with GTX860 gpu. So you need a GPU to reproduce the results.