FlappyAI

FlappyAI is a fun project of mine.
It was time to play with some knowledge I've learned in my free time and AI classes. I wanted to start easy so I chose a very simple game to learn, Flappy Bird. The player has two actions, jump or do nothing, and the goal is to pass as many pipes as possible.

The machine learning takes inspiration from the deep reinforcement learning algorithm used by DeepMind Technologies in 7 Atari games.
DeepMind Article

The model uses a form of Q learning, using the environment as input to maximize rewards for specific state-action pairs without a given policy. It uses an epsilon-greedy approach to encourage it to try new actions but as every simulation passes, the epsilon is decreased to limit experimentation and to start exploiting what it has learned.

Highlights

• Model Free Reinforced Learning with epsilon-greedy Q learning
• Python/PyGame
• Tensorflow
• OpenCV