Tensorflow implementation of Deep Q-Network (DQN) and Behavior Cloning (BC) to learn how to defeat humans in a FlappyBird game.
- Tensorflow implementation of DQN similar to the paper Human-level control through deep reinforcement learning [Mnih et al., 2015].
- Some visualization tools for analysing experimental results.
- Possibility to learn from expert dataset (Behavior Cloning).
The environnement used for the experiments is the flappyBird_cnn Gym environnement from this repository [blavad].
Which parts of the inputs were decisive when the AI won against human at Flappy Bird Game ?
We can visually explain actions taken by the agent via Gradient-based Localization [Servaraju et al., 2016].
More details: Internship report (in French)