Flappy Bird with Deep Reinforcement Learning

In this paper, a convolutional neural network model recently developed by Minh et al 2015 is applied to evaluate the Qfunction from raw pixel values from the screen. We took advantage of this method for Flappy Bird, a mobile game which is well known for being hard for humans to play. This method is capable of approximating Q-function to allow generalization to unseen states, not only it leads to faster convergence, but also it makes proposed method to be generalize enough to apply for different problems in different domain without changing much.