Flappy Bird Game Based on the Deep Q Learning Neural Network

Abstract

In the past decades, with the rapid development of technology, people discovered ways for machines to learn. Machines can be trained to recognize things, play games, create sounds, or find the best choices. There are several models and tools to train the machine to make the machine learn through supervised or not supervised, even independent learning. Through neural networks or other methods, the machine can be trained for many episodes. Making awards or punishments can make the machine construct ways to decide the most valuable solution. Among these functions AI can achieve, game is the most direct platform to apply machine learning, since awards and punishments can be applied very easily: through the score. In this study, we choose to use a Deep-Q-Learning Neural Network (DQN) to train our AI to achieve our goal: Using AI to play Flappy Bird through deep learning. Our task is different from other game training, such as navigating an AI to find the best solution in different choices. In this task, the player (i.e. bird) cannot get any award or punishment through a single action, but it can get an award by passing each obstacle. The goal of the AI is to pass as many obstacles as it can, by choosing to fly upward or idle.