Reinforcement Learning for Improving Flappy Bird Game

Abstract

Currently, Artificial Intelligence becomes popularity among the human daily life, like games, Internet and so on. Authors has shown that in the game filed the Artificial Intelligence always have better performance than human beings, so in this article, the author wants to use AI to carry out on an old- fashion fame called Flappy Bird. This study aims to determine the specific method why AI has better performance than human beings. In this context, the author based on the process of experiments: It mainly used reinforcement learning model (Acting based on feedback from the environment, through continuous interaction with the environment, trial, and error, to ultimately accomplish a specific purpose or to maximize the overall benefits of the action) and supervised learning model (the process of making the machine learn a large amount of sample data with labels, training a model and making the model get the corresponding output according to the input) to improve the Flappy Bird and both two method are belonging to the machine learning. In addition, this study alters the layout of the game, including pipe, appearance of agent, and background of the game in order to make a more fashionable game. Furthermore, this study increases the number of agents, which makes it easier for agent to achieve higher score. Last but not the least, author establish a archive point, which means if the player face operation mistake and lead to game over, they bird will relive before passing the last pipe.