Deep Reinforcement Learning in Games Based on Extracted Features

Abstract: FlappyBird is a popular mobile game that captured many people's attention because itwas easy to understand but difficult to perform --- players were often right on the edge ofsucceeding, which led to a strong desire to play again. The purpose of this project is to investigatethe possibility of using a neural network trained with reinforcement learning to play the game usingextracted features rather than raw images. Specifically, we use the Q-learning algorithm to train aneural network on a set of pre-defined game features. Our results show that the agent can navigatethe game and collect points, but there are limitations in its performance due to factors such assuboptimal hyperparameters, the agent lacking information about the game's dynamics, and limitedcomputing power. Despite these limitations, the agent was able to navigate the game and collectpoints, but there is still room for further exploration. Overall, this project demonstrates the potentialfor using reinforcement learning and feature extraction techniques to train agents for simple mobilegames like FlappyBird.