Game Playing (2048) Using Deep Neural Networks

Abstract

A one-player game, 2048 is also known as stochastic puzzle. This fascinating and engaging game has gained widespread acclaim and drawn researchers to create gaming software. The game 2048 has evolved into an engaging and difficult platform for assessing the efficacy of machine learning techniques because of its simplicity and complexity. Convolutional neural networks were used to create some computer players, but they performed poorly. In this work, the authors create a 2048 agent based on the reinforcement learning method and neural networks. The authors want to outperform other neural-network-based competitors in terms of results. Additionally, cutting-edge software created using this methodology for 2048 achieves the best performance out of all learning-based algorithms.