Level Generation for Angry Birds with Sequential Autoencoder

Abstract

In this paper, we propose a deep generative model based level generation method for the video game Angry Birds. Although Angry Birds is a popular target for level generation, it is difficult to generate a stable level automatically because the level is governed by the gravity and there is a high degree of freedom in generating the level, and automatic generation using deep generative models is rarely done. In this study, we propose to encode levels sequentially and process them as text data, while existing methods process levels as images using a tile-based encoding method. The experimental results show that the existing methods fail to generate stable levels with high probability, while the proposed method succeeds in generating stable levels with high probability, and also generates levels with diversity.