Enhancement of Angry Birds Level Generation from Sketches Using Cycle-Consistent Adversarial Networks
Enhancement of Angry Birds Level Generation from Sketches Using Cycle-Consistent Adversarial Networks
复制标题
DOI：
10.1109/gcce50665.2020.9291893
复制
发表时间：
2020-10
期刊：
影响因子：
--
通讯作者：
Mury F. Dewantoro;Febri Abdullah;Pujana Paliyawan;R. Thawonmas;F. A. Bachtiar
中科院分区：
文献类型：
--
作者：
Mury F. Dewantoro;Febri Abdullah;Pujana Paliyawan;R. Thawonmas;F. A. Bachtiar
本文介绍了我们的工作，以提高一个国家的最先进的水平发生器（草图到水平发生器），产生水平的愤怒的小鸟一样，从绘制草图的游戏。为了实现这一任务，使用了周期一致性对抗网络（Cycle-GAN）。CycleGAN使用两个数据集进行训练：草图和典型的层次结构。Google的Quick，Draw！数据集，后者来自2017年和2018年AIBIRDS水平生成比赛的获胜水平生成器。经过训练的CycleGAN的输出用作草图到级别生成器的输入。我们的研究结果表明，建议使用CycleGAN预处理技术允许草图到级别生成器更成功地从任意草图生成级别。
This paper presents our work to enhance a state-of-the-art level generator (Sketch-to-Level Generator) that generates levels for an Angry-Birds-like game from drawn sketches. To achieve this task, Cycle-Consistent Adversarial Networks (Cycle-GAN) are used. CycleGAN is trained using two datasets: sketch drawings and typical level-structures. The former are taken from Google’s Quick, Draw! datasets, and the latter from the winning level generator at the 2017 and 2018 AIBIRDS level generation competitions. The output of the trained CycleGAN is used as the input of Sketch-to-Level Generator. Our results show that the proposed preprocessing technique using CycleGAN allows Sketch-to-Level Generator to more successfully generate levels from arbitrary sketch drawings.