How to improve the quality of GAN-based map generators

  • Daniele Fernandes E Silva IFFar / UFPel
  • Rafael Piccin Torchelsen UFPel
  • Marilton Sanchonete De Aguiar UFPel

Resumo


Procedural Content Generation algorithms aim to create unique and variable dungeon maps, ensuring that players encounter infinite maps in the game. This capability is essential to prevent repetitive environments, keeping players engaged and providing them with new challenges and discoveries. Machine learning techniques, such as Generative Adversarial Networks (GANs), have proven effective in generating data, although they may have specific limitations. This paper proposes a GAN-based approach for generating dungeon maps and introduces three optimizations to enhance the training process. Our approach achieves remarkable results in producing valid and varied maps compared to existing methods. We demonstrate that our approach outperforms other approaches by generating more valid maps with increased variability.
Palavras-chave: dungeon game, generative adversarial networks, level generation, neural networks, procedural content generation, unsupervised learning
Publicado
06/11/2023
FERNANDES E SILVA, Daniele; TORCHELSEN, Rafael Piccin; AGUIAR, Marilton Sanchonete De. How to improve the quality of GAN-based map generators. In: SIMPÓSIO BRASILEIRO DE JOGOS E ENTRETENIMENTO DIGITAL (SBGAMES), 22. , 2023, Rio Grande/RS. Anais [...]. Porto Alegre: Sociedade Brasileira de Computação, 2023 . p. 106–113.