Aprimoramento da Geração de Mapas em Jogos Digitais por Meio de Estratégias Avançadas em GANs

Abstract

Procedural content generation (PCG) plays a key role in creating virtual worlds and optimizing digital game production, allowing the automatic generation of levels and maps without manual intervention. With advances in deep learning, generative adversarial networks (GANs) have emerged as a promising approach for PCG, offering the possibility of creating diverse and playable levels. However, significant challenges remain, such as mode-collapse, training instability, and difficulty ensuring that the generated maps are valid and consistent for gameplay. This research explores the tuning policies of GANs applied to map generation for top-down games, investigating strategies to optimize the diversity and validity of the generated maps. The central problem addressed in this thesis is improving the quality of levels generated by GANs. It seeks to answer the following question: How can generative adversarial network models be enhanced to maximize the generation of diverse and playable levels in digital games? To this end, several aspects of the training of these networks are analyzed, including mitigating mode-collapse, assessing the quality of the generated maps, increasing the expressiveness of the dataset, and penalizing the network with heuristics that favor gameplay. The main objective of this work is to define a set of techniques and methodologies that optimize the adjustment of GAN models for the procedural generation of levels in digital games. To this end, strategies are proposed that increase the expressiveness and quality of the dataset, minimize network training problems, and ensure that the generated levels are valid and varied. In addition, the aim is to integrate interactive and heuristic approaches to make the generative process more efficient and adaptable to the needs of different types of games. This research uses an approach that involves experimenting with different GAN architectures, including VanillaGAN, DCGAN, and WGANGP, and incorporating techniques such as spectral normalization, bootstrapping, and heuristic penalization. Metrics are explored to evaluate the validity and playability of the generated maps and strategies to condi tion the generation of levels in a controlled manner. The experiments are conducted with different datasets, comparing traditional PCG approaches with the proposed solutions. The results demonstrate that the strategies developed in this thesis signifi cantly improve procedural level generation and mitigate mode-collapse and instability problems. The experiments show that spectral normalization and heuristic penalization increase training efficiency in model tuning, enhancing its ability to generate playable and diverse levels. In conclusion, this research highlights the importance of refining adjustment policies for GANs in procedural map generation, highlighting the potential of these networks to optimize level design in digital games. However, challenges remain, such as the need to develop better mechanisms to control the variability and complexity of the generated maps. Future work could explore the combination of GANs with other deep learning approaches, such as more advanced conditional generative models, and apply these techniques in three-dimensional games and other genres, in addition to top-down vision games.