MAP GENERATION IN DOOM USING GAN

• Main Author: Jansen
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/69196
• Institution: Institut Teknologi Bandung

Most game development is time consuming and need an expensive cost. Developer of the game will also produce game quality based on their skills and creativities. Therefore, a system is needed to create game contents in a short amount of time and require less resource and costs. One of the ways is using a Procedural Content Generation. Various research has been conducted for Procedural Content Generation development to create a good game content. One of the ways is using artificial intelligence to create game contents. Game map is one of the game contents. Game map is also a crucial part in developing a game. This research is focus on creating a Procedural Content Generation to generate game map using Generative Adversarial Network. This network is used because it can generate a high quality and realistic image similar to the images designed by the developer. This research was conducted by extracting images on the DOOM game file. Extracted images was then preprocessed to get a suitable size before used as a model input. Model learning will use three Generative Adversarial Network models before perform evaluation for each model. Generative Adversarial Network models used in this research are DCGAN, LSGAN and WGAN-GP. After perform model learning, results show that DCGAN and LSGAN generate images less similar to a real image set, while WGAN-GP generate images most similar to real image set. DCGAN and WGAN-GP model learning were also showed not convergence and unstable compared to LSGAN.