Generative Adversarial Network Rooms in Generative Graph Grammar Dungeons for The Legend of Zelda

Generative Adversarial Networks (GANs) have demonstrated their ability to\nlearn patterns in data and produce new exemplars similar to, but different\nfrom, their training set in several domains, including video games. However,\nGANs have a fixed output size, so creating levels of arbitrary size for a\ndungeon crawling game is difficult. GANs also have trouble encoding semantic\nrequirements that make levels interesting and playable. This paper combines a\nGAN approach to generating individual rooms with a graph grammar approach to\ncombining rooms into a dungeon. The GAN captures design principles of\nindividual rooms, but the graph grammar organizes rooms into a global layout\nwith a sequence of obstacles determined by a designer. Room data from The\nLegend of Zelda is used to train the GAN. This approach is validated by a user\nstudy, showing that GAN dungeons are as enjoyable to play as a level from the\noriginal game, and levels generated with a graph grammar alone. However, GAN\ndungeons have rooms considered more complex, and plain graph grammar's dungeons\nare considered least complex and challenging. Only the GAN approach creates an\nextensive supply of both layouts and rooms, where rooms span across the\nspectrum of those seen in the training set to new creations merging design\nprinciples from multiple rooms.\n