Generating Procedural Dungeons Using Machine Learning Methods

Abstract

Procedural content generation (PCG) is a powerful tool to optimize creation of content in the game industry. However, it can lead to lack of control and mischaracterization of the game design, creating unbalanced or undesired situations. To overcome such problems, machine learning can be used to map important patterns of a game design and apply them in the PCG. Considering such aspects, this paper proposes a strategy for procedurally generating dungeons using ML techniques. We use Unity ML-Agents tool for the implementation, since dungeons are environments largely used in the industry that also require more control over its creation. The strategy used in this paper has proven to generate dungeons that respect room positioning design choices and maintains the game characterization. We conclude, after conducting a survey with users, that the generated dungeons presented reliable maps and showed to be more enjoyable and replayable than manually generated ones following the same design principles.