Abstract:
We present an unsupervised process to generate full video game levels from a model trained on gameplay video. The model represents probabilistic relationships between shapes properties, and relates the relationships to stylistic variance within a domain. We utilize the classic platformer game Super Mario Bros. to evaluate this process due to its highly-regarded level design. We evaluate the output in comparison to other data-driven level generation techniques via a user study and demonstrate its ability to produce novel output more stylistically similar to exemplar input.
DOI:
10.1609/aiide.v12i1.12861