Guzdial, Riedl: AI-designed games match humans on challenge

A system that learns to design games by watching gameplay video produced novel titles that players could not reliably tell apart from human-made games on challenge, according to a paper by Matthew Guzdial and Mark Riedl published in IEEE Transactions on Games.

The approach departs from classical automated game design, which the authors write has "relied on the authoring of search spaces by a designer, defining the space of all possible games for the system to author." Instead, Guzdial and Riedl "learn representations of existing games from gameplay video and use these to approximate a search space of novel games."

In a human subject study, they report, the generated games were "indistinguishable from human games in terms of challenge," and one novel game was rated "equivalent to one of the human games in terms of fun, frustration, and likeability." Neither the participant count nor the number of games tested appears in the abstract.