Breeding a diversity of Super Mario behaviors through interactive evolution

Creating controllers for NPCs in video games is traditionally a challenging and time consuming task. While automated learning methods such as neuroevolution (i.e. evolving artificial neural networks) have shown promise in this context, they often still require carefully designed fitness functions. In this paper, we show how casual users can create controllers for Super Mario Bros. through an interactive evolutionary computation (IEC) approach, without prior domain or programming knowledge. By iteratively selecting Super Mario behaviors from a set of candidates, users are able to guide evolution towards behaviors they prefer. The result of a user test show that the participants are able to evolve controllers with very diverse behaviors, which would be difficult through automated approaches. Additionally, the user-evolved controllers perform as well as controllers evolved with a traditional fitness-based approach in terms of distance traveled. The results suggest that IEC is a viable alternative in designing diverse controllers for video games that could be extended to other games in the future.