Imitating human playing styles in Super Mario Bros

We describe and compare several methods for generating game character controllers that mimic the playing style of a particular human player, or of a population of human players, across video game levels. Similarity in playing style is measured through an evaluation framework, that compares the play trace of one or several human players with the punctuated play trace of an AI player. The methods that are compared are either hand-coded, direct (based on supervised learning) or indirect (based on maximising a similarity measure). We find that a method based on neuroevolution performs best both in terms of the instrumental similarity measure and in phenomenological evaluation by human spectators. A version of the classic platform game “Super Mario Bros” is used as the testbed game in this study but the methods are applicable to other games that are based on character movement in space.

► We compare the capacities of different learning methods to imitate human gameplay.
► Ground truth about human-likeness is crowd-sourced over the Internet.
► All learning methods appear more human-like than non-trained controllers.
► Indirect imitation methods are superior to direct methods in both performance and human-likeness.
► Neuroevolution comes out as the top choice on both metrics.