The first level of super mario bros. is easy with lexicographic orderings and time travel . . . after that it gets a little tricky

Abstract

This paper introduces a simple, generic method for automating the play of Nintendo Entertainment System games. The approach is practical on a single computer and succeeds on several games such as Super Mario Bros.. The approach is amusingly elegant and works by inferring a player’s notion of progress from a short recording of their inputs to the game. The learned progress is then used to guide search over possible inputs, using an emulator. The objective function is generated by learning lexicographic orderings on memory locations. The search strategy favors locally preserved objectives rather than globally preserved objectives. As such, the learned behaviors may differ from human behaviors. – AI-generated abstract.