A team of computer programmers from IT University of Copenhagen has created MarioGPT, a novel technique to encoding and generating Super Mario Bros. levels that is based on the GPT-2 language paradigm. In a publication on the arXiv pre-print site, the group describes their work and how others might utilise their technology.
Mario Brothers is a 1983 video game in which two Italian plumbers emerge from a sewer and seek to rescue Princess Peach, who has been abducted and imprisoned by Bowser. To save her, the brothers must navigate (through game user input) over a sequence of pipes and bricks obstacles. The landscape varies as they move, depending on the level they have attained in the game. The team in Denmark has reproduced one component of the game—the amount of levels that may be traversed—in its new endeavour.
To convert user requests into graphical representations of Super Mario Brothers game levels, the researchers employed Generative Pre-trained Transformer 2 (GPT-2)—an open-source language model established by an OpenAI team. To do this, they wrote some Python code to assist the language model in understanding what needed to be done, and then trained it using samples from the original Super Mario Bros. game and one of its sequels, “Super Mario Bros.: The Lost Levels.”
MarioGPT produces a new level of the game after querying for play style. Prompt responses may contain words like “many pipelines, numerous opponents, little blocks, low height.” The code and instructions for encoding and producing levels were shared on GitHub by the team. Users may create as many levels as they like.
In related news, a team from New York University Tandon collaborated with a colleague from the University of the Witwatersrand to produce and play versions of the game Sokoban, in which players push boxes about in a warehouse—they detail their work in a publication on the arXiv preprint service.
You might also be interested in reading, Engineering in black phosphorus for floquet bands
