Super Mario no longer needs your help to save the princess now that scientists have developed an AI that can learn to play the game on its own. Researchers from the University of Tübingen computer science department have submitted their entry for the AAAI Video Competition 2015 People’s Choice Award showcasing their efforts to create an intelligent Mario who eats coins when he’s hungry, explores when he’s bored, and learns the rules of his world as he goes along.
How Mario learns to jump on Goombas
To accomplish all that, Mario keeps track of his surroundings and develops rules based on his experiences interacting with the rest of the game. The video shows him (the Super Mario 3D version specifically) saying at one point that he doesn’t know anything about Goombas, little mushroom monsters, and then jumping on the next one he sees to kill it, concluding that Goombas ‘maybe die’ when he jumps on them (translating Mario’s knowledge base into natural language is another part of the research). Repeated experiments reinforce the rule until Mario decides that a Goomba will ‘certainly die’ if he jumps on it.
The same experimentation lets him figure out how other game elements work, and the researchers can tweak how happy and how hungry he is to motivate him to explore more. They can also give him direct instructions like ‘look for an enemy’ which the AI is capable of understanding and acting on. For those who are interested in the nitty gritty, the team’s video has more details on the specific algorithms being used to produce natural language, store information, check rules and all the rest.
Mario AI isn’t playing to win
One interesting distinction between this AI and other video game AIs that we’ve seen before, like the ones developed by Deepmind that have Elon Musk so concerned, is that it doesn’t appear to be that interested in ‘winning’ the game. The Deepmind algorithm that was showcased last year was also able to determine the rules of the game on its own (using only visual input in its case) and then used that knowledge to exploit weaknesses in the game and rack up high scores. The University of Tübingen AI has its own built-in motivations, what the researchers refer to as hunger and happiness, a model that could be more generally applicable when AI research eventually moves beyond situations with clearly definable external goals.
