TOKYO: Super Mario’s newest adventure is not about saving a princess, conquering Bowser or even completing a round of golf. It’s about being empowered to break free of the shackles of player control, become self-aware and sate his endless hunger for coins and monster-stomping.
The University of Tubingen in Germany has built a custom clone of the timeless video game which it uses to teach and converse with an artificial intelligence (AI) version of Mario, who in turn develops a personality and behaviour of his own.
The Mario Lives! project, created as an entry for the annual AI Video Competition, shows the chubby mushroom-lover coming to grips with the mortality of his enemies (“If I jump on Goomba, then it maybe dies”, he says after being instructed to obliterate the enemy and report on what he learns) and analysing potential future scenarios to plot jumps without falling to his death.
The AI also learns to react to the researchers playing with his emotions, that is lowering his “hunger” value (which he must rectify by “eating” coins) or making him “curious” (in which case he explores the environment and learns the values of various power-ups and game features).
There is something unnerving about the usually-friendly character reporting a feeling of unhappiness or naively admitting to having no knowledge of the effect his jumps might have on others, but at least he’s trapped within the confines of the game and not making Skynet-style judgements on the effect of jumping on his human masters. Yet.
Mario’s fame and the predictability of his games’ design have made him a popular subject for research and computer projects, particularly among game modders, tool-assisted speed-runners and amateur level designers (in fact, the popularity of the latter is such that Nintendo will this year release an official level design game for its Wii U console called Mario Maker).
While the Mario Lives! project shakes things up by removing the need for a player to help Mario on his quest at all, this is not the first time the Japanese-Italian plumber has been used to show off innovations in AI planning, learning and potential futuristics.
At the 2013 SigBrovik conference – an event for programming geniuses and computational tricksters to show off their more humorous projects – scientist Tom Murphy demonstrated an application that can teach itself how to play simple video games by analysing scores, win states, fail states and other variables.
When applied to Super Mario Bros., the AI eventually mastered the first level and – although it was initially confused by just about everything – it even figured out some exploits that would be impossible for a human player to pull off.
The software didn’t work for all types of games however, as evidenced in its application to Tetris, in which the AI inevitably decided the only way to succeed was to pause the game and never resume it.
As for the future of cognitive modelling and video game agents, Mario Lives! researcher Fabian Schrodt tells The Verge that the team’s next project will pair Mario and Luigi together so they can learn from each other and further distance the AI from its supposed human masters. Stephen Hawking was right to be afraid.
