Advertisement
Tech by VICE

AI Can Now Demolish Humans In ‘Mortal Kombat’

Finish him.

by Jordan Pearson
Nov 10 2016, 12:00pm

Screengrab: YouTube

Video games can teach computers a lot, like how to "see" obstacles and solve problems. Artificial intelligence has already learned how to play classic games like Super Mario Bros. and titles on the Atari 2600, but these simple games no longer pose a challenge for today's machines.

This is according to a team of graduate students from the Israel Institute of Technology who recently designed what they call the Retro Learning Environment (RLE), a platform that allows researchers to set AI loose on a range of games on Nintendo's 90s console, SNES. Games they got the algorithm to play included F-Zero, Wolfenstein, and Mortal Kombat.

According to a paper the team recently posted to the arXiv preprint server, which is awaiting peer review, their algorithm had trouble with learning how to play most of the games, but after a process of trial and error, the machine was able to beat a human at one title: Mortal Kombat. (The paper describes the AI's opponent as an "expert human player.")

Read More: Video Games Are So Realistic That They Can Teach AI What the World Looks Like

The idea of a machine turning a person into ice and them smashing them to bits, even digitally, is a bit unsettling. But learning how to win in a game has many real-world applications for AI.

"If algorithms can play complex games, then we can project that into the real world to solve real problems," said Shai Rozenberg, one of the authors of the paper. "Similar to how a child might learn to play a game, the computer only sees the pixel information from the screen. They learn how to spot objects to avoid, and how to solve problems to gain the most rewards."

In other words, learning to play a videogame gives an AI transferrable skills that could be useful when, say, navigating a hallway without bumping into a trashcan.

The AI technique used by Rozenberg and his colleagues is called deep learning, wherein algorithms train themselves to accomplish tasks—really, recognizing patterns—through a process of trial-and-error. With every error, these algorithms slightly reorganize themselves to get a little bit closer to a solution until they can get it right with some reliability.

According to Rozenberg, the RLE can accommodate more advanced gaming systems than even the SNES, and the team has their sights set on systems like PlayStation next. But, since they quickly realized that their algorithms couldn't beat most SNES games reliably, they decided to stop there until those problems be solved.

"Moving forward, it is possible and even relatively easy to adapt our learning environment to more complex games, including things like Grand Theft Auto," Rozenberg said.

But, for now, AI needs to get better at nailing fatalities before it can start deftly jacking rides in Grand Theft Auto V's Los Santos.

Get six of our favorite Motherboard stories every day by signing up for our newsletter.