MIT Creates Brutal AI That Can Beat Humans at Video Games and Learns on Its Own

The world was shocked in 1996 when the Deep Blue supercomputer beat Garry Kasparov, a world-class chess champion, at one of the most sophisticated strategy games ever devised. In retrospect however, it seems laughable that anyone was surprised by this development. Although chess is a sophisticated game, it’s also the perfect game for a computer. Deep Blue wins at chess through brute force rather than intelligence, by running through millions of possible games every time a player moves, and choosing the best counter-move. No human can compete with that.

But there are other games that you wouldn’t expect a computer to beat. While computers rule at methodical strategy, they’re not so hot when it comes to say, fast paced video games. At least until now.

MIT recently unveiled an AI that can easily beat humans at a video game called Super Smash Bros. Melee. In case you’ve never played it before, it looks like this:

To the uninitiated that may look like two human players going at it, but that’s not the case. One of those players is a computer. That doesn’t look like the kind of game that artificial intelligence could dominate at, but that’s what this program can do, and it does it pretty consistently. In this case, it went against a few of the best players in the world and won.

The AI is called Phillip, and it was created by MIT’s Computer Science and Artificial Intelligence Laboratory. According to the researchers who made Phillip, “The SSBM environment has complex dynamics and partial observability,  making it challenging for human and machine alike.  The multiplayer aspect poses an additional  challenge.” But this AI, which is capable of learning on its own, ruthlessly dispatches human players in unique ways.

“It uses a combination of human techniques and some odd ones too – both of which benefit from faster-than-human reflexes,” wrote Vlad Firoiu, who led the project. “It is sometimes very conservative, being unwilling to attack until it sees there’s a opening. Other times it goes for risky off-stage acrobatics that it turns into quick kills.”

Fortunately the program has a hilarious weakness. According to Firoiu, “If the opponent crouches in the corner for a long period of time, it freaks out and eventually suicides,” which he suggested “should be a warning against releasing agents trained in simulation into the real world.”

Keep that in mind for when the robot rebellion occurs. Just hide somewhere and wait for the machines to sort themselves out.