Ů��С��Ƶ

AI has earned another victory against humans, this time in Nintendo fighting game Super Smash Bros. Melee.

A team led by at Massachusetts Institute of Technology trained an AI to play the game using deep learning algorithms and then pitched it against ten highly-ranked players. The AI came out on top against every one of them.

Super Smash Bros. is a cult Nintendo series where players battle classic video game characters like Super Mario and Zelda. The aim is to knock out (KO) the opponent by sending them out of bounds. The Super Smash Bros. Melee game was originally released in 2001 for the Nintendo Gamecube console.

The game may not be as complex as strategy games like Go, which Google DeepMind’s AlphaGo mastered in 2016, but Firoiu says it poses a different challenge for AI because you can’t work out many moves in advance. “You can’t plan far ahead with Smash like you can with, for example, Go,” he says. “To add to the difficulty, the attacks you perform can be used against you by your opponent.”

And unlike many video games AI has already conquered, like Pacman and Space Invaders, Super Smash Bros. is multiplayer, pitting two players against each other.

SmashBot beginnings

The new AI builds on previous efforts to make a Super Smash Bros. AI. A couple of years ago, security researcher Dan Petro was challenged by a friend who said it would be impossible to make an AI that could defeat him at the game. “I took that as a challenge,” says Petro.

Petro made a system called SmashBot based on his own experience playing the game. “I directly programmed an optimum strategy into SmashBot,” he says.

Firoiu and his colleagues saw Petro’s bot and asked if they could use his groundwork to take a Super Smash Bros.-playing AI to the next level. Petro had already built the infrastructure needed for an AI to interact with the Super Smash Bros. Melee game and control a character.

The researchers trained their AI using reinforcement learning. They first had it fight the in-game AI, which players can battle in one-player mode, then improved it by making it play against itself.

Finally, they took the AI to two tournaments and asked professional players to try to defeat it. The AI played as the popular character Captain Falcon, partly because this character doesn’t have any projectile attacks, which the AI wasn’t trained to deal with. The AI won more fights than it lost against each of the 10 high-ranking players it fought, who ranked from 16th to 70th in the world.

The whole process was fast. “After a couple of hours the AI was good enough to beat the in-game AI, and after a couple of weeks it could beat the top-ranking humans,” says Firoiu.

Reaction speeds

at the NYU Game Innovation Lab says it is not surprising that an AI has conquered Super Smash Bros. Melee, as computers excel at the fast reaction times that give players an advantage in this kind of game. “Compared to other games, fighting games rely very little on long-term planning and very much on quick reactions,” he says.

The AI plays with a reaction speed of around 33 milliseconds, compared to over 200 milliseconds for humans. The researchers are considering restricting the AI’s reaction speed to see if they can build a system that is strategically superior when playing at human speed.

Meanwhile, the AI still has a fatal flaw that the top-ranking players didn’t notice. If an opponent just crouched at the side of the stage, the AI didn’t know what to do. “It then killed itself,” says Firoiu.

This should act as a warning for researchers trying to apply AI to unfamiliar situations, says Firoiu. “If an AI encounters something that it’s not seen before, it can fail remarkably and spectacularly.”

Journal reference: arXiv,