Machine Learning AI Demolishes World’s Top Super Smash Bros. Players
Watch out world, machines are beating flesh and blood opponents in a variety of different contests. There was Watson, a smarty pants question answering system developed by IBM that took on and defeated the world's best Jeopardy opponents. Then Google's DeepMind division built an artificial intelligence program called AlphaGo that became the world's top Go player. And now there's an AI that seemingly has no equal in Nintendo's Super Smash Bros. game.
The AI team was led by Vlad Firoiu at Massachusetts Institute of Technology (MIT). Using deep learning algorithms, the teamed honed the AI's skills in Nintendo's popular brawler featuring iconic game characters. The goal in the game is to knock the opponent out of bounds, which becomes increasingly easier to do after doling out physical damage using a number of different hits and special powers.
Super Smash Bros. isn't as complex as a game like Go, but in terms of building an AI that can compete with human players, the brawler presents a different kind of challenge. Firoiu explains that unlike strategy games, you can calculate moves in advance.
"You can't plan far ahead with Smash like you can with, for example, Go," Firoiu told New Scientist. "To add to the difficulty, the attacks you perform can be used against you by your opponent."
This new AI is a second attempt at becoming a boss in Super Smash Bros. An earlier version developed by security researcher Dan Petro was unsuccessful in asserting its dominance. A friend of Petro's told him it would be impossible to build an AI that could beat him at the game. Petro took that as a challenge and developed a system called SmashBot based on his own experience playing Super Smash Bros.
The bot got the attention of Firoiu and his colleagues. They asked Petro if they could build on top of his bot, and he agreed. The result is an AI that uses reinforcement learning to smash its opponents. It has a reaction speed of around 33 milliseconds, versus around 200 milliseconds for humans, giving it an inherent advantage.
So, what comes next? The researchers are considering dialing back the AI's reaction time to see if they can further tweak the program to beat humans when playing at their speed.
The AI team was led by Vlad Firoiu at Massachusetts Institute of Technology (MIT). Using deep learning algorithms, the teamed honed the AI's skills in Nintendo's popular brawler featuring iconic game characters. The goal in the game is to knock the opponent out of bounds, which becomes increasingly easier to do after doling out physical damage using a number of different hits and special powers.
Super Smash Bros. isn't as complex as a game like Go, but in terms of building an AI that can compete with human players, the brawler presents a different kind of challenge. Firoiu explains that unlike strategy games, you can calculate moves in advance.
"You can't plan far ahead with Smash like you can with, for example, Go," Firoiu told New Scientist. "To add to the difficulty, the attacks you perform can be used against you by your opponent."
This new AI is a second attempt at becoming a boss in Super Smash Bros. An earlier version developed by security researcher Dan Petro was unsuccessful in asserting its dominance. A friend of Petro's told him it would be impossible to build an AI that could beat him at the game. Petro took that as a challenge and developed a system called SmashBot based on his own experience playing Super Smash Bros.
The bot got the attention of Firoiu and his colleagues. They asked Petro if they could build on top of his bot, and he agreed. The result is an AI that uses reinforcement learning to smash its opponents. It has a reaction speed of around 33 milliseconds, versus around 200 milliseconds for humans, giving it an inherent advantage.
So, what comes next? The researchers are considering dialing back the AI's reaction time to see if they can further tweak the program to beat humans when playing at their speed.
