The beginning of the end of the world at the hands of robots will not start in some dark, dank server room in the back of a research facility’s computer lab. It has already begun, and it’s taking place in Super Smash Bros. Melee.
That’s because of group of madmen and madwomen over at MIT have developed a neural network that is learning - and learning quick - how to kick serious ass at the Nintendo brawling game. They taught the AI how to identify enemies, and then incentivized it to play in novel ways that work around a human’s randomness. Gone are the days of computer-controlled enemies just following patterns and input cues. No, this one is basically alive. And it’s alive from the inside, rather than just observing objects on the screen like you or I would, this AI is talking to the code and learning from it as it goes.
The neural network is unusual from computer enemies in that it is both calculating and reckless. Its creators at CSAIL explain that sometimes the bot won’t even attack, choosing to “turtle” and waits for opportunities to attack. Add that to its electronic speed and reflexes, and SSBM experts are having a hard time beating it.
“It is sometimes very conservative, being unwilling to attack until it sees there’s a opening,” research team leader Vlad Firoiu told Techcrunch. “Other times it goes for risky off-stage acrobatics that it turns into quick kills.”
The end of the world at the hands of videogame-playing neural networks is still a bit off, however. It turns out that the program can’t deal with projectiles (which makes it less effective with those characters who shoot things), and you can make it panic into self destruction if you just hide in the corner.