First Chess, then Go, and now Super Smash Bros. Melee?! Are the AIs finally getting the better of us at everything?
Not quite, but close enough!
Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory have created a Super Smash Bros. Melee AI superior to the regular "target practice" you can already fight in the game. And it’s good enough to hold its own against some global top 100 players!
Super Smash Bros. Melee was chosen because, according to the paper's abstract, its "environment has complex dynamics and partial observability, making it challenging for human and machine alike. The multiplayer aspect poses an additional challenge"
The team, led by Vlad Firoiu, trained a neural network model to play the game by feeding it gameplay data and incentivizing play that resulted in the computer’s victory. Rather than watching the screen and learning patterns, it is more like an in-game computer player that’s learned everything from scratch.
“It uses a combination of human techniques and some odd ones too – both of which benefit from faster-than-human reflexes,” wrote Firoiu in an email to TechCrunch. “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.”
However, it doesn't fare well with projectiles and it has a secret weakness:
“If the opponent crouches in the corner for a long period of time, it freaks out and eventually suicides,” Firiou wrote. (“This should be a warning against releasing agents trained in simulation into the real world,” he added)
Below is a video of the AI playing against several top 100 global players, against whom it won more than it lost!
While not a reason to fret for your gamer skills (yet?), the research demonstrates the interesting way in which AIs can cope in new environments. Moreover, game devs could make use of such neural networks to provide a more serious challenge to competitive and professional players to practice their skills.