Super Smash Bros. Melee
A team of researchers working at the Massachusetts Institute of Technology’s Computer Science and Artificial Intelligence Laboratory has taught a neural network model how to play Super Smash Bros. Melee, a game chosen because of its intricacy and depth. The abstract of the paper based on their research makes direct reference to the title’s complex dynamics, as well as the added complication of partial observability.
The researchers trained the AI by supplying it with coordinates of crucial objects and items like other players and ledges that it could fall from, according to a report from Tech Crunch. Strategies that resulted in victory were incentivized, fostering the sort of play that could worry even top-level competitors.
The AI was pitted against several players who are ranked among the top 100 Super Smash Bros. Melee players worldwide, and managed to win more games than it lost. This is undoubtedly a big achievement — although there are a few major holes in the way the computer plays the game.
For one, it’s apparently unable to work with projectiles, forcing it to use fan favorite character Captain Falcon who does not have access to any such ranged attacks. It also has a strange quirk that forces it to jump to its death whenever an opponent heads to a corner of the stage and crouches for a long period of time, which team leader Vlad Firoiu cites as evidence that AIs trained in simulation might have unexpected behaviors when let loose in the real world.
Of course, those comments speak to the fact that this particular project wasn’t really about Super Smash Bros. Melee
