A.I. Learned How to Compete Against Esports Pros in Highly Complex Games
Artificial intelligence has been dunking on humans at perfect information board games like Chess and Go for years now. In fact, the era of refining A.I.’s problem solving-skills using these games may have finally ended, and far more complicated games are required to continue pushing A.I.’s limits. 2018 saw major strides in this regard.
The most notable esports-playing computer was the “Open AI Five,” a team of bots that absolutely balls at the strategy game, Dota 2. The Elon Musk and Sam Altman-founded non-profit trained its A.I. dream team on the equivalent of 180 years’ worth of Dota 2 games, using 128,000 processor cores and 256 graphics processors. This resulted in a squad of bots that held its own against the games’ top, human players but didn’t quite clinch the victory.
Dota 2 is what’s known as a “multiplayer online battle arena” game, which involves two teams of five players each trying to destroy the other’s home base. Games can last anywhere between 15 minutes to an hour, and are won with a mix of technical prowess, a keen sense of what your opponent is up to, and decisive decision making. The fact that the Open AI Five can kind of do all of this to an extent is a breakthrough.
This is #12 on Inverse’s list of the 20 Ways A.I. Became More Human in 2018.
There are some caveats: In order to train the bots, OpenAi dumbed down Dota 2’s rules by limiting the roster of available characters, and removing some of the game’s details. However, the bots have slowly become capable of increasingly complex strategies and over the course of just a few months.
Artificial Intelligence researchers use games to train A.I. because the games themselves act as a surrogate for increasingly complicated forms of problem solving and decision making. Games like chess and Go are also attractive because they are elegantly complicated but also “perfect information” games, where all the relevant information about a players’ pieces and moves are seen by their opponent.
2018 proved that perfect information games are in fact too-easily mastered by computers, and that in order to continue refining A.I., researchers will need to turn to much, much better games.