As part of another achievement for Artificial Intelligence, it can now beat humans in a popular video game — StarCraft II — and attain the highest-ever level in the game.
DeepMind’s AlphaStar AI
AlphaStar AI, developed by Google’s AI wing DeepMind, could play StarCraft II fully, defeat humans, and ultimately reach the highest level in the game, which is called the Grandmaster level.
DeepMind’s AI was able to attain a score of 99.8% out of all the other players on Battle.net and surpassed all three StarCraft II races: Protoss, Terran, and Zerg.
Several machine learning techniques such as neural networks, self-play methods, multi-agent learning, and imitation learning were used to train AlphaStar and finally attain the breakthrough.
AlphaStar Played StarCraft Previously!
For those who don’t know, AlphaStar (back in January) played StarCraft II at preparatory level against two of the world’s top players of the game.
While the AI defeated the players at one point, the players too defeated it at another, suggesting both needed improvements.
This time, AlphaStar faces similar conditions as humans face during gameplay.
Conditions include viewing the world through a camera, stronger limits on its actions, and gameplay on an official game server with the same conditions as that of humans.
How Is Winning StarCraft II A Breakthrough?
StarCraft II is one of the most popular and complex video games present out there.
Being able to win the game over humans is an achievement for AI because the complex and open-ended nature of the game is a challenge for AI.
Another aspect of the game is that it has up to 1026 possible actions. This requires AI to look for all possible actions and make a decision and apply it to win or lose.
Previous AI Achievements
As a reminder, researchers have developed AI algorithms that have successfully beaten humans in the game of chess, Go, checkers, and even Dota 2.
Recently, Facebook’s AI division developed Pluribus to beat professional poker players in the popular six-player no-limit Texas Hold’em poker format.