The AI versus Human

DeepMind and Blizzard cooperate and develop an AI to play Starcraft II. Today they live steamed a demonstration about the performance of their AI, Alphastar. They showed replays of AlphaStar playing against two professional players. There are three races in Starcraft II, at the moment, AlphaStar can only play Protoss against Protoss at the moment. In general, AlphaStar can play Starcraft well; however, it is not invincible like AlphaGo, in the live game, the human professional player defeated AlphaStar and the game was not a very close game either.

Training AI to play Starcraft is a very interesting topic that attracts, not because I like playing the game, but more importantly because it create an environment that can show how an AI solves a problem with imperfect information. Starcraft is a very complex game that has fog of war that creates an environment with imperfect information. In he human world, when people are playing a game with imperfect information, they design their strategies based on their expectations about other players in the game, including other players' expectations about themselves, and other players' expectations about their expectations about other players and so on. It would be interesting to how an AI forms its expectations about its human opponents. To be invincible, AlphaStar has to form an algorithm that precisely make an estimation about its human opponent's expectation. It is a great experiment for people to study human behaivours, especially the patterns about how people form their expectations.