After 5 minutes and 36 seconds of battling, Grzegorz Komincz, known in the gaming world of StarCraft II as MaNa, announced "GG," an expression that video game players often use when they are defeated. His victorious opponent was AlphaStar, an artificial intelligence (AI) introduced by DeepMind Technologies, a British artificial intelligence company. StarCraft II, in which AlphaStar defeated MaNa, is a video game based on the “fog” of war, meaning that some information, like the locations of your enemy’s armies, is hidden. Therefore, players need to think, plan, decide, then act. By destroying all of its opponents’ armies, AlphaStar easily defeated MaNa, one of the top professional players in StarCraft II, with a 5-1 score; in December 2018, AlphaStar also beat MaNa’s teammate TLO (Dario Wünsch) with a 5-0 score.
Can you believe that it has only been two years since DeepMind’s AI AlphaGo mastered the board game Go and crushed the world champion Lee Sedol 4-1? Go, invented in China about 2,500 years ago, is known for its complex and unpredictable style of play, since there are 361 moves a player can make on every turn. The result shocked everyone when AlphaGo beat Sedol, and just recently, we witnessed another huge success in AI development.
Compared to Go, StarCraft II is even more challenging in terms of strategy and planning. There are approximately 10 to the 26th power (10 with 26 zeroes) actions a player can choose to take after each turn, which means AlphaStar needed to make the best decision in seconds and react in real life; this required a way more complicated technique compared to Go because the AI was required to think and act like a real human. Also, due to the imperfect information in the game, AlphaStar needed to actively discover the locations of enemies’ armies. Despite these challenges, AlphaStar had some advantages that no human can possess.
AlphaStar learned the game by watching other people play, and then refined its skills by playing against itself. In such a complex game, it is impossible to find a best way to win. So the machine learning agent basically split into hundreds of selves, and each had a task to accomplish. The hundreds of selves trained for hundreds of years of in-game time, and then they were gathered together to form the AI program, AlphaStar, who then would have 200 years of playing experience. The human players it fought against were all much younger than two hundred years old, obviously, so the match-ups were quite unfair.
A graphical representation of AlphaStar’s processing. The system sees whole map from the top down and predicts what behavior will lead to victory. Image: DeepMind/Verge
AlphaStar’s second advantage was that it can multi-task very well. There are two maps shown on its screen: a whole map in which it can see its units’ locations in the game, and a map that has the locations of every visible enemy’s units. From DeepMind’s data, AlphaStar switched its attention between the two maps 30 times per minute when it played against TLO in December.
The purpose of AlphaStar is not just to defeat humans in computer games; it is to accomplish more difficult tasks like “macro” planning (long-term), “micro” planning (controlling a big number of units effectively), real-time reaction, detective ability (since StarCraft II is based on the “fog” of war), and game theory by debating the best strategy to use. These techniques are most likely to be applied in future inventions. Although a lot of people doubt the victory proves that AI is a better StarCraft II player than humans, this is undoubtedly a triumph in the progress of technology.