Observers have noticed an "artificial advantage" in the actions of the algorithm that humans do not have.
Artificial intelligence AlphaStar, created by DeepMind (owned by Alphabet Inc.), has beaten two professional players in the real-time strategy game StarCraft II. This is stated on the company's website.
Algorithm matches with two professional players - Dario Wunsch (Dario "TLO" Wünsch) and Grzegorz "MaNa" Komincz) - were held back in December, but this week, January 24, the last match took place, before which the program properties were changed. As a result, the victory went to a man for the first time.
Previously, DeepMind's algorithms beat the strongest go player and predicted protein shape better than professional biologists. However, in each of these cases, all available information was open to artificial intelligence. RTS like StarCraft II is different: it is a classified game. The machine does not know what is happening behind the fog of war, and cannot predict the player's actions, so it has to adapt to his maneuvers immediately.
The developers of the program used real people who were provided by Blizzard to teach it to play. By adopting the behavioral patterns of the players, she learned to defeat the artificial intelligence built into the game of the highest difficulty 95 percent of the time. Then her agents began to compete with each other in a knockout tournament. It took place in an accelerated mode, so each of them played for about 200 years.
Choosing the most successful agent, the authors put it against TLO and MaNa. First, he won five out of five matches against TLO, and then another version of the neural network defeated MaNa the same number of times. At the same time, the algorithm performed fewer actions per minute than its opponents. This is probably due to the small advantage that he still possessed. Unlike a person, he saw not a part of a known map (the same screen), but the whole of it, so he did not have to switch between different areas. Moreover, viewers noticed that he was giving orders to three detachments located in different zones at the same time, which a person is definitely not capable of.
In the last match, the developers made concessions and disabled this AI ability - this time, MaNa won. However, the authors noticed that the version of the algorithm that needed to move the camera was trained in the tournament for only seven days, which is half of the previous training period. Mana himself praised the actions of the program: "If I had played with a real person, he would not have been able to control his stalkers so delicately (a unit of one of the races - Ed.)."
The developers said that games like StarCraft II only serve as a training ground for artificial intelligence. In the future, he will take on other goals. For example, DeepMind had previously stated that their algorithm would diagnose eye diseases.