DeepMind AI secretly hides in the public scale of StarCraft II 1v1

Google's DeepMind Brings Its State-of-the-Art Artificial Intelligence StarCraft II again. We have already seen the AI, called "AlphaStar", take pro StarCraft II players in the matches to the test, but AlphaStar is now ready to face the public and players jumping on the European multiplayer scale 1c1.

Like the last time, AlphaStar is built with the cooperation of Blizzard (StarCraft IIdeveloper), and the official SC2 The website contains the details of the new incarnation of AlphaStar. The game's user interface now has a "DeepMind registration button" in the 1v1 Versus menu, which allows you to mix instances of AlphaStar into the human pool of multiplayer players. AlphaStar will play anonymously at 1v1 scale, so you will not know if you play AlphaStar or a human (I mean, I guess you could try calling your opponent). According to Blizzard, "playing anonymously with AlphaStar helps to ensure that it is a controlled test, so that the experimental versions of the agent offer a gaming experience as close as possible to a match in normal 1v1 scale ". Players will be paired with AlphaStar according to the usual match-making rules, and a victory or defeat will count as it would against a human.

The message contains a number of implementation details for this new version of AlphaStar, which seems to be a huge improvement over the version played. StarCraft II the pros in January. First, a number of improvements have been made so that the speed capabilities of the AI ​​correspond more to those of a human player. As an AI company, the stated purpose of DeepMind in this experiment is to play SC2 on an equal footing and teach AI things like thinking and long term planning – basically the strategy. At a very high level, one could say the two major components of everything StarCraft victory are "speed" and "strategy". Previous DeepMind's artificial intelligence experiments were turn-based games such as Chess and Go, where the speed at which you can move parts mattered little. As a real-time game, speed is an important factor in SC2 victory, and in previous games, AlphaStar has sometimes shown superhuman speed that gives it an unfair advantage and blurs the results of the experiment.

While humans are forced to play StarCraft By moving their fingers on a keyboard and a mouse, AlphaStar was wired directly into the game via an API created by Blizzard. For humans, SC2 It's about running multiple plates at once, like managing your base extension, positioning units, controlling your armies during the battle and doing it all the way through the limited camera in the game. With AlphaStar's direct control of the game, it would be relatively easy to create a winning AI at incredible speed and multitasking, with super-fast reaction times, perfect control of every unit of the game, and complete visibility of everything. what is happening on the map. . Limiting the speed and access to the AlphaStar game is essential to ensure that all wins are due to a superior strategy.

Blizzard says that this new version of AlphaStar "now perceives the game as a camera-like view," which was not always the case in January. At the time, when AlphaStar was playing against Grzegorz "MaNa" Komincz, the AI ​​bot had first won 5-0 with an unfair and unrestricted view of the match. Playing with a global view would provide more information that a human player is normally allowed to grab, with faster reaction times and easier multitasking. In the only game MaNa v AlphaStar where AlphaStar was more limited by the game's camera, he lost. In this new version, Blizzard notes that "AlphaStar does not receive any information about its opponent unless it is in the field of view of the camera and can only move the units to its position."

AlphaStar should also be excluded from the control of the superhuman unit he showed during the January matches. The speed of a player to control StarCraft is measured in "APM" or "Actions per minute", where each camera movement, unit click or basic function counts as an action. In January, DeepMind only limited the AlphaStar APM in five-second increments, meaning it could reach APM in superhuman burst for a few seconds at a convenient time. When a battle begins and there are dozens of units to control, this superhuman APM burst could easily make the difference between victory and defeat. In the new version, the peak APM has been capped. According to Blizzard, the new APM requirements "are more restrictive than the January DeepMind demonstration games and have been applied in consultation with professional players."

This new version of AlphaStar also seems a lot more complete. He can now play in the same way as any of the three races in the game, while in January he was only trained to play one race, Protoss. There is also no new version of DeepMind. Blizzard's message states that "DeepMind will analyze the performance of a number of experimental versions of AlphaStar to allow DeepMind to gather a wide range of results during the test period."

DeepMind promises at some point to publish these results in a peer-reviewed scientific article, as well as reruns of AlphaStar games. Good luck to everyone! Go ahead and win one for the human team.