Google’s AlphaStar AI crushed StarCraft 2 pro players

Message
Games
January 25, 2019, 2:15 p.m.

Google’s DeepMind team – working on artificial intelligence – has announced that its AlphaStar program has defeated two professional StarCraft 2 players from Team Liquid. In both cases, the matches ended with the score of 5:0.

AlphaStar – self-portrait.

IN SHORT:

  1. AI AlphaStar defeated Team Liquid’s professional StarCraft 2 players – Dario “TLO” Wünsch and Grzegorz “MaNa” Komincz;
  2. In both cases, the artificial intelligence won 5–0, although Komincz later managed to win in an extra match;
  3. The AI ​​studied for 14 days, gaining experience comparable to 200 years of continuous play during this time.

The creators of the Google DeepMind team have announced that their “student” – AI AlphaStar – crushed pro players Dario “TLO” Wünsch and Grzegorz “MaNa” Komincz from Team Liquid in StarCraft 2. Two separate series of five games were played – in both cases the clashes ended with the score of 5: 0 for the artificial intelligence. The conditions and rules of the game were comparable to those of professional competition, but matches could only be played with the Protoss race.

Initially, nothing announced such results – watching the matches played by AlphaStar Wünsch was absolutely convinced that he would be able to cope with artificial intelligence, even despite the aforementioned limitation (Protoss are not his favorite race – the skill level “TLO” when playing them is GrandMaster). However, when he started playing, the AI ​​each time used a completely different, unconventional strategy that he couldn’t defend against.

The graph shows the skills of the AI ​​agents (measured in MMR - Match Making Rating) against the SC2 leagues and the two professionals.

The graph shows the skills of the AI ​​agents (measured in MMR – Match Making Rating) against the SC2 leagues and the two professionals.

Where does AlphaStar’s good result come from? The neural network underlying the program initially learned the rules Starcraft 2 watch replays of human matches. Following AI training it has entered its second phase – artificial intelligence “divided” and began to play with itself. The completed meetings resulted in the formation of subsequent subroutines that utilized the knowledge gained, which further increased the variety of tactics. Each of the so-called “Agents” meanwhile have gained experience equivalent to… 200 years of continuous play Starcraft 2. Finally, the most effective discovered strategies were selected and contained in a single “agent” running on a single desktop GPU.

The face of the bespectacled gentleman in the audience - a picture is worth a thousand words.

The face of the bespectacled gentleman in the audience – a picture is worth a thousand words.

Interestingly, in terms of response time, the AI ​​was no better than that of professional gamers. His APM (actions per minute) rating was around 280, which DeepMind developers say is a much lower result than human pros. The statistics for the delay between observing a given state of affairs and reacting to it are similar – it took on average about 350 milliseconds. However, the AI ​​makes decisions differently than people and perceives the game differently – the AI ​​was looking at the whole map at once, without having to move the camera. The developers also developed a version that was supposed to control the camera like humans – this one trained for 7 days, but was defeated by “MaNa” (work is still in progress).

Finally, it should be noted that the AI ​​that Grzegorz Komincz faced was wiser about the experience gained from clashes with “TLO”. “MaNa” also managed to salvage the honor of our race by winning a single match, streamed live (below) by Blizzard and DeepMind. It therefore seems that there is still hope for humanity, even if we must admit that our chances do not look very good.

  1. StarCraft 2 – official site
  2. DeepMind – official site

Leave a Comment