In a nutshell
- AI AlphaStar defeated two pro StarCraft 2 players from Team Liquid - Dario "TLO" Wünsch and Grzegorz "MaNa" Komincz;
- In both cases the AI won 5:0, although Komincz managed to score a win in one additional game played afterwards;
- AI learned for 14 days, gaining the experience comparable to 200 years of continuous play during that time.
Developers from Google's team at DeepMind reported that their "pupil" – AI AlphaStar – wrecked two pro StarCraft 2 players: Dario "TLO" Wünsch and Grzegorz "MaNa" Komincz from Team Liquid. Two separate series of five games were played – in both cases the battles ended with a 5:0 win for the artificial intelligence. The conditions and rules of the game were comparable to those of professional games, but matches could only be played with the Protoss.
Initially, nothing pointed at such an outcome – observing the games played by AlphaStar, Wünsch was absolutely convinced that he would take down the AI, even despite the aforementioned limitation (Protoss are not his favourite race – TLO’s level when playing them is ranked as GrandMaster). However, when he started to play, the AI used a completely different, unconventional strategy, which he was unable to counter.
Where did AlphaStar’s impressive score come from? The neural network, which was the basis of the program, initially learned the principles of StarCraft 2 by observing players in matches played by people. Then the AI training entered the second phase – the artificial intelligence "split up" and began to play with each other. Completed matches resulted in the creation of new sub-programs, that used the acquired knowledge, which further increased the variety of tactics. Each of the so-called "agents" gained experience equivalent to... 200 years of continuous playing StarCraft 2. Finally, the most effective of the discovered strategies were selected and included in a single "agent" launched on a single desktop GPU.
Interestingly, in terms of response time, the AI was not better than professional players. Its APM (actions per minute) ratio was about 280, which, according to DeepMind programmers, is a much worse result than with human pros. The stats regarding the delay between observation of a given state of affairs and responding to it look similar – it took the AI about 350 milliseconds on average. However, artificial intelligence makes decisions differently than humans and perceives the game differently – the AI watched the whole map at once, without having to move the camera. However, it should be noted that the map was still covered by the fog of war. Developers have also developed a version that had to control the camera just like people do – it trained for 7 days, but was defeated by MaNa. (the AI is still under development).
Finally, it is worth noting that the AI that Grzegorz Komincz faced, was smarter with the experience gained in the battles against TLO. MaNa eventually managed to save the honor of our race, winning in a single match, streamed live (recorded below) by Blizzard and DeepMind. It seems that there is yet hope for humanity, although it must be admitted that our chances are slim.
- StarCraft 2 – official website
- DeepMind – official website