Google’s DeepMind won two, five strokes. It is seen as a crucial step forward for AI research.
For fans of movies like Terminator, it sounds like a bad idea to train artificial intelligence in winning over people on a battlefield.
But that is, nonetheless, what Google is doing.
And in December, their artificial intelligence (AI) won DeepMind so convincingly over two professional computer players in the advanced real-time strategy game Starcraft 2.
The history of artificial intelligence is marked by a number of decisive victories in various games. And I hope even though there is still a lot of work to be done that people in the future will look back on this as another step forward for AI, says one of the leaders of Google DeepMind, David Silver, to the tech media. – The Verge.Se DeepMind StarCraft II Demonstration fra StarCraft på www.twitch.tv
Normally difficult for artificial intelligence
Starcraft 2 is a game in which different factions in the distant future are fighting for military and resource domination on alien planets.
Computer games like Starcraft are very advanced and harder for the artificial intelligence to win than classic strategic games like Chess and Go.
Participants start in different places on a map and do not know where the other participants are. You win by being the best at utilizing the path resources to build your army so that it can defeat the opponent.
It all happens in real time. One cannot pause the game. And thus, it becomes much harder for the artificial intelligences that can not predict the next 10,000 moves.
DeepMind won with ease
But Google DeepMind couldn’t stop it. In the game, the AI was named AlphaStar and it easily beat the professional players Dario “TLO” Wünch and Grzegorz “MaNa” Komincz.
According to The Verge, professional Starcraft commentators called AlphaStars’ “phenomenal” and “superhuman” games.
The Verge also writes that it was especially the way the artificial intelligence controlled individual entities in the game that were impressive.
Concerning gameplay Kotaku describes Dario Wünsch AlphaStars strategies as untraditional and completely impossible to enter into the rhythm of.
Was five different players
According to Kotaku, an explanation for it might be that AlphaStar was in fact five different “agents”, parts of the AI trained in their own way. But all based on a so-called neural network that had observed people playing Starcraft 2.
According to AI researcher Dave Churchill, whom The Verge has spoken to, the victory is a big step forward.
- I think the participant’s (AI’s) strength is a significant achievement, and it came at least a year before the most optimistic guesses I have heard among AI researchers, he says to the media.
Won a single match
The player Grzegorz “MaNa” Komincz won a single subsequent battle over the AI after it was modified to have to use the camera in the game like humans – instead of being able to see the whole card simultaneously.
To Kotaku, Grzegorz says “MaNa” Komincz that it was up to him how much his game usually depended on provoking human error.
We will see if that strategy works against a possible coming showdown with Skynet.