The international magazine Nature published a study on the 14th of this month in which the game AI & # 39; Alphastar & # 39; developed by Google's Deep Mind research team, such as researcher Orion Binyals, has reached the level of Grandmaster in StarCraft II. Covered the game. Nature provided
Artificial intelligence (AI) has conquered & # 39; Starcraft & # 39; after chess and Go. StarCraft is the oldest real-time science fiction strategy (RTS) game in the world.
The international magazine Nature published a study on the 14th of this month in which the game AI & # 39; Alphastar & # 39; developed by Google's Deep Mind research team, such as researcher Orion Binyals, has reached the level of Grandmaster in StarCraft II. Covered the game. Grandmaster levels are among the top 0.2%.
Chess and go are played according to certain rules. Thus, when AI learns many existing patterns, it predicts and responds to the number of cases the opponent will place in the actual game. But games are more complicated than Go or Chess and more complicated than two or more players. In particular, a strategy game like StarCraft is not easy because it predicts the opponent's strategy and at the same time elaborates my strategy so that the opponent doesn't understand.
AlphaStar is composed of synthetic artificial neural network (CNN), which analyzes visual images, and circular artificial neural network (LSTM) that can be predicted using previous data. Just as a person looks at the game screen with his eyes, he understands the situation, predicts his opponent's strategy, and develops his own tactics. We learned over 800,000 game playback data.
In addition, the DeepMind team trained AlphaStar with 16 different AI algorithms and 14-day iterations without human intervention. In human terms, it's like playing a game for about 200 years.
In January, Alphastar won a 10-1 victory over British pro players Darius Duncy (TLO) and Gregory Commints (MaNa). At the time, experts said Alphastar was a bit more advantageous because it played games with different interfaces.
Since then, DeepMind has updated AlphaStar to allow users to play with the same interface. And by the end of October, Alphastar reached Grand Master level in all three races including Terran, Protos and Zerg.