In 2015, Google created its artificial intelligence program DeepMind and taught it to play Go, an ancient Chinese board game. This game for two players is a kind of mix between chess and checkers, so it is relatively easy to learn and only includes a few possible actions.
DeepMind defeated Go’s European champion, Fan Hui, in five consecutive games without disheveled. Later, Google created a new version of the DeepMind system even better, which beat the previous algorithm in 100 consecutive games. It was clear by then that DeepMind had mastered the game Go.
A new challenge: Master the strategy game Starcraft II
Overcoming Go’s challenge, Deepmind has set its sights on a totally different game, the real-time strategy game StarCraft II: Wings of Liberty 2010. In StarCraft II, players have to control armies across terrain. They have to build infrastructure, maintaining a balance between short-term profits and long-term benefits. A potion of the map becomes visible only if you have your troops on that region. So players must act on intuition in many cases.
Today, some games between DeepMind and two professional players have been broadcast live on YouTube and Twitch. In two series of the best of five games, the DeepMind AI has defeated the two expert players by 10-1.
“This is, of course, an exciting time for us,” David Silver said in the live broadcast followed by more than 55,000 people.
“For the first time we saw an AI capable of defeating a professional player,” he added.
DeepMind created five versions of its AI, called AlphaStar, and trained them with game material between humans. The different AIs then played against each other in a league, accumulating the best of the five the equivalent of 200 years of gaming experience. Thanks to this experience, AlphaStar has managed to beat professional players Dario Wunsch and Grzegorz Komincz, who occupied the 44th and 13th positions in the world respectively.
Dominate Starcraft II serves more than just to sign up
The success of AlphaStar comes with some considerations: The AI is only able to compete on a single map and only knows how to handle a race. Professional players had to deal with different versions of AlphaStar in each game, so they found it difficult to get used to their style of play.
“I did not expect something like that,” says Niels Justesen, of the University of Information Technology in Copenhagen, Denmark. “Especially because previous attempts to learn to play StarCraft from beginning to end have been far from the human level,” he added.
“I think StarCraft is like running a business, particularly a logistics operation,” says Julian Togelius of New York University. “It’s about planning the research and development, and taking the products to the right place at the right time, avoiding bottlenecks.”