2 min

Tags in this article

, , ,

Google sister company DeepMind has trained a self-learning computer that is very good at the real-time strategy video game StarCraft II. The Artificial intelligence (AI) defeated 99.8 percent of human players.

StarCraft is a real-time strategy game, in which gamers have to collect resources, deploy troops and fight against opponents. It requires both tactical reactions and comprehensive and long-term strategies.

It is precisely because of all these different elements, like the strategic and tactical aspects, that these kinds of games form an interesting platform to test AI. DeepMind did this using a specially developed AI. It concerns AlphaStar, according to the Financial Times.

AlphaStar is now proving to be exceptionally successful. The AlphaStar bots ended up in the StarCraft League at the ‘grandmaster’ level. This means that the AI outperformed 99.8 percent of human players. Moreover, human gamers often didn’t even know they were playing against an algorithm.

What’s the point?

As mentioned before, AI’s are often tested in games. Previously, AlphaGo, another AI from DeepMind, won, for example, versus world champion Lee Sedol in the board game Go. But mastering StarCraft may be a more interesting use case than winning Go.

In StarCraft, the AI does not have perfect information about what the opponent can and will do. That’s why AlphaStar behaves more like how it would behave in the real world, which means that an AI like this may also have applications out of video games.

Research leader Oriol Vinyals explains that AlphaStar’s neural network architecture can set out a very long series of possible actions. It has to, because games can take up to an hour and consist of tens of thousands of options, based on imperfect information. According to Vinyals, what has been learned with AlphaStar can also be used for weather forecasts and personal assistants, for example.