DeepMind's AlphaStar AI is one of the world's best StarCraft 2 Protoss players

AlphaStar, an AI designed by Google DeepMind specifically to master StarCraft 2, looks like it’s well on its way to succeeding. Fed on data from past matches and games with StarCraft 2 players, DeepMind thought the AI was up to the challenge of taking on a pro. It turns out it definitely was.

In a series of recorded matches from last year, it defeated StarCraft pros 10-0, though humanity got its revenge in a streamed exhibition match where it was finally defeated by Team Liquid’s Grzegorz "MaNa" Komincz.

AlphaStar is a Protoss player, but the AI’s first match-up was against Zerg pro, Dario “TLO” Wünsch. In a livestream where the matches were shown off for the first time, TLO said that he’d put himself in the top 1 percent of Protoss players, but wouldn’t say he was at the pro level. The AI won all five matches.

One of the most obvious advantages you’d expect an AI to have is reaction speed, but AlphaStar’s average APM was actually lower than TLO’s, and it won not through preternatural speed, but by taking advantage of strategies presumably picked up as it digested the data from countless matches. It helped, of course, that it had the benefit of more StarCraft 2 experience than any human could ever acquire—nobody has the time to watch all those matches—as well as greater coordination.

Unlike TLO, MaNa is one of the world’s best Protoss players, but again the AI prevailed. It’s a good thing it hasn’t been programmed to boast, or it would be all over Twitter, calling out humanity and mocking our skills. But while this was all apparently done under tournament conditions, it was all prerecorded. When it came to the live match, it buckled. AlphaStar had a stronger start, but it made some blunders that MaNa was able to take advantage of, eventually wiping the AI out. Now AlphaStar will presumably seek revenge, and the vicious cycle of AI-on-human violence will continue.

DeepMind says that the techniques used to make AlphaStar a great Protoss player can be applied to the other races, so expect to see AI battles with Zerg and Terran armies at some point, but there are also applications beyond fighting StarCraft 2 pros. DeepMind cites natural language processing and any complicated sequence modelling as areas where AlphaStar could potentially assist. Specific examples including things like weather prediction.

A Reddit AMA with DeepMind, MaNa and TLO will take place later today, and you can read a blog post from DeepMind on the AI here.

TOPICS

Fraser is the UK online editor and has actually met The Internet in person. With over a decade of experience, he's been around the block a few times, serving as a freelancer, news editor and prolific reviewer. Strategy games have been a 30-year-long obsession, from tiny RTSs to sprawling political sims, and he never turns down the chance to rave about Total War or Crusader Kings. He's also been known to set up shop in the latest MMO and likes to wind down with an endlessly deep, systemic RPG. These days, when he's not editing, he can usually be found writing features that are 1,000 words too long or talking about his dog.