DeepMind's StarCraft 2 A.I. Ranks Higher Than 99.8% of Human Players

DeepMind's StarCraft 2 A.I. Ranks Higher Than 99.8% of Human Players

If you lost against AlphaStar, we're sorry.

A few months ago, an A.I. racked up 90 games of StarCraft 2—30 each as the Protoss, Terran, and Zerg—and managed to reach the Grandmaster rank for all three races. The A.I., dubbed AlphaStar, is the product of years of research conducted by DeepMind. Today, DeepMind went public with AlphaStar by way of a new research paper in Nature and a more digestible summary on its website.

DeepMind's final iteration of AlphaStar ranked above 99.8% of all active StarCraft 2 players. As the highest achievable rank in StarCraft 2, Grandmaster rank consists of the top 200 one-versus-one players in a given region (AlphaStar's games were played on European servers).

It took a novel innovation in machine learning techniques in order to get AlphaStar into Grandmaster. A common problem with A.I. designed to converge on better strategies is that they can also effectively "forget" what it has learned and then lose to past versions of itself. In a league where every A.I. agent is trying to win, that can lead to a situation where the A.I. end up stuck cycling through a few strategies that happen to beat each other. To prevent this, DeepMind reworked its AlphaStar A.I. training league to include "exploiter agents" that prioritize seeking out flaws in the strategy of win-maximizing agents.

It's an approach inspired by how actual top players train, and AlphaStar benefitted from the input of StarCraft 2 pros. Team Liquid's Dario "TLO" Wünsch is listed as a co-author on the paper: in addition to playing matches against AlphaStar for the purposes of research, Wünsch helped inform the limits placed on AlphaStar to ensure it isn't discernibly making moves faster than a human player possibly could.

In 2017, DeepMind partnered with Blizzard to make an open source environment for A.I. research within StarCraft 2—all 90 of AlphaStar's game replays and an open access version of the new paper are also available for free download. DeepMind is owned by Alphabet, the parent holding company of Google.

As always, the key takeaway of research like this is the breakthrough in methodology. A few good players did have to lose games against AlphaStar in the wild for the sake of science, but if it's any consolation, those folks are still better than AlphaStar at doing literally anything besides playing StarCraft 2.

Sometimes we include links to online retail stores. If you click on one and make a purchase we may receive a small commission. See our terms & conditions.

Mathew Olson

Reporter

Mathew Olson is a writer formerly of Digg, where he blogged and reported about all things under the umbrella of internet culture (including games, of course). He lives in New York, grew up under rain clouds and the influence of numerous games studios in the Pacific Northwest, and will talk your ear off about Half-Life mods, Talking Heads or Twin Peaks if you let him.

In other news

The Last of Us Part 2 and Hades Lead 2020's Game Awards Nominees

Supergiant and Naughty Dog have racked up the nods at this year's Keighley-tastic show.

Frost Giant Studios Is Another New Team Formed by Ex-Blizzard Employees

The founders of Frost Giant hope to make an RTS that'll resonate with a wide audience.

The Pikmin 3 Deluxe Demo Has a "Hidden" Boss, Speedrunner Discovers

There's a bit more of Pikmin 3 in that demo than intended, it seems.

You may also like

Cuphead's Delicious Last Course Pushed Back Until It's Ready

The final sip will need to steep a while longer.

Super Mario Maker Support on Wii U is Coming to an End Next Year

That means it'll also be removed from the eShop soon.

Microsoft Is Working to "Identify and Resolve" Performance Issues in Xbox Series X Games

A Microsoft spokesperson says that developers are "just now scratching the surface" of what the new Xbox consoles can do.

Metro Exodus Is Getting a Next-Gen Version Next Year

4A Games is also working on the next Metro and exploring potential for multiplayer.