Gamasutra: The Art & Business of Making Gamesspacer
View All     RSS
February 19, 2019
arrowPress Releases







If you enjoy reading this site, you might also want to check out these UBM Tech sites:


DeepMind AI faces off against (and defeats) StarCraft II pro players

January 24, 2019 | By Alissa McAloon

January 24, 2019 | By Alissa McAloon
Comments
    Post A Comment
More: Console/PC, Video



"StarCraft afforded some pretty unique research challenges that were, for us, the obvious choice as a next step for our agents to perform at a very good level."

- DeepMind's Oriol Vinyals explains StarCraft's appeal for AI research.

A DeepMind AI agent trained on the ins and outs of Blizzard’s StarCraft II went head to head with a pair of StarCraft Pro players to showcase the potential of the Google-owned artificial intelligence tech.

While DeepMind has had showdowns against human players in other games in the past, today’s live-streamed event offers developers an interesting look at how an AI agent takes on a complex RTS like StarCraft II, partnered with useful commentary on how those AI agents learn and improve their StarCraft II skills.

All this is made possible by a special version of StarCraft II Blizzard released a while back specifically geared toward training AI agents. Those agents play StarCraft a little different than actual human players. For one, AI agents are able to view the entire map at one time (though the fog-of-war effect still obscures inactive areas) so, unlike human players, decisions about what part of the screen to observe don’t come into play.

The face-off was streamed by DeepMind today and, though the bulk of matches themselves took place back in December, a panel of experts were on hand to offer moment-to-moment analysis of each individual battle. Ultimately, the DeepMind agent AlphaStar won all 10 of its pre-recorded matches, though a new version of AlphaStar lost to the pro player MaNa in a live match at the end of the stream.

Some interesting tidbits discussed in the stream include the fact that, despite inhuman reflexes and a semi-omnipotent view of the map, AlphaStar made fewer actions-per-minute and had a slower average reaction time than the professional players it went up against. More data like that can be found in the video above, though Engadget also has a strong writeup on the Stream’s events for those with a preference for text.

DeepMind researchers have used a number of games throughout the years but, as DeepMind’s Oriol Vinyals explains in the stream, StarCraft provides an excellent environment for AI agents to learn and evolve.

“First and foremost, the mission at DeepMind is to build an artificial general intelligence. To do so, it is quite important to benchmark how our algorithms, or our agents as we call them, perform on a wide variety of tasks,” said Vinyals. “And then, for us, StarCraft afforded some pretty unique sort of research challenges that were, for us, the obvious choice as a next step for our agents to perform at a very good level. And the one perhaps that I would highlight […] is the fact that in StarCraft you don’t see the board all the time, there’s this notion of imperfect information, so that means that an agent must kind of predict, estimate what the other player is doing all the time. And there’s all sorts of interesting things that emerge.”



Related Jobs

IllFonic
IllFonic — Tacoma, Washington, United States
[02.18.19]

Senior to Lead Engineer
Sony PlayStation
Sony PlayStation — San Mateo, California, United States
[02.18.19]

Sr. Program Manager
Sony PlayStation
Sony PlayStation — San Mateo, California, United States
[02.18.19]

Sr. eSports Manager (Competitive Gaming Content)
Curriculum Associates
Curriculum Associates — San Francisco, California, United States
[02.15.19]

Senior Software Engineer - Learning Games (Unity)









Loading Comments

loader image