DeepMind’s StarCraft II AI Will Play Public Matches on Battle.net

Google’s DeepMind division has been hard at work applying artificial intelligence to problems like computer vision and climate change, but there’s still some room for games. DeepMind first dominated the game of Go, and then it moved on to StarCraft II, beating some of the world’s top players early this year. Now, you might have a chance to play against the AlphaStar AI, but you’ll probably get wrecked.

Before challenging pro players, DeepMind simulated more than 200 years of StarCraft II gameplay to train the bot. It’s a convolutional neural network that started by absorbing replays of pro StarCraft II matches. Using competing models, DeepMind trained several “agents” that can build and battle as well as human players — better, in fact. AlphaStar won 10 of its 11 matches against pro players.

The previous matches were an impressive display of AI prowess. AlphaStar understood resource allocation, cover, and unit micromanagement better than most human players. Naturally, its ability to seamlessly control multiple units also helped. Although, it didn’t have to take as many actions as human players to win.

The experiment will take place on Blizzard’s European servers where a small number of humans will be matched with AlphaGo in 1v1 play. They won’t know that, but players have to opt-in to have a chance of going up against the AI. Sadly, there is no way to ensure you get to play a match against AlphaStar.

Blizzard will let DeepMind run several different agents on Battle.net, and they’ll operate differently than the last demo. The new AlphaStar will be able to play as or against any of the three StarCraft II races (it was Protoss only before). It also relies on a normal camera view of the game, whereas the previous agents used a bird’s eye view of the entire map. DeepMind has capped the actions per minute (APM) of AlphaStar as well.

DeepMind is primarily interested in testing AlphaStar in matches where players vary their strategies, and keeping the matching secret ensures a controlled test. After presumably being murdered by the AI, players will see their ranking affected as if they’d played a human opponent. DeepMind will use the results of this test to inform future AI research, and the results of the matches will appear in a forthcoming scientific paper.