AlphaStar at StarCraft II
On 19 December 2018 DeepMind's AlphaStar beat the professional Grzegorz "MaNa" Komincz 5-0 at StarCraft II, after the same result against Dario "TLO" Wünsch; the company told of both matches on 24 January 2019. On 30 October 2019 Nature published the paper: AlphaStar played at Grandmaster level with all three races.
Why it matters
Real-time strategy means imperfect information, play without turns and around ten to the twenty-sixth legal actions at each step. After the showing DeepMind itself removed the machine's advantages, and it was a version under human-like limits that earned a rating among people.
Limits DeepMind named in January: the matches were Protoss against Protoss on one map, CatalystLE; the agent saw the whole map through a raw interface without controlling a camera; in a live exhibition match MaNa beat a camera prototype trained for 7 days. AlphaStar averaged around 280 actions a minute. The agent league trained for 14 days on 16 TPUs per agent, up to 200 years of play each. In the Nature version the agent sees through a camera and is capped at 22 actions per 5 seconds (a cap agreed with TLO), played anonymously on Battle.net and, by the paper's abstract, ranks above 99.8% of officially ranked players. The record does not claim the day of the TLO match: DeepMind does not give it.