How DeepMind are planning to conquer StarCraft 2 with a learning AI

StarCraft 2 DeepMind AI

One of the stranger announcements at yesterday’s BlizzCon opening ceremony was DeepMind, the Google subsidiary behind the AlphaGo computer which defeated Go grandmaster Lee Sedol in March, would be attempting to defeat a human opponent in the near future. In today’s panel on the subject, DeepMind’s lead research scientist Oriol Vinyals explained how and why the task is so difficult, as well as announcing the API for StarCraft 2 would be made available for other AI researchers early next year.

For more of the news from behind the curtains at BlizzCon 2016 see the news hub here.

Vinyals said that, though complex games like chess and Go had been cracked by AI, StarCraft introduces several levels of concepts above what computers have had to deal with before.

Fog of war introduces the concept of imperfect information, the game’s economy involves resource management and trade-offs for expansion and defense, and everything takes place on a large action space requiring hierarchical actions. However even basic features of the game present challenges to a completely blank AI, such as the strengths and weaknesses of three asymmetrical races, weighing up long-term goals against short-term micro wins and the realtime simultaneous actions of player and foe.

Kevin Calderone, senior software engineer from Blizzard, explained that DeepMind uses an unscripted form of AI that depends on learning at a far accelerated rate compared to humans, but still using a lot of the same processes we do.

To help it get to terms with all these aspects of StarCraft, the team have set up a learning environment for the AI – where it sends units out, observes their progress using several different metrics as seen above, analyses the success of their actions and evaluates what it can do better – as well as letting it analyse replays from actually good human players.

Here’s where you, the humble player at home, can be a part of the initiative as Calderone says just by playing on the ladder, your replays can get added to DeepMind’s repository of data to sift through in order to pick out winning – or losing – patterns of behaviour.

Calderone mentioned the current AI-programming scene for Brood War – something Blizzard has been strangely quiet on since announcing their team-up with DeepMind to make an AI that can play StarCraft – would benefit from the official StarCraft 2 API being made available in Q1 2017 if they wanted to try to port their progress across to the series’ flagship title.

AI researchers have been engaging in a silicon war of their own at the Artificial Intelligence and Interactive Digital Entertainment’s annual StarCraft AI tournament for more than half a decade, however showmatches against human opponents so far haven’t yielded any supreme bots yet.

“While we’re still a long way from being able to challenge a professional human player at the game of StarCraft II,” Vinyals said, “we hope that the work we have done with Blizzard will serve as a useful testing platform for the wider AI research community.”

PCGamesN logo Free newsletter