After training an artificial intelligence model to master chess and Go, Alphabet Inc.’s DeepMind subsidiary has embarked on a new project: cracking the code on a hit video game seen by researchers as one of the “grand challenges” of AI.

The group has built a dedicated deep learning system called AlphaStar for the game in question, StarCraft II, that it detailed late Thursday.

The system is hailed as the most sophisticated of its kind to date. In a series of matches that DeepMind held last month, AlphaStar became the world’s first AI to beat a professional-level human player without any game restrictions.

What makes it such a significant milestone is the complexity of StarCraft II. Unlike chess and Go, the game provides a so-called imperfect information playing environment where certain key details are hidden. This makes it conceptually more similar to the tasks AI models must handle when they’re used for real-world, practical applications.

StarCraft II has a complex strategy element that ups the difficulty level even further. Matches take place on a three-dimensional map where two opposing players collect resources, construct buildings and assemble virtual armies in an attempt to overrun one another.

Winning under these conditions requires an AI to master skills such as long-term planning that are also necessary for some real-world deep learning use cases. “Like many real-world problems cause-and-effect is not instantaneous,” DeepMind’s researchers explained. “Games can also take anywhere up to one hour to complete, meaning actions taken early in the game may not pay off for a long time.”

DeepMind trained AlphaStar in two phases. First, it fed the AI footage of matches played by human players. The group then put AlphaStar in charge of a virtual StarCraft II league and tasked it with generating artificially intelligent competitors to play against each other.

“As the league progresses and new competitors are created, new counter-strategies emerge that are able to defeat the earlier strategies,” DeepMind researchers wrote. “While some new competitors execute a strategy that is merely a refinement of a previous strategy, others discover drastically new strategies consisting of entirely new build orders, unit compositions, and micro-management plans.”

In the match series that DeepMind held last month, AlphaStar’s agents beat the two professional-level human players who took part 10 to 1. The Alphabet subsidiary plans to share some of the methods that facilitated this victory with the broader AI community in a future academic paper. The idea is to enable researchers outside DeepMind to harness AlphaStar’s training techniques in  their own projects.

Photo: DeepMind