DeepMind Technologies Inc. made an unexpected appearance at last year’s Blizzcon, an annual convention held by legendary game studio Blizzard Entertainment Inc. for its fans. In between presentations for hit games such as Overwatch and Hearthstone, DeepMind took the stage to announce that it was launching a joint project with Blizzard that would turn Starcraft II into a research environment for artificial intelligence.

Now after months of silence, that project is finally here. DeepMind announced today that it is releasing the Starcraft II Learning Environment, or SC2LE, a toolset that makes it easy for AI developers to use Starcraft II to test their AI. The toolset includes a machine learning API developed by Blizzard, a dataset of anonymized game replays, and a series of reinforcement learning mini-games for AI benchmarking. SC2LE also includes an open source version of PySC2, a Python component that allows developers to use Blizzard’s machine learning API as a Python RL Environment.

In a blog post, DeepMind explained that Starcraft II is a perfect fit for testing AI because it offers a wide range of challenges, and an AI has to prioritize its objectives in order to be successful.

“For example, while the objective of the game is to beat the opponent, the player must also carry out and balance a number of sub-goals, such as gathering resources or building structures,” the DeepMind team explained. “In addition, a game can take from a few minutes to one hour to complete, meaning actions taken early in the game may not pay-off for a long time. Finally, the map is only partially observed, meaning agents must use a combination of memory and planning to succeed.”

Image: DeepMind Technologies Inc.

DeepMind said that SC2LE’s mini-games, which are designed to test specific skills, have shown promising results in training and evaluating AI, but the company said that its strongest AI agents could not win a full Starcraft II match against even the easiest opponents.

During one match, the AI failed to ensure that some of its units were still gathering resources, which the team said is “a task that humans find trivial.” According to DeepMind, the AI did eventually learn to perform more actions after further training, but in order for an AI to be competitive with human players, the team said that there will need to be “further breakthroughs in deep RL and related areas.”

DeepMind’s new Starcraft project is not the first game-based research platform for AI. For example, Microsoft Corp. released a tool last year that allows developers to test out their AI using the popular game Minecraft.

Video games have become a popular testing ground for AI because they allow developers to control a number of variables while still giving the AI the freedom to interact with a semirealistic environment.

Photo: Starcraft II | Blizzard Entertainment