Then, Blizzard will release a series of SC2 replays of professional and amateur games to give the AIs a dataset to learn from visual information. Image Source: Google

U.S. Anaheim, Calif. – Google Inc. (NASDAQ: researchesGOOGL) owned DeepMind machine learning startup is collaborating with Blizzard game studio to create the ultimate artificial intelligence research environment inside StarCraft II.

DeepMind will not build an unstoppable AI on its own. Instead, both companies will release a series of programming tools on early 2017 that will allow researchers and hobbyist around the world build and train their bots to play Starcraft II.

Google’s DeepMind researcher Orion Vinyals made the announcement during BlizzCon 2016 at Anaheim, California. Vinyals was the top-ranked SC2 player in Spain’s leaderboards before becoming a top scientist in the British AI startup. Vinyals believes the results of the investigation could translate to the real life.

“The skills required for an agent to progress through the environment and play StarCraft well could ultimately transfer to real-world tasks,” said Vinyals.

How will they start the research?

DeepMind will release two APIs to kickstart the research: one to turn SC2 into basic visuals that can better feed a machine learning system, and a second one to improve how bots can read game data.

Then, Blizzard will release a series of SC2 replays of professional and amateur games to give the AIs a dataset to learn from visual information, just like humans do.

Any researcher and active member of Starcraft community can get the APIs and the replays to contribute to the AI efforts and build their agents.

Both companies agree that there is still a long way to go before an AI could defeat a top-ranked SC player, but the foundations have been laid for the ultimate showdown between pro-gamers and bots.

DeepMind shocked the world back in March when its AlphaGo AI defeated Go world-champion Lee Se-dol

Go is a puzzling 2,500-year-old Chinese board game. It requires highly advanced degrees of intuition to play at a dangerous level, and AlphaGo managed it with a system built on neural networks and machine learning.

StarCraft II is not Go, though. Blizzard Entertainment’s real-time strategy hit is one of the most fiercely competitive games around. Mastering the game requires thinking fast and ahead of the opponent, as well as being able to manage many things at the same time.

Gamers call it micromanaging vs. macromanaging: the micro is all about moving individual units, scouting and outmaneuvering the opponent during the battles; and the macro refers to gathering resources building bases, improving the army and planning the economy.

googledeepmind-ai-personal-assistant-talks like humans
Google and DeepMind develop a new AI personal assistant. Image credit: 1 red Drop.

Creating high SCII AIs would have significant implications for the game

The research has the potential of creating customized virtual coaches and AI opponents to raise the level of play.

The game’s present AI system operates very different to DeepMind’s neural networks. It is hand-crafted and works differently at each difficulty level. It also “cheats” because it has access to information a human player could not have.

For example, an AI can see the entire map and issue orders to all units at the same time. So, the challenge is to create a StarCraft II AI player that is stronger than a professional gamer while having the same limitations.

Source: The Verge