Back to Top

DeepMind AI has reached Grandmaster tier in StarCraft 2

The next StarCraft 2 esports star could be an AI, as one just achieved Grandmaster tier in the game. The artificial intelligence, built by Google’s DeepMind, is now in the top percentile of StarCraft 2 players, and the first AI to master a videogame without needing a simplified version.

Known as AlphaStar, the AI made headlines in January when it beat two StarCraft 2 pros very convincingly. The victories, though impressive, were over-shadowed by the fact AlphaStar had some slight advantages, namely the amount of meaningful actions it could execute, and a far greater view of the entire field-of-play, over its human opponents. Aware of these criticisms, DeepMind has been working on ways to make the AI learn just as quickly using the same resources that we have, and managed it with AlphaStar recently mastering the entire game.

In a new DeepMind blog (via ArsTechnica), the team explains the process. Using a mix of reinforcement learning, whereby the program plays a version of itself over and over, imitation learning, involving the copying of existing game data, and other machine-learning techniques including playing humans in regular match-making, AlphaStar has become completely proficient playing as or against any of Protoss, Terran, and Zerg. This is to the point the AI is now in the top 99.8% of all StarCraft II players, playing through the Battle.net interface with added constraints so the experience is as similar to ours as feasible.

The project was developed in collaboration with members of the StarCraft 2 community. StarCraft pro Dario “TLO” Wünsch was heavily involved, and is credited as helping develop AlphaStar’s interface by the team. “I’ve found AlphaStar’s gameplay incredibly impressive – the system is very skilled at assessing its strategic position, and knows exactly when to engage or disengage with its opponent,” Wünsch said. “And while AlphaStar has excellent and precise control, it doesn’t feel superhuman – certainly not on a level that a human couldn’t theoretically achieve. Overall, it feels very fair – like it is playing a ‘real’ game of StarCraft.”

According to the DeepMind blog, there’ll be more work in games using AI like this in the future: “At DeepMind, we’re interested in understanding the potential – and limitations – of open-ended learning, which enables us to develop robust and flexible agents that can cope with complex, real-world domains. Games like StarCraft are an excellent training ground to advance these approaches, as players must use limited information to make dynamic and difficult decisions that have ramifications on multiple levels and timescales.”

You can read the full paper on AlphaStar here, and you can view all AlphaStar’s matches through Battle.net here. In other StarCraft news, a cartoonified remaster of the first game was released as part of the 20th anniversary recently.

Back to Navigation