Deepmind AI by Google takes on Starcraft by Blizzard after mastering Go

Blizzard and Google are coordinating to open up Starcraft to AI and machine learning researchers around the world. A Starcraft API allows for commands to be delivered to the Starcraft engine. The commands are restricted to the human range of clicks per minute. The API is expected to be made available to researchers next year. Intelligent agents will be using simplified image layers from the game, directly learning about the current state of the game from these images. Here is how the AI “sees” a game.

A game like Starcraft represents the next step for the progress of artificial intelligence. As against previous games handled by machines, such as Chess or Go , all the information of a current game is not available to the machine. Scouts have to be sent to explore regions, and the map keeps changing over the duration of the game. Additionally, the AI will have to train itself to rapidly adapt its plans based on the actions of the opponent. [caption id=“attachment_345538” align=“aligncenter” width=“640”]  An RTS game poses some unique challenges for AI.[/caption] The intelligent agent will also have to demonstrate effective memory and the spontaneous generation of long term plans. Starcraft has been chosen as the game to make a Deepmind bot out of, because of its long shelf life as a competitive e-sports game. Starcraft was one of the first games to have highly paid elite professional players, and the Starcraft gaming scene remains active after 20 years of the original release of the game, demonstrating the technical soundness of the underlying design of the game. The game-like scenario provides a bridge for potential real world applications of Google Deepmind in the future. [caption id=“attachment_316058” align=“aligncenter” width=“640”]  The Tensor Processing Unit[/caption] The Google Team will be working with Blizzard to create “curriculum scenarios,” which are training missions to get an AI up and running by researchers of any skill capability. Different algorithms and advances can be tracked and benchmarked for their effectiveness. Additionally, researchers will be allowed to generate their own scenarios, using editing tools built into Starcraft II. Google Deepmind works on Google’s own Tensor Processing Unit .