Neuroscience-inspired AI company DeepMind will be partnering up with StarCraft II this Thursday for a special livestream that will showcase what researchers have learned from AI training with the RTS game.
DeepMind says games "are core" to its research programme and in 2017, it partnered with Blizzard to release open-source tool PySC2 to AI and machine learning researchers around the world. RTS games are of interest to AI researchers because they "pose hard problems to solve in areas of computer science such as planning, dealing with uncertainty and spatial reasoning". Consequently, StarCraft II has "emerged by consensus of the community" as a "grand challenge for AI research".
According to DeepMind, simulated game environments are widely used by the AI community because they provide the right kind of platform for developing and testing smarter, more flexible AI algorithms quickly, particularly as they offer "limitless training data, measurable progress, and huge parallelism". It also allows researchers to measure the generality of their algorithms, which is a good indicator of if, and how, they might transfer to other complex challenges.
"StarCraft is a game where, just like rock-paper-scissors, there is no single best strategy," a press statement said. "As such, an AI training process needs to continually explore and expand the frontiers of strategic knowledge."
At BlizzCon 2018, research scientist Oriol Vinyals shared the progress that DeepMind had made since PySC2 was released, including an update on how its AI was progressing in StarCraft II, starting by grasping the basic rules of the game and performing basic macro-focused strategies.