DeepMind, the Google-owned AI startup behind AlphaGo, is collaborating with Blizzard to open the popular Starcraft II game to AI and machine learning research. The goal is to foster advancements in artificial intelligence to the extent that it can solve complex problems without human guidance.
DeepMind writes that games such as Starcraft II, Go, and Atari provide provide ideal training grounds for AI, “allowing us to develop and test smarter, more flexible AI algorithms quickly and efficiently, and also providing instant feedback on how we’re doing through scores.” What makes Starcraft II an interesting testing ground for AI is that it, to a certain extent, replicates “the messiness of the real world”, forcing agents to dynamically navigate challenges while responding to new pieces of information and prioritizing accordingly.
In Starcraft, players select one of three races, each of which possess distinct units and abilities, which in turn determine their respective optimal gameplay approaches. Players must also make economic calculations, harvesting minerals and gas in order to produce armies and buildings, and judiciously allocating their limited resources while doing so.
In contrast to games such as chess and go where players have complete visibility over the playing field (i.e. perfect information), Starcraft players can only see the parts of the map that are within range of their units and buildings, meaning they are largely unable to see what their opponents are doing. These facets of gameplay mean that successful players must “demonstrate effective use of memory, an ability to plan over a long time, and the capacity to adapt plans based on new information.”
In sum, to successfully navigate this environment, AI agents coordinate multiple factors (e.g. their mouse clicks, vision, and available resources), learn through trial-and-error, and plan hierarchically. These are all challenges that a deep reinforcement learning approach is meant to address.
The environment the two companies are collaborating to develop will be open to all researchers by next year. It features an API that allows AI agents to control individual units, and an image-based interface that represents feature layers (including map, minimap, unit type, terrain heightfield, and unit health) as low-res RGB images. To test AI agents’ abilities to compete intelligently against professional players rather than through sheer speed, the number of “actions per minute” they can undertake will be constrained within the limits of human dexterity and agility. While DeepMind and Blizzard have partnered to create “curriculum” scenarios for researchers train their agents, third-parties can also develop their own scenarios and tasks.