Tackling the board game Go was cool and all, but how will AI react to a Zerg rush? The fun to play, hard to master Starcraft 2 is becoming the new testbed for artificial intelligence. The rules are still simple. Beat the other guy/gal. But getting there will be full of methodical planning, bluffing and economy management.
At Blizzcon, DeepMind and Blizzard Entertainment announced a collaboration to open up the sci-fi world of Starcraft 2 to AI and Machine Learning researchers around the world.
In its heyday, Starcraft was the pinnacle of E-sports. It’s still popular today, but the explosion of competitive gaming has brought many more games to the forefront.
If you haven’t been following the world of AI, DeepMind is on a journey to push what is possible with AI. They develop programs that can learn how to solve the toughest problems without being told how beforehand.
Games are the perfect testing ground. Complex problems are already here. All researchers need to do is create an API to let the AI play it. And DeepMind plans to release it in the first quarter of next year. Check out how AI will play Starcraft 2 below.
AI’s will tackle problems ranging from build order and economy management to attacking. How will AI do the first time they encounter a feint or an all-out push? Can they adapt to overcome these challenges? Researchers will build systems using a combination of techniques to help AI take the skills honed on the board game of Go to the virtual domain of Starcraft 2.
But won’t AI quickly surpass what humans can do in Starcraft 2? Not so fast, says researchers. The new interface coming to Starcraft 2 is going to limit the machine-learning systems. Take the number of commands they can execute per minute. The best Starcraft 2 players already look damn near like robots anyways. But researchers want to match how fast a machine can execute commands closely to that of human players.
Why Starcraft 2? I’ll let Oriol Vinyals, a research scientist at DeepMind, explain:
“StarCraft is an interesting testing environment for current AI research because it provides a useful bridge to the messiness of the real-world. The skills required for an agent to progress through the environment and play StarCraft well could ultimately transfer to real-world tasks.”
Remembering will play an important role for machine-learning AI in Starcraft 2. You can see in the video above, one of the panels says ‘vision.’ Each player can only see parts of the map within range of its own units. When an AI scouts an area and discovers enemy units (or bases) it will need to remember that information and determine the best possible response.
Human players know certain types of buildings early in a match usually means a certain type of strategy. AI will learn the same over time.
DeepMind and Blizzard aren’t just throwing AI’s to the wolves here. They are working together to create what DeepMind calls “curriculum” scenarios. This will let researchers see how their AI’s perform against increasingly complex tasks. The API will also include the ability for researchers to create their own tasks using the existing Starcraft 2 editing tools.
Don’t expect machine-learning to turn pro in Starcraft 2 anytime soon. Remember, these AI’s are tackling the game without any previous knowledge. Its first games will look a lot like all of ours. A bunch of ‘what does this button do’ and ‘what’s this unit?’ But eventually, AI’s should have no problem taking on the best of the best.
Or, they’ll lose their virtual minds after hearing “not enough minerals” for the millionth time.