Blocky, addictive videogame Minecraft could help researchers train better artificial intelligence.
The game, conceived by Swedish developer Mojang, was acquired by Microsoft in late 2014 and it is now at the centre of the tech giant’s latest AI project. In July, the company launched the open source platform Project Malmo, which offers a modified version of Minecraft as a workshop for AI experiments.
Microsoft researchers think the game’s free-roam universe would be ideal for teaching budding AIs how to navigate complex situations similar to the world we live in.
“In many cases, to play a computer game, one only needs to pick up a single skill. That doesn’t get close to the variety of situations we encounter in the real world,” says Katja Hofmann, a researcher at Microsoft’s UK lab in Cambridge. “Minecraft is fantastic because it’s an open world game and you can play it in so many different ways.”
“The aspect we are really interested in is that, within the game, you have to interpret the environment and move to a goal. And you can also build things, but in a logical way: you can’t put a block in midair—you have to build the foundations, to build a roof, and so on.”
While Minecraft is great to teach machines how to walk around a crevice or how to hop up a hill, Hofmann underlines that the game is not accurate enough to run experiments where hyperrealistic laws of physics are a must. (So no, we probably will not see any chunky self-driving Humvee crisscrossing through Project Malmo.)
Still the mixed bag of actions available in Minecraft—players can talk, fight, search for and gather materials, using tools to craft objects and buildings— are exactly the kind of polymathic training necessary to sow the seeds of artificial general intelligence.
AIs deployed in Malmo’s Minecraft will also have the possibility of living faster lives, as a feature called “Overclocking” allows researchers to run trials in an environment moving at a quicker-than-usual pace.
“It is as if you were taking drugs and the world around you sped up, and you had to react more quickly,” Hofmann says. “ That will help researchers collect much larger amounts of data, and to train their AIs faster than would otherwise be possible.”
The option would only work for experiments dealing with discrete actions—such as moving quickly from one block to another, as opposed to walking smoothly to an arbitrary location in the space.
In many cases, to play a computer game, one only needs to pick up a single skill. Minecraft is fantastic because it’s an open world game and you can play it in so many different ways. Katja Hofmann
Beyond that, what might make the AI community excited about Project Malmo is how the platform enables different AI agents to interact among themselves and with humans in the same world. And sharing the same (blocky) testing ground will make it much easier for researchers to compare how different AIs perform in identical situations.
Hofmann believes that the project’s full potential is now for users to explore. After all, if Minecraft caught the eye of Microsoft researchers, it was also because of its thriving “modding” online community— showcasing how the game lent itself to elaborated designs and tasks.
“ We just put the Malmo platform out there on github” Hofmann says. “Let’s see how quickly people pick it up.”
Watch Katja Hofmann explaining the whole thing in a 40-minute talk
Follow Katja on Twitter
Co-working space and blog dedicated to all things data science.