Making AI Play Lots of Videogames Could Be Huge (No, Seriously)
Source: Jake Muncy
It’s almost a given that you’ll ride in an autonomous car at some point in your life, and when you do, the AI controlling it just might have honed its skills playing Minecraft.
It sounds crazy, but open-world games like Minecraft are a fantastic tool for teaching learning algorithms—which power the next generation of advanced artificial intelligence—how to understand and navigate three-dimensional spaces. Achieving that is a major stepping stone toward creating AI that can interact with the real world in complex ways.
It’s easy to consider videogames mindless escapism, but because they generate such vast amounts of information—think of the expansive world players create in Minecraft—they are exceptionally well suited to teaching an AI how to perceive the world and interact with it. “It’s hard for a human to teach AI,” says Xerox researcher Adrian Gaidon, because they are “worse than the worst toddlers in the world—you have to explain everything.”
Beyond a certain point, humans just don’t have the time and patience to teach an AI how to behave. Videogames don’t have that problem. You may grow frustrated with them, but they never grow frustrated with you.
Fooling an AI
Researchers typically teach the so-called “deep learning” algorithms that underpin modern artificial intelligence by feeding them staggering amounts of data. These systems gorge on information, seeking patterns. If you want to teach an AI like AlphaGo to play Go, you feed it every record of every Go game you can find. For something like a board game, this is the easiest part of the task. The machinations of even the most complex board game can be rendered pretty easily by a computer, allowing AlphaGo to learn from a sample size in the millions.
For more complex tasks like, say, driving an automobile, gathering enough data is a huge logistical and financial challenge. Google has spent untold sums testing its autonomous vehicles, racking up millions of miles in various prototypes to refine the AI controlling the cars. Such an approach isn’t feasible for researchers who don’t have the limitless resources of a company like Google or Baidu. That makes videogames increasingly appealing. You can gather vast amounts of data relatively quickly and cheaply in a game world.
This idea came to Adrien Gaidon about 18 months ago when he saw a trailer for the latest installment of Assassin’s Creed. “I was shocked, because I thought it was the trailer for a movie, whereas it was actually CGI. I got fooled for 20 seconds, easily. It’s the first time that happened to me.”
If modern game engines could so easily fool him, he thought, perhaps they could fool an AI, too. So he and his team at Xerox started using the videogame engine Unity to feed images of things like automobiles, roads, and sidewalks to a deep-learning neural network in an effort to teach it to recognize those same objects in the physical world.
Researchers have seen success with this. Before tackling Go, Google’s AI mastered Atari games. Other AI projects have conquered Super Mario World levels. Using game engines with three-dimensional rendering, and training AI within those spaces, however, represents a level of complexity that’s only recently become possible.
“The real benefit of a game engine is that, as you generate the pixels, you also know from the start what the pixels correspond to,” Gaidon says. “You don’t just generate pixels, you also generate the supervision [AI] requires.”
So far, Gaidon says his work at Xerox has been very successful: “What I’m showing is that the technology is mature enough now to be able to use data from computers to train other computer programs.”
Embodied in Minecraft
Microsoft also sees the value in this. It recently announced that later this year it will release Project Malmo, an open-source platform that “allows computer scientists to create AI experiments using the world of Minecraft.” Beyond its complexity and open-ended freedom, Minecraft offers new ways of experimenting with AI embodiment, says Katja Hofmann, Project Malmo’s lead researcher.
“When you play Minecraft, you are really directly in this complex 3-D world,” Hofmann says. “You perceive it through your sensory inputs, and you interact with it by walking around, by placing blocks, by building things, by interacting with other agents. It’s this kind of simulated nature that’s similar to how we interact with the real world.”
Hofmann and her team hope their tools push research in even more radical directions than Gaidon’s team is pursuing. Using skills learned in a program like Malmo, AI could, she believes, learn the general intelligence skills necessary to move beyond navigating Minecraft‘s blocky landscapes to walking in our own. “We see this very much as a fundamental AI research project, where we want to understand very generically how agents learn to interact with worlds around them and make sense of them,” she says. “Minecraft is a perfect spot between the real world and more restricted games.”
The transition from simulation to reality is complex, though. Avatars in games typically don’t move like real people move, and game worlds are designed for ease and legibility, not fidelity to real life. Besides, the basics of how any agent, human or otherwise, builds their understanding of spatial reality remain something of a mystery.
“We’re really at the very early stages of understanding how we could develop agents that develop meaningful internal representations of their environments, says Hofmann. “For humans, it seems like we make use of integrating the various sensors we have. I think linking various sources of information is one of the interesting research challenges that we have here.”
“The Hallucinations of Sensing Machines”
When science finally figures out just how AI develops an internal representation of a given environment, people might be surprised at what form it takes. It may look like nothing ever seen before. “This may look very different from what actually happens in our brains,” Hofmann says.
This should come as no surprise. Humans wanted to fly, but achieving it looked nothing like how birds fly. “We are inspired by how birds fly or how insects may fly. But what’s really important is that we understand the actual mechanisms, how to create the right pressures, for example, or the right speed in order to lift an object off the ground.”
And so it will be with AI. Computers already view the world in a fundamentally different way than humans. Take, for instance, recent work by London’s ScanLAB Projects revealed how the laser-scanner “eyes” of an autonomous car might view a city. The results are utterly foreign, a “parallel landscape” of ghosts and broken images, urban landscapes overlain with “the delusions and hallucinations of sensing machines.”
Likewise, as Google’s recent showcase proved, AlphaGo understands the ancient game of Go in a way no human ever could.
What, then, will the world look like when viewed by the next generation of “sensing machines?” The models, methods, and technologies built out in algorithms by experience in virtual space—what will they see when applied to our cities, our parks, our homes? We’re teaching AI to understand the world in more robust ways. Videogames can help these machines reach that understanding. But when that understanding comes, we might not recognize it.
Correction appended [4:45 P.M. PT 4/18]: A previous version of this story incorrectly spelled Katja Hofmann’s name.
| }
|