Meta Arcade Trains AI to Adapt Through Video Games
Meta Arcade, a suite of arcade games that can be configured and used as training tasks for deep reinforcement learning, was created by a team at Johns Hopkins APL under the Defense Advanced Research Projects Agency (DARPA) Lifelong Learning Machines (L2M) program.
“We needed to develop a tool like Meta Arcade to study and advance our AI research,” said Bart Paulhamus , chief of APL’s Intelligent Systems Center. “By releasing it to the public, APL is accelerating the development of trusted AI for our nation’s most critical challenges. Now, AI researchers can focus their time on AI research, not tool development.”
Transcript
00:00:01 Meta Arcade is the framework that allows a researcher to construct a game that they can use to train an AI. Games are actually a really good tool for training AI because they are more simplified than the real world they also have a metric for progress. Meta Arcade comes with a set of pre-existing games and what's different about these games compared to something you might buy at the store to play as a human is that the games are configurable so you can change anything you want about the game you can change the colors, you can change how fast the game runs, everything about it is up to you. Rather than being hyper specialized on a single task the AI has to demonstrate something more it has to show that it can adapt to different games. The underlying game can change and it can continue to learn. It helps us explore all these aspects of tackling AI but we have some control over the world that the AI lives in.
00:01:04 In the past a lot of progress has been made on single tasks, so it'd be a specific video game controlling a specific robot. What researchers are looking at now is how can you create an AI That's more general, right? How can you play any video game? How can you move any robot? And that's why we needed a set of games that were similar enough that you could learn one and then transfer what you've learned to a new game, but we're still distinct and still had variety. We're building a framework to set up experiments where you have a lot of different problems that only a single AI is trying to solve. Most of the work, especially in research, is about tackling existing problems that's how you measure progress. But I think it's worth stepping back and thinking about, are we solving the right problems? The benchmarks that we've been using to track progress are becoming outdated. Meta Arcade is
00:02:03 saying to the community hey we needed to make something new because the existing problems that are out there, they're just not sufficient for the research questions that we're asking. We're starting in this realm of video games where maybe that's not super impactful now but if as soon as you move to another discipline where hey now our AI understands the situation better, is more capable than the best human authority, that's huge. And that's why games are so powerful for moving AI forward.