A team of computer gurus representing the College of Engineering (CoE) and the College of Computing & Informatics (CCI) is tackling one of DARPA’s fringiest competitions yet—a grueling, three-year test of time, energy, and frustration to conceive new ways to expand the nation’s electromagnetic spectrum by teaching machines not only to manage the spectrum, but to collaborate with each other while doing it.
Called the DARPA Spectrum Collaboration Challenge (SC2), the first-of-its-kind contest seeks to avert a looming crisis in the radio frequency spectrum as countless wireless-enabled devices overwhelm some assigned frequencies while leaving others standing by idly.
SC2 is about abundance: creating more of something where there is less and using all of it more efficiently. Think of it as roughly analogous to asking your napping teenager to walk the dog and make dinner tonight because you’re far too busy to do it yourself.
Last year, during the second year of the challenge, Drexel’s “Dragon Radio” team took 8th place globally, securing impressive bragging rights and a chance to move forward through the next few stages. Now in its final year, the contest has become increasingly complex. DARPA—the Defense Advanced Research Projects Agency in Washington, DC—has been turning up the heat on the contest specs, and the Drexel team has seen its standing diminish as a result.
Not that they’re going anywhere. They are in it, team members said, until the bitter end this fall.
“DARPA actually gave a talk at its last set of meetings about the nature of their competitions,” said Kapil Dandekar, professor in CoE’s Department of Electrical and Computer Engineering and the Dragon Radio team leader. “They like to set the bar very high. So high that the most typical outcome of the competition is failure. You don’t do what you set out to do. But in trying, you develop technologies that are ultimately useful.”
Out of 30 international teams with the wherewithal to enter the contest in December of 2017, only 15 remain. Drexel is one of them. With funding from the National Science Foundation, the Drexel team is headed into the final competition at the Mobile World Congress Americas in October, where a $2M prizewinner will be named.
Along with Dandekar, faculty team members include ECE Department Head Steven Weber; and CCI’s Geoffrey Mainland, associate professor of computer science and lead software developer for the team, who is doing the “lion’s share” of work on radio development. A revolving cast of undergraduates and PhD candidates filtering through the Drexel Wireless Systems Lab step in as needed. The team is always recruiting new members.
In SC2, teams succeed when their radio design communicates and cooperates better in the congested wireless spectrum than other competitors. This rising-tide-lifts-all-ships approach sets the stage for groundbreaking innovation, even as the radios battle for supremacy in an autonomous arena called the Colloseum. Metrics, feedback from DARPA, and scores that measure data flows on test events have all provided ways for teams to gauge their progress over the past two-and-a-half years.
“It’s really about making the most intelligent radio system that ever existed in history,” said Kyunglok Baik, an undergraduate in ECE. “Being a pioneer in the field—that’s the point of the challenge, and to solve it is our mission.”
Alex Lackpour, a PhD candidate under Dandekar, added, “It takes a long time to do this. It’s a very hard problem that we’re trying to solve. This is stuff that we’ve never done before. It’s never been done. And what DARPA does is hold our feet to the fire while we’re trying to do it.”
Spectrum: a finite resource
The electromagnetic spectrum accommodating the nation’s wireless connectivity needs—from laptops to smart kitchen appliances to TV networks to underwater marine sensors to radar, and on and on—is divided into regional and geographic bands in the spectrum. So, spectrum must be managed like any other national resource. However, the divisions—many of them set by the US government nearly a century ago—are inflexible and outdated. Certain frequencies get overwhelmed during peak use, impairing both speed and performance, while others languish.
The strain on the system is beginning to tell. This unsustainability predicts a looming scarcity of spectrum and “a serious problem for our nation,” according to DARPA’s contest website.
SC2 is an attempt to find a solution, focusing the brightest minds in software defined radio and artificial intelligence (AI) on the challenge. One contest goal is to imagine new ways to access frequencies that are underutilized to provide band relief; another is to shift management of that access to machines and AI capability that would increase efficiency by many orders of magnitude.
“So, the starting point was this static human in the loop who comes up with this elaborate plan about how frequencies can be used over here so this system won’t interfere with that system over there,” said Lackpour. “SC2 says, instead of doing that manually with a human driving that planning, what if we can imbue the radio networks with AI and machine learning? How much better can we do?
“We’re building our own custom software-defined radios, and people at other universities all around the world are building them, and they have to interact with one another and learn how to collaborate in order to use the spectrum more efficiently,” he added.
Some of the decisions about which parts of the spectrum to occupy at any given moment are determined by the machines themselves, which is where Baik comes in. He is charged with developing the algorithms that allow the machines to learn and communicate as they manage spectrum. But Baik doesn’t know which competitors they’ll be facing during periodic scrimmages, or how they’ll interact with the Dragon Radio device. This uncertainty is one of the most daunting challenges of the contest.
“Machine communication has to happen in on a one-millisecond basis. That, humans cannot do. Machine learning really kicks in in this case,” Baik explained. “Based on given data that you collected previously, you can interplay which teams are sending which signals and what kind of pattern, and also how to predict them. This gives us an opportunity to occupy that empty space of spectrum that is available. That’s the competition.”
Congress in Los Angeles
DARPA has a history of staging research contests that boggle even the most accomplished minds. There was the recent competition to develop a car that would self-drive across the desert, and one seeking a robot that could navigate through a cluttered, first-responder environment (like the Fukushima nuclear power plant in Japan) under limited remote control.
As with those earlier contests, the cost and time commitments for SC2 are prohibitive for many. That Drexel is competing at all is testimony to its intellectual grit.
“There is no one answer on how to do this. There isn’t just one radio that does it,” said Lackpour. “A big part of this is the AI component. What you can’t do is go in with one pre-set behavior and expect to do well. You have to introduce autonomy. You have to imbue this radio, the software-defined radio which has got a cognition capability, with the ability to perceive its environment and to make decisions about what is perceived, and then put that into action.
“It requires a lot of innovation,” he added. “You have your firm foundation from your textbooks and published papers, but that’s just the beginning. You have to do new research.”
Along with Drexel, academic teams from Northwestern, Vanderbilt, Purdue, and Georgia Tech, among others, are participating. The contest is open to academic, industry, and independent teams from all over the world. For example, Northrop Grumman’s Autonomous Intelligence and Robotics Laboratory is but one industry opponent.
Last December, the top six SC2 competitors split a $750K cash prize. Dragon Radio missed that pay line by just two places when it came in 8th place globally.
Dragon Radio is currently preparing for the Mobile World Congress Americas in October, to be held in Los Angeles, one of the largest annual gatherings of the wireless telecommunications community in the United States.
Always recruiting to meet the demands of competition, Dragon Radio encourages students at all levels with interests in machine learning and software-defined radio to join. Team leaders stress the rare opportunity the contest provides, and emphasize what it can do for an engineer’s skillset.
“If you want to learn, if you want to get your hands on new, cutting-edge technology and new research ideas, then this is a good project for you,” said Lackpour. “It’s really challenging. You’re going to go way beyond what you’ve done in your courses. Few people get to do something like this.
“It’s a great chance to do the next thing.”