“You’ve graduated from the school of spectral hard knocks,” Paul Tilghman, a U.S. Defense Advanced Research Projects Agency (DARPA) program manager, told the teams competing in the agency’s Spectrum Collaboration Challenge (SC2) finale on 23 October. The three-year competition had just concluded, and the top three teams were being called on stage as a song that sounded vaguely like “Pomp and Circumstance” played overhead.
“Hard knocks” wasn’t an exaggeration—the 10 teams that made it to the finale, as well as others who were eliminated in earlier rounds of the competition—had been tasked with proving something that hadn’t been demonstrated before. Their challenge was to see if AI-managed radio systems could work together to share wireless spectrum more effectively than static, pre-allocated bands. They had spent years battling it out in match-ups in a specially-built RF emulator DARPA built for the competition, Colosseum.
By the end of the finale, the top teams had demonstrated their systems could transmit more data over less spectrum than existing standards like LTE, and shown an impressive ability to reuse spectrum over multiple radios. In some finale match-ups, the radio systems of five teams were transmitting over 200 or 300 percent more data than is currently possible with today’s rigid spectrum band allocations. And that’s important, given that we’re facing a looming wireless spectrum crunch.
But as Tilghman has also stressed repeatedly during SC2, when a DARPA grand challenge ends, it doesn’t mean the technology is ready to go to market. These challenges are more about proving that a new technological idea is possible. And in that regard, SC2 has by all standards succeeded.
So, great—AI-controlled radios can work together to share spectrum amongst themselves and pack more data into a given amount of spectrum. But where does the technology go from here, now that the DARPA challenge is finished?
In two panels at MWC Los Angeles, the mobile industry convention where the finale occurred, industry experts, employees of the Department of Defense (DoD) and Federal Communications Commission (FCC), and members of the top three teams all weighed in, with several clear themes emerging.
One big talking point was that there’s still room for more collaboration. “Now that the competition is over,” Tilghman said, “I know many of the teams are looking forward to working together with their competitors.” It was a sentiment shared by the team members during their panel. “We’ve got really smart guys here, and we haven’t been able to ask ‘Why is your radio doing this?’” said John Shea, a member of the University of Florida team, GatorWings, that came in first place.
The competitive nature of the challenge also meant teams didn’t necessarily build their systems to be as truly collaborative and altruistic as they may need to be in the real world. At the end of the day, the top teams recognized that they also had to score the most points to win, and so developed AI systems that would also seek to opportunistically edge out the other radio systems while also making sure they remained collaborative enough to meet SC2’s demands. GatorWings’ system, for example, could recognize when it was collaborating with the best other teams and enter a “Competition Mode” that caused it to selfishly hoard spectrum from other well-performing teams. Similarly, third place team Zylinium’s system had a “Hulk Mode” that achieved the same effect.
But there are also larger technical questions that must be resolved beyond training the AI systems to be more collaborative. Members of the GatorWings team talked about how they spent two years building and understanding the custom software-defined radios for their system—and these team members each had 20 years of experience in RF engineering. The winning teams agreed that there’s still plenty of work to be done in refining the radios, including improving systems’ “brawn,” that is, how well their signals could resist interference and still get their own data through.
Teams also believe there’s value in first training their AIs in less-intensive simulations before throwing them into a fully-emulated environment like Colosseum. Simulations will train AIs more quickly, cutting down on the time needed to train them to be more accurate and robust.
Panelists from the DoD and FCC discussed the work to be done to make players comfortable with the idea of sharing spectrum. “We’re being asked to share with technologies we simply don’t typically share with,” said Fred Moorefield, the acting deputy chief information officer for the DoD’s Command, Control, Communications and Computers, and Information Infrastructure Capabilities. He raised the example of a broadband provider that must collaborate with military satellite communications. Julius Knapp, chief of the FCC’s Office of Engineering and Technology, said it could be tricky to get competing service providers to share the same spectrum in the real world, as collaborative radio efforts require sharing information about your wireless communications that companies might not be keen to disclose.
For at least one opportunity, the finale of SC2 seems well-timed: the deployment of 5G networks. Several bands of spectrum are being developed for this next generation of wireless, including for the first time, millimeter wave spectrum. But 5G also relies on immensely valuable mid-band spectrum, which remains coveted and in short supply. Developing collaborative spectrum technologies alongside 5G could make the wireless generation more successful.
It’s impossible to predict when or in what manner we’ll see autonomous spectrum sharing in the wireless networks of the future. But research will definitely continue. And in the spirit of what’s next, it’s worth mentioning that Colosseum itself, the testbed at the heart of SC2, will be starting a new life.
With MWC Los Angeles wrapping up, the testbed will be packed up and trucked to Northeastern University in Boston, where it will become part of the National Science Foundation’s Platforms for Advanced Wireless Research program.
A version of this post appears in the December 2019 print magazine as “AI-Enabled Spectrum Technology: What’s Next?”