While AI can learn how to get devices to share spectrum, better hardware is still needed. Dr. William Chappell explained how it works at IMS 2019.
When you enter a crowded room, you see groups of people talking to each other. Generally, you can hold a conversation with someone near and you filter out the background noise. If someone in another conversation gets too loud, it might interfere with your conversation. We all share the same audio spectrum, yet we're usually able to converse.
Dr. William Chappell
at IMS 2019 in Boston.
Not so with wireless devices. They either transmit or receive with no awareness of the electromagnetic environment. Even with such knowledge, wireless devices must react to the environment and yield spectrum to others when required. A plenary address by DARPA's Dr. William Chappell on June 4 at the International Microwave Symposium in Boston described how artificial intelligence (AI) can help, but better hardware is also needed.
"The era of free spectrum is over," declared Chappell. "There's no slice of spectrum that's underutilized anymore." Given that premise, Chappell explained how negotiating a spectrum full of wireless devices can benefit from artificial intelligence (AI).
Machines have been learning for years, with breakthroughs coming in 1997 when IBM's Deep Blue computer beat the reigning world chess champion Garry Kasparov. Chess, with its 8x8 board and simple rules, is a relatively straightforward game. In 2016, a machine beat the world champion AlphaGo player. That required considerably more computing power. The machine was able to learn the game rather than being told the rules as was the case with chess.
What makes chess and AlphaGo easy for machines to learn comes from the fact that the entire board is visible. All facts are known at any given time. Games like DOTA 2 are considerably more difficult because some aspects of the game are hidden. Chappell explained that a computer, Open AI Five was able to learn by playing against 180 years' worth of games against itself every day. Even that, however pales in comparison to learning how wireless devices can share spectrum because devices enter and exit that shared spectrum randomly and some may need higher priority than others. For example, military or emergency communications take priority over posting cat photos.
"The real world is messy," said Chappell. That's why DARPA developed the RF Colosseum, a test bed for making sense out of the RF chaos, which operates 24 hours a day, every day as part of DARPA's Spectrum Collaboration Challenge (SC2, Figure 1). SC2 consists of five wireless networks where teams compete for prizes to show how to best manage spectrum sharing with minimal interference. Computers learn from interference caused by hundreds of software-defines radios (SDRs). Unlike a game, which ends when there's a winner, the RF environment has no stopping point.
"The strategy must continually adjust to new conditions," said Chappell. It constantly gets smarter." Chappell showed, using video, how frequency hopping doesn't scale to a scenario where hundreds of devices are trying to share spectrum. That's where AI comes in. It learns the RF environment and adjusts for new conditions. So far, spectrum sharing has improved throughput from 13 Mbps to 19 Mbps in some scenarios with a goal of improving to 52 Mbps. A demonstration planned for Mobile World Congress Americas 2019, to take place in Los Angeles, should show further improvement towards to goal.
Realizing that RF engineers are more interested in hardware than in another AI story, Chappell asked "What does this have to do with RF components? The mind and body of the RF physical layer must work together to improve spectrum sharing. To find out what's needed, Chappell spoke with several RF engineers.
"We need better analog radios," said Chappell, who cited a need for improvements from 14 bits to 19 bits. That's needed to switch from analog to digital filters. With better tunable, linear filters, the additional bits might not be needed. Figure 2 highlights what's needed, in the form of components and matierials.
"We have a long way to go," Chappell concluded.