Skip to content Skip to navigation

Research & Ideas

Search this site

Can AI prevent mobile devices from dropping calls?

Today, a weak signal is no big deal, but next-generation wireless devices will need flawless service to perform vital tasks such as telesurgery, or helping guide autonomous cars.

Illustration of three signal waves in blue, pink, and yellow

Artificial intelligence could be used to tune 5G wireless signals and prevent them from taking dangerous dips. | Illustration by Kevin Craft

It happens all too often. A mobile call drops right in the middle of your sen…

Most of the time, it’s merely an annoyance. But one day wireless signals will be tied to critical applications in which a blip could be disastrous. Imagine what might happen in a telesurgery, for example, when a surgeon is controlling a robotic device via a wireless connection. Or, on a highway dedicated to autonomous vehicles, if one car suddenly stops or slow downs and needs to wirelessly signal its intent to all the other cars in the vicinity to avoid a pileup.

Scenarios such as these will play out over time as wireless carriers deploy 5G networks designed to handle more than the calls, texts and streaming data flowing over current 3G and 4G systems. When 5G begins rolling out next year, and becomes universal by the mid-2020s, it will serve wireless devices and applications that demand uninterrupted connectivity. Maintaining that connectivity will require that the radios in mobile devices be far more flexible than anything we know today.

A team of engineers at Stanford has devised a way to insure that flexibility. They have designed wireless radios that can automatically sense their circumstances and use artificial intelligence to tune their transmitters and receivers to optimize performance and adapt themselves to the needs of any scenario. The researchers call their flexible creations “self-driving radios.”

Not unlike tabletop radios of old, these next-generation radios are controlled by internal settings that govern the inputs and outputs that affect signal reception. Researchers refer to these settings as “knobs.” They aren’t physical knobs, but rather digital adjustments that control important variables such as what frequency a radio is communicating on, how much power it needs and what bitrate is used for data transfers. The myriad combination of knobs and the sheer number of settings for each present an almost limitless array of options to the radio designers.

“In essence, it’s too much information, and we’ve turned to artificial intelligence to help,” says Sachin Katti, a professor of computer science at Stanford and leader of the team.

Today, if connectivity is fading, it could be from any number of causes ranging from channel congestion to poor coverage. In such cases, predetermined algorithms, built into the radios, tweak the knobs to help the radio find a better channel or adjust the transmit power and data rate. In 3G and 4G radios, the control algorithms are preprogrammed and hard-coded into the radio chips, a labor-intensive process that limits the number of scenarios that can be optimized on any given radio.

Samuel Joseph, a doctoral candidate in Katti’s lab and co-lead author of a paper the team is presenting at this week’s HotMobile conference in Santa Cruz, California, said this preprogramming approach works fine with less-demanding current wireless technologies, when interruptions are mostly just annoyances.

5G radios will have to be far more nimble. They will be asked to deliver data at 100 megabits-per-second to as many as a million devices per square kilometer on vehicles traveling up to 300 miles per hour. And, they will have to do it at energy efficiencies 100 times that of 4G, while dropping fewer than one connection in a million.

Katti’s team has responded with artificial intelligence algorithms that will enable self-driving radios to find the right knob settings for each and every eventuality. Their approach is based on reinforcement learning, a branch of artificial intelligence in which computers teach themselves to perform specified tasks. Co-lead author Rakesh Misra, who earned his PhD working in Katti’s lab, said the self-training system can cycle through innumerable what-if scenarios and write code on the fly to deal with situations much more quickly than human programmers.

What’s more, the Stanford AI system is able to adapt even after the self-driving radio is deployed.

Should a radio find itself running a vital application in unfamiliar or dangerous circumstances, the team’s algorithm senses the situation and writes new code to adjust the receiver’s knobs in real time, without missing a beat.

In performance testing for routine operation, the Stanford team’s AI-based algorithms perform at least as well as algorithms hand-designed by experts. When faced with more challenging situations, their algorithms actually outperform the hand-designed alternatives.

The Stanford researchers are presenting their self-driving radio prototype at the HotMobile conference, where radio chip manufacturers, equipment vendors and service providers are demonstrating the components of future 5G networks.

“The flexibility to adapt will be key to 5G’s future,” Katti says.