A team of researchers says it has developed the next generation of space communications, using a reinforcement learning neural network (RLNN), a form of machine learning. The researchers use an algorithm to enable cognitive radio functions so that a communications system can autonomously adapt itself. With an RLNN, a radio can optimize its operating parameters or select the best wireless channel in its vicinity in order to avoid interference created by space weather, solar storms, and other noise that interferes with signals.
The architecture has been successfully evaluated on the Space Communications and Navigation (SCaN) Testbed aboard the International Space Station. NASA uses the vehicle to research radio communications and GPS. The test bed includes three software-defined radios that can be reprogrammed for communications research experiments.
“Multiobjective Reinforcement Learning for Cognitive Satellite Communications Using Deep Neural Network Ensembles,” the article outlining the approach, was published in the May edition of the IEEE Journal on Selected Areas in Communications.
The team includes researchers from NASA’s Glenn Research Center, Penn State University, and Worcester Polytechnic Institute (WPI). Five are IEEE members: Sven Bilén, Paulo Victor Rodrigues Ferreira, Timothy Hackett, Randy Clinton Paffenroth, and Alexander Wyglinski.
TRAINED TO ADAPT
Radio spectrum, which is finite, must accommodate all forms of wireless communications including cellphone calls and data traffic—which are increasing at an unprecedented rate. Software-defined radios can be used to implement cognitive radios in order to allow spectrum to be used more efficiently. For example, a cognitive radio transceiver can scan for unoccupied bands and change its transmission and reception parameters to different frequencies during heavy data loads without interruption. It also can listen for interference on busy channels and calculate a way to reduce it so more people can use the channels.
Congestion of the radio frequency spectrum in space is similar to that on Earth. Those using the spectrum are restricted to certain segments of the airwaves, but there are so-called white spaces that anyone can use as long as there’s no licensed activity on it. Those hoping to use those white spaces must compete not only with one another but also with naturally occurring phenomena, like space weather and electromagnetic radiation emitted by the sun and other celestial bodies.
Current space communication systems use radio-resource selection algorithms, but they’re rudimentary and work with a preprogrammed look-up table for adaptive communications, according to the article. Operators have to manually switch ground stations and spacecraft over to another part of the spectrum whenever the frequency is interrupted. That process doesn’t work well in changing environments.
With the aid of machine learning, the researchers say, cognitive radio will be able to alter its transmission frequency or cancel out distortions.
“This is the first time anyone has ever conducted successful cognitive radio experiments using machine learning algorithms in space,” says Wyglinski, an IEEE senior member. He’s a professor of electrical and computer engineering at WPI, as well as president of the IEEE Vehicular Technology Society.
“We’ve hit the right nail on the head in terms of combining computing platforms and advances in algorithms for machine learning with a very challenging environment like space,” he says.
By using reinforced learning, he adds, an artificial neural network can be trained to adapt to the conditions of space through multiple trials and experiments. That’s because the algorithm is set up to learn in a manner similar to the human brain, by weighing inputs and optimizing those weights to achieve a goal.
Reinforcement learning would allow the cognitive radio to realize that it is encountering an area of poor transmission and either move to a new channel, use another type of transmission scheme, or increase the power of its signal, Wyglinski says.
“When humans have problems with radio reception, we tweak knobs and turn dials to get better performance. We use our brains to figure that out,” he says. “That's what the cognitive radio does. It’s the brain. RLNN helps with that decision-making, that adaptation learning, if you will. This is quite significant.”
The proof-of-concept design was tested through computational simulations as well as ground- and space-based experiments. Simulation results for the new cognitive radio engine design achieved 80 percent average accuracy with respect to the overall performance relative to an ideal system in the six experiments conducted on the ISS, Wyglinski says.
The ground system is composed of software-defined radios, commercial modems, and RF equipment emulating the targeted space-to-ground channel. The on-orbit communication system—including a space-based, remotely controlled transmitter—resides on the ISS and operates with a receiver at NASA’s Glenn Research Center, in Cleveland. The goal of the flight testing was to provide a baseline performance and potential of computational engines for communications systems and to inspire future research, according to the article.
The algorithm also has commercial applications, Wyglinski says. It could be integrated into existing terrestrial communication systems to enhance their performance in challenging environments.
The first page of the article is available for free from the IEEE Xplore Digital Library.
This article has been corrected from an earlier version.