UVA Engineering Researcher Brings Machine Learning to 5G

When Nikolaos Sidiropoulos and his colleagues submitted a paper on the application of data-driven machine learning to wireless communications in 2017, it was still a novel idea. Wireless communications had formerly been almost exclusively the domain of traditional engineering design methods. Much has changed since then. Sidiropoulos' pioneering paper was named to a list of the Institute of Electrical and Electronics Engineers Communications Society's “Best Readings in Machine Learning in Communications,” and he was invited to guest-edit a special issue of the society's Journal on Selected Areas in Communications on the topic. “The response to our call for papers has been unprecedented,” Sidiropoulos, professor and chair of the Charles L. Brown Department of Electrical and Computer Engineering and an IEEE fellow, said. “Normally, we receive 40 submissions for an issue like this. We received 130. There is now consensus that data-driven methods have an important role to play in wireless communications.” The recognition by IEEE of Sidiropoulos' work is significant. IEEE is the world's largest technical professional society and serves more than 420,000 professionals involved in all aspects of the electrical, electronic and computing fields and related areas of science and technology. Its membership includes computer scientists, software developers, information technology professionals, physicists, medical doctors, and many others in addition to IEEE's electrical and electronics engineering core. Gearing Up for 5G Among the many challenges facing engineers designing wireless communications systems are optimizing the resource allocation among multiple users and minimizing the mutual interference generated by their signals. Multiuser power control is a typical example. The power of transmitters operating in a single area over the same frequency band — including cellphones as well as antennas on towers — must be coordinated to ensure the best possible overall system performance. “If everyone transmits at maximum power and uses the same bandwidth at the same time, you have polyphony,” Sidiropoulos said. “In this situation, no one is heard.” There are a number of well-established algorithms that can accomplish tasks like multiuser power control, but they are computationally intensive, making them too slow for real-time applications. Wireless communication systems are constantly in a state of flux. To account for changing channel conditions and fluctuating user numbers, adjustments must be made in milliseconds. With the advent of 5G networks, the need for a solution has become even more pressing — and more daunting. “We are dealing with much higher bandwidths and carrier frequencies, and more nonideal engineering devices and components,” Sidiropoulos said. “The sheer scale and complexity of the problem is rising rapidly. We are looking for novel solutions that scale up.” Moving the Burden of Computation Offline The method that Sidiropoulos and his colleagues proposed was to use machine learning to learn the relationship between network conditions and the adjustments that the algorithms recommend. They used multiuser interference as an example. “Our approach was to treat the algorithm as a map and, using historical data, deploy machine learning to determine how it maps network conditions to power control outputs,” he said The results of that offline learning would be encoded in a neural network that could be dropped into a wireless communications system. In their paper, Sidiropoulos and the team focused on determining whether their approach was practical — which meant determining the complexity of the neural network that would be needed to successfully approximate the results from the algorithm. They found that a small network would be sufficient to obtain high approximating accuracy and that substituting a neural network for a state-of-the-art interference management algorithm could achieve orders of magnitude improvements in computational time. “The ultimate goal is to marry the empirical success of machine learning with a century of engineering insights on wireless communication,” Sidiropoulos said. “Our paper was a first step in this direction, which, as the response to our call for papers demonstrates, has now become a major focus of research.”