Deep Learning for Wireless Communication Systems
Abstract
Deep learning has demonstrated huge success in a variety of domains. This success can be attributed to two factors: (a) First is addressing the challenge in modeling complexity. For example, in image classification, the underlying statistics of data and the inference problem are hard to analytically model. For such cases, deep learning has been able to provide good solutions. (b) The second is in discovering complex algorithms. For example, in the games of chess and Go, where the underlying statistics of the problem are analytically well defined, but the space of an algorithm is huge and grows exponentially over time, deep learning has been able to discover algorithms that significantly outperform the stateoftheart ones. In designing communication algorithm, we face both challenges. The underlying communication channels cannot be analytically modeled in several scenarios, for example, when the user is highly mobile. The space of communication algorithms scales exponentially in the block length; for example, a channel coding function of k bits requires a mapping from 2k messages to 2k codewords. The current wireless communication systems often rely on heuristicsbased algorithms to mitigate the analytical challenges in deriving physical layer algorithms. Motivated by the successes of deep learning in solving challenging tasks of image classification, Go, and protein folding, our central hypothesis is that deep learning serves as a powerful tool to design novel communication algorithms and could result in new paradigms for wireless networks in the 2030s and beyond. This hypothesis was formulated, in large part, on the basis of our preliminary work indicating that nonincremental gains are achievable for interactive and multiterminal communications [1, 2, 3, 4, 5, 6, 7, 8, 9]. Specifically, we demonstrate that neural networkbased codes outperform the stateoftheart codes by several orders of magnitude for channels with feedback. An extended version of our preliminary work is currently deployed in Bluetooth communications, achieving more than 2? communication range, 4? better power consumption, and 1000? reliability [10]. Similarly, for multiuser interference channels, we demonstrate that the classical interference alignment scheme can be enhanced significantly by introducing the nonlinearity and adapting the scheme to practical channels. In this project, we aim to realize the promise of the deep learning results into a Proof of Concept functioning realworld prototype capable of coping with realworld channel conditions, including channel impairments and device imperfections. We particularly focus on tactical multiuser communications, channels with feedback, and multiuser communications with feedback. We will deliver a neural code library which includes neural networkbased codes trained to be robust and adaptive to practical channel models, static and mobile, with impairments and jamming noise. We expect 5100? improvement in reliability depending on the SignaltoNoise ratio and channel statistics, with the latency comparable to the latency of traditional schemes. We will also demonstrate our invention in an overtheair implementation by modeling several realworld challenges (eg., channel impairments, frequency and time synchronization hazards, I/Q imbalance) with softwaredefined radios. We have a lab for overtheair experiments with an extensive track record of demonstrating neural networkbased communication algorithms in overtheair environments and practical communication systems [7, 9, 10, 11].
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 09, 2023
- Source ID
- W911NF2310062
Entities
People
- Hyeji Kim
Organizations
- Army Contracting Command
- United States Army
- University of Texas at Austin