AITF: Learning and Adapting Sparse Recovery Algorithms for RF Spectrum Sensing

点击翻译按钮获取中文摘要

Wireless communications technology plays a critical role in society, supporting personal communication, business, defense and security, family connectivity, and entertainment. As new applications emerge, the demand for connectivity is increasing at a rapid pace. This development has put a strain on the available resources for communications: the spectrum is a finite natural resource; the most useful portions for mobile applications lie roughly from 700 MHz to 6 GHz. Large portions of it have been allocated to primary users, many of whom play socially critical roles. To accommodate the growing demand for wireless connectivity, there is a need for devices that can sense and utilize available spectrum in an opportunistic manner, while not interfering with primary users. The challenge is to do this in a time- and energy-efficient manner, on mobile devices. A general approach, which promises order-of-magnitude improvements in energy efficiency for rapidly detecting large interfering signals, uses a combination of new hardware to take a small number of measurements of the spectrum as a whole, and nontrivial algorithms to interpret those measurements. This project develops from machine learning to learn algorithms that are adapted to the specific characteristics of the hardware sensor, improving handling of non-linearities, yielding lower power sensors with improved sensitivity. The researchers are mentoring graduate and undergraduate students, whose work crosses disciplinary boundaries, and disseminating the results through new course development and a new textbook. The project studies methodologies for learning and adapting algorithms for sparse recovery for RF spectrum sensing, leveraging a known connection to artificial neural networks, in which the structure of the algorithm dictates the topology and weights of the network. These weights can then be adapted and optimized to fit the characteristics of a physical sensor. A major promise of this approach is the ability to adapt to modeling errors, while simultaneously producing recovery methods that are more sensitive, more robust, and implementable in a simple and efficient manner. The project is developing a principled and transparent methodology, including theoretical characterizations of when and why it is possible to learn reconstruction and support recovery procedures that are effective in both typical and worst-case senses. The project studies these problems both for linear inverse problems and for nonlinear problems, both for sensing the spectrum at a single time, and for integrating information over time. The project experimentally evaluates the impact of these methodologies on the efficiency and sensitivity of hardware sensors, realized as integrated circuits.