CAREER: Development of Learning Frameworks for Nonlinear Massive MIMO Systems

San Diego State University Foundation, San Diego CA

Investigators

Abstract

The need for enabling massive connectivity using energy- and cost-efficient massive multiple-input multiple-output (MIMO) wireless transceivers, mobile devices, sensors, and actuators has motivated engineers to use low-cost and non-customized hardware. However, these hardware components are highly susceptible to generating nonlinear distortions. Moreover, each component distorts the signals of interest in its own way with a possibly unknown nonlinear transfer function that needs to be compensated by signal processing techniques. Studies of massive MIMO have demonstrated its tolerance to hardware impairments. In addition, community research efforts have also focused on new performance frontiers in signal processing for nonlinear massive MIMO. This CAREER project is framed into these efforts by developing novel learning frameworks for signal processing in nonlinear massive MIMO systems in which structured optimization of their performance is not feasible or too complex to implement. The project aims to construct a holistic approach to estimate, detect, optimize, and adapt in nonlinear MIMO systems with the development of model-based and model-free machine learning, deep learning, deep reinforcement learning, and meta-learning frameworks. The outcomes of the project can transform the area of MIMO signal processing and propel massive MIMO into the next stage of development. Moreover, the project will integrate research and education activities through the development of undergraduate and graduate level courses and an integrated wireless testbed for research and training. The project also aims to broaden the participation of students in research and educational activities in electrical engineering with an emphasis on communications and signal processing. This project will develop new learning frameworks to resolve the aggregation of nonlinearities in wireless transceivers and to optimize massive MIMO performance in highly dynamic and ill-defined operational environments with possibly unknown distortions. The project is organized into four interconnected thrusts: i) Learning to estimate and detect with machine learning and deep learning by incorporating the domain knowledge of nonlinear MIMO for devising efficient model-based and data-driven learning models, ii) Learning to optimize signaling designs with deep learning and deep reinforcement learning, focusing on optimizing pilot and signaling designs for model-based and model-free nonlinear MIMO systems, iii) Learning to adapt with offline and online meta-learning, focusing on developing learning models that will enable fast adaptation to a newly observed nonlinear MIMO system configuration with only few training samples, and iv) Integrating research into coursework development for teaching wireless communications and signal processing to undergraduate/graduate students with hands-on experiments using built-in software-defined radio and off-the-shelf wireless testbeds. Through the development of learning to estimate, detect, optimize, and adapt frameworks, this project will promote a better understanding of signal processing in nonlinear MIMO systems and will have broad applications in other learning-based wireless systems. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.