NSF-AoF: SOLID: System-wide Operation via Learning In-device Dissimilarities

Cellular communication systems continue to incorporate new multiple-antenna technologies. In particular, third, fourth and fifth generation cellular systems saw advancements in the use of multiple antennas at the base-station infrastructure and multiple antennas in the devices. A main application of these antennas was to support multiple-input multiple-output (MIMO) communication, which is known to increase spectral efficiency and thus the data rates that can be achieved by devices in a given bandwidth. The numbers of antennas and the ways the antennas are used can vary across device models even from the same manufacturer. At the same time, the types of devices supported in cellular systems is growing beyond smartphones to include other highly mobile platforms like aerial vehicles, automobiles, and robots. The differences in the hardware between devices, coupled with the high device mobility, makes it challenging to configure the antennas to provide MIMO communication with the highest performance. This project develops machine learning-inspired solutions to empower devices to learn optimal configurations collaboratively. System-wide Operation via Learning In-device Dissimilarities is a cooperation among experts in wireless communications at North Carolina State University (NC State) and Tampere University (TAU). The overall objective of the proposal is to employ machine-learning-assisted collaborative solutions for MIMO beam prediction and codebook optimization in a large-scale dynamic system. The key challenge of such networks is the extreme diversity of the devices’ hardware (e.g., antenna designs and configurations). The existing distributed ML approaches do not explicitly include this type of client heterogeneity and do not fully support the temporal and spatial heterogeneity of data, network resources, and deployments. The project team will develop a novel integrated-learning and wireless-networking framework, which will enable the design and optimization of advanced MIMO beam-management solutions specifically tailored to the highly diverse and dynamic system. This project will result in new algorithms for collaborative device-centric beam management for 5G+/pre-6G MIMO communications in non-stationary environments with highly mobile and heterogeneous agents. The specific technical contributions occur in several directions: (a) Distributed user-centric learning for optimizing codebook-based MIMO communications; (b) Novel representation of device heterogeneity in an ML-friendly way; and (c) Network-resource optimization to facilitate distributed learning. The immediate impact will be improved communication efficiency in 5G+/pre-6G networks. The longer-term impact will be the establishment of the core principles for designing fast and reliable methods of distributed ML training deployed over wireless systems with diverse hardware and resources.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.