6G will build a new framework for cellular communications, sensing and machine
learning that requires a fundamental redesign of 5G along several key dimensions.
6G will need to deliver ultra-reliable wide area coverage, furnishing a tradeoff
between extremely high data rates and low latency guarantees.
6G will also introduce communication/sensing co-design principles to achieve
centimeter-level localization, including inside buildings and in both urban and
rural areas, to allow the network to learn precisely where every device is.
Machine learning techniques will become fundamental to operations from the
physical layer through the network stack, including configuration and deployment,
aided by the new network sensing capabilities.
The key metrics will include power consumption and reliable coverage, as opposed
to ever-increasing peak data rate.
We envision an increasingly open and software-defined cellular network that builds
on the O-RAN paradigm to provide a platform for continuous and more rapid innovation,
as opposed to 5G and earlier.
In the 6G@UT research center at UT Austin,
our four closely related key 6G research directions are
- Deeply embedded machine learning techniques across the protocol stack
as well as spatial and temporal scales for site-specific adaptability and network automation.
- Pervasive sensing will feed machine learning algorithms continuously
tuning and reconfiguring the network, while “sensing-as-a-service” will be offered to subscribers and applications.
- New spectrum above 100 GHz and new topologies for improved coverage,
including massive low-earth orbit satellite constellations and self-backhauled small-cell deployments.
- Network slicing and sharing architectures will enable new revenue
streams and sharing of network/spectrum resources for diverse tenets with different requirements.
We expect innovation to lie at the intersection of these areas, e.g. applying machine learning in conjunction with localization information to enable efficient resource sharing.
Within the 6G@UT research center, Prof. Evans' group
is working on the following projects:
- Thrust 1: Deeply embedded machine learning
- Reinforcement Learning for Multicell Beamforming and Reconfigurable Intelligent Surfaces.
(current with NVIDIA).
In many cellular deployments, communication performance is limited by interference.
For multicell multiantenna basestations, we have developed algorithms that design
minimum user power control settings (uplink) and optimal digital beamformers (downlink)
to meet a certain SINR target.
In each basestation, the algorithm estimates out-of-cell interference and models
quantization noise from the data converters to solve the convex optimization problem.
We propose to extend our approach for hybrid analog/digital beamforming architectures
for millimeter wave bands.
To solve this non-convex problem, we propose to use RL with received SINR and spatial
location for each user as the state; actions will estimate real-valued (continuous)
angles of arrivals/departures and large beamforming/combining matrices; and the reward
will be a sum of the user SINR values in dB.
We also propose to incorporate reconfigurable intelligent surfaces (RIS) and control
the RIS via RL to maximize SINR.
A RIS, which is a thin metasurface composed of discrete elements, passively manipulates
incident electromagnetic waves through controlled reflective phase tuning.
The phase shift leads to a unit-modular constraint on each RIS element, which leads to
a non-convex problem to maximize communication performance.
- Thrust 3: New Spectrum and Topologies
- Full-Duplex Millimeter Wave Communications and Integrated Access and Backhaul for 6G.
Full-duplex systems have gained enormous attention, due to their potential to double
the spectral efficiency, reduce the latency and enhance the reliability/coverage since
the transmission and reception occur at the same resource block in time and frequency.
Due to these advantages, massive MIMO full-duplex systems are currently proposed in
3GPP Release 17.
Due to the in-band transmission and reception, full-duplex systems are vulnerable to
the loop-back self-interference which is up to 1000-10000x times the received signal
power resulting in a severe degradation such as the ADC saturation.
The goal of this project is to design robust hybrid analog/digital beamforming algorithms
to improve the spectral and energy efficiency while minimizing the losses incurred by
the self-interference.
We propose to develop these beamforming algorithms for cellular user equipment/base
station communications as well as integrated access and backhaul.
- Rate optimization with Reconfigurable Intelligent Surfaces.
Reflectors in the environment have been sources of interference, and reconfigurable
intelligent surfaces (RIS) have the potential to flip the narrative.
RIS, which are thin metasurfaces composed of discrete elements, passively manipulate
incident electromagnetic waves through controlled reflective phase tuning.
Our initial work develops gradient-based approaches to configure RIS phases to optimize
a single user’s rate assuming global availability of channel state information.
Given the lack of a global optimum due to the unit modular constraint on each RIS element,
we develop and compare heuristics to try to maximize spectral efficiency or the equivalent
channel power as a proxy.
Future work will relax the assumptions to optimize sum rates in more practical and
challenging RIS-assisted settings.
Mail comments about this page to
bevans@ece.utexas.edu.