Research Projects
Stop the VideoResearch Projects
STATUS: In Progress
YEAR: 2023
TOPIC AREA: Connected and autonomous systems Vehicles and infrastructure
CENTER: PSR
Deep-learning-based radio channel prediction for vehicle-to-vehicle communications
Project Summary
Project number: PSR-22-10 TO 065Funding source: Caltrans
Contract number: 65A0674
Funding amount: $100,000
Performance period: 6/1/2023 to 5/31/2024
Project description
Vehicle-to-vehicle (V2V) communications are an essential component of the future of
transportation. It is essential for any type of assisted or autonomous driving, e.g., for a
vehicle warning those driving behind it that an emergency braking is imminent, to cars
talking to each other to arrange smooth lane changes. Yet the general adoption of V2V has
been slow – a fact that is partly due to economic reasons but is also impacted by limited and
unpredictable performance of such systems. This, in turn, is related to the difficult operating
environment – both in terms of signal propagation between transmitter (TX) and receiver
(RX), but also the high density of devices, leading to strong interference, and thus possible
packet loss. Further research to improve reliability and latency is thus urgently needed. One
of the main challenges for V2V communications systems is that the allocation of resources
for the communication (spectrum, power) needs to be done based on the current state of
the propagation channel, but the vehicles have only access to past measurements. It is thus
critical to find suitable channel prediction methods that allow the vehicles to infer the
current state from those past observations.
Channel prediction has a long history in cellular communications, and has also been
applied to some V2V scenarios. Most of the past work was based on simplified channel
models, and classical tracking/extrapolation methods have been proposed in the literature,
but performance in real-world environments has been insufficient. This motivates the use of
Machine Learning (ML) for this purpose, since ML has been shown to be a good tool for
tackling otherwise intractable problems with complicated data structure. Yet ML for V2V
channels suffers from a lack of available channel measurement data, and a disconnect
between the neural network structures chosen for the prediction and the main propagation
physics.
This project will investigate machine learning (ML) based vehicular channel
extrapolation on actual measurements, instead of simplified channel models as done in the
past. The measurements are for multi-antenna (8x8 MIMO) channels, enabling the
exploitation of the directional information of the channels and beamforming for improving
SNR and reducing interference. At the same time, development of such V2V MIMO channel
predictions with ML has – to the best of our knowledge – rarely been investigated in the past.
To close these gaps, we will develop new ML algorithms not only for channel extrapolation,
but also transmission parameter adaptation and scheduling. In particular, we will consider
the topics of neural network design, training strategy, data augmentation, and whether a
separate channel prediction and scheduling is superior or inferior to a joint approach. With
these novel ML designs and carefully designed time schedules, the effectiveness and
robustness of the proposed solutions will be accessed by the actual V2V measurements
collected in multiple scenarios, e.g., car-to-car, car-to-truck, truck-truck in campus, city street,
and highway.
A key outcome of this project will be a methodology for optimizing resource assignment
in V2V communications. The first part of the methodology, namely the channel prediction, is
independent of the specific type of system; this is important because currently there is a
debate about whether the IEEE 802.11p/WAVE system will be used, or the 5G NR; our
proposed work will be useful in either case. The actual resource allocation algorithms
developed in the second part will show differences in absolute performance depending on
which system they are used with, but in either case will provide a significant improvement
over the state of the art, where suboptimum resource allocation is a major source for
inefficiencies and outages in V2V communications. These outcomes will eventually lead to
the following broader impact: 1) Improved traffic safety, i.e., packet drops and insufficient
coverage, may lead to vehicles not learning safety-critical information, or warning of a
pile-up, in time to react. It is worth remembering that on-board sensors of a car are not
sufficient to avoid all accidents (this is a key motivation for V2V communications in the first
place). 2) Energy savings, since more reliable communications allows cars to drive more
closely to each other in a convoy formation, which in turn reduces energy consumption due
to reduced wind drag. At the same time, this also leads to reduction of traffic jams, because
the capacity (throughput) of the road is improved.
P.I. NAME & ADDRESS
Andreas MolischProfessor, Ming Hsieh Department of Electrical Engineering; USC Viterbi School of Engineering
3740 McClintock Avenue
Hughes Aircraft Electrical Engineering Building (EEB) 500Los Angeles, CA 90089-2565
United States
[email protected]
Major Funders
Metrans Associate Partners
