Machine Learning Technique for Beam Management in 5G NR RAN at mmWave Frequencies

Abstract

Ericsson has an interest in investigating if the fast-growing concept known as
machine learning can be applied to beam management, in a 5G NR environment
using mmWave frequencies. Because of the high path-loss at mmWave frequencies
and high throughput demands of 5G NR systems it is crucial to the UE to always
stay connected to the most suitable beam, to provide highest possible throughput.
To obtain the required fine alignment of each single beam, optimization of beam
management operations, such as beam tracking is essential.
The type of machine learning algorithm used is called reinforcement learning.
The algorithm will aim to always connect the UE to the most suitable beam -
by comparing RSRP values from a selection of beams, that are picked based on
the current serving beam. The machine learning algorithm will initially pick a
candidate beam set based on baseline (which is explicitly programmed), however
after multiple iterations, when the algorithm is considered experienced, decisions
will instead be based on machine learning.
The algorithm will be trained from scratch over 50 different seeds i.e. 50
different environments with different properties to increase the reliability of the performance of the machine learning. The performance of the machine learning
algorithm will be evaluated by comparing the cell downlink throughput of machine
learning and baseline. When reviewing the result, it is clearly illustrated that reinforcement learning can be applied to beam management in mmWave environment to boost the average cell downlink throughput compared to baseline.

Popular Abstract

The mobile industry is looking to achieve much higher data rates when making the
transition from the fourth generation (4G) of mobile technology to 5G. One way of
accomplishing this is by using the frequency spectrum where the frequencies have
a wavelength of millimeter length, the so called millimeter wave spectrum. This
spectrum was previously not used because signals on these frequencies are difficult to deal with, but with more advanced techniques it has become a possibility. One of these advanced techniques that enables the usage of this spectrum is called beamforming and is based on focusing the transmitted signals in a more narrow direction than what was done earlier. These focused signals are called beams. The cell tower can modify the signals that it sends out so that they are focused in the direction of where the intended receiving cell phone is positioned. Due to the narrow beams, the cell phone’s movements needs to be taken care of and the direction of the transmitted beams needs to be updated. To help the cell
tower in selecting which direction to transmit in, the cell phone can report channel measurements on a set of beams and tell on which one it experiences the best signal quality. Therefore, as long as the cell tower provides a set of beams where at least one beam results in the cell phone experiencing good signal quality, the movement of the cell phone can be handled and good signal quality can be sustained. If the cell tower selects wrong beams in the set, it results in the cell phone experiencing non-optimal or bad signal quality. With this in mind, the goal of this thesis has been to select a set of beams for a moving cell phone to measure signal quality on and that results in the cell phone experiencing good signal quality. The selection of beams has been done by using a machine learning algorithm. Machine learning is a sub-field of artificial intelligence and with the help of machine learning the algorithm could experiment with different beams in the set and observe if the selected beams gave a good result or not. It could then learn from its decisions to get better over time and eventually only select beams that had been proven successful in the past. The machine learning algorithm’s performance was then compared to an already existing algorithm for selecting the beams and the results shows that the machine learning algorithm can reach significantly better performance. The two algorithms have been compared by looking at how good throughput the cell phone experiences, which is a measurement of how well the signal quality is.

• author: Bill, Joel and Fahlén, Gustav
• supervisor: Ove Edfors
• organization: Department of Electrical and Information Technology
• course: EITM01 20191
• year: 2020
• type: H2 - Master's Degree (Two Years)
• subject: Technology and Engineering
• keywords: Machine Learning, Reinforcement Learning, Beam Management, Beam Tracking, mmWaves
• report number: LU/LTH-EIT 2020-742
• language: English
• id: 9003743
• date added to LUP: 2020-02-06 13:20:14
• date last changed: 2020-02-06 13:20:14