Page 129 - ITU Journal Future and evolving technologies Volume 2 (2021), Issue 4 – AI and machine learning solutions in 5G and future networks

ITU Journal on Future and Evolving Technologies, Volume 2 (2021), Issue 4

SIMULATION OF MACHINE LEARNING‑BASED 6G SYSTEMS
IN VIRTUAL WORLDS

1
1
1
2
1
1
Ailton Oliveira , Felipe Bastos , Isabela Trindade , Walter Frazão , Arthur Nascimento , Diego Gomes , Francisco Müller 1
and Aldebaro Klautau 1
1 Universidade Federal do Pará ‑ LASSE — www.lasse.ufpa.br, Av. Perimetral S/N, Belém, Pará, Brazil., Universidade
2
Federal do Sul e Sudeste do Pará ‑ IGE — www.ige.unifesspa.edu.br , Marabá, Pará, Brazil.
NOTE: Corresponding author: Ailton Oliveira, ailton.pinto@itec.ufpa.br
Abstract – Digital representations of the real world are being used in many applications, such as augmented reality. 6G
systems will not only support use cases that rely on virtual worlds but also bene it from their rich contextual information to
improve performance and reduce communication overhead. This paper focuses on the simulation of 6G systems that rely on
a 3D representation of the environment, as captured by cameras and other sensors. We present new strategies for obtain‑
ing paired MIMO channels and multimodal data. We also discuss trade‑offs between speed and accuracy when generating
channels via ray tracing. We inally provide beam selection simulation results to assess the proposed methodology.
Keywords – 6G, arti icial intelligence, machine learning, MIMO, ray tracing

1. INTRODUCTION in which the ML/AI model is executed within the virtual
world simulation loop and the one in which the ML/AI
Machine Learning (ML) and, more generally, Arti icial In‑
telligence (AI), are currently under investigation to op‑ model is out of the loop and the simulator can then write
timize the performance of future communication net‑ iles to be later used for training ML/AI models. An exam‑
works [1]. The applications include, for instance: phys‑ ple of the irst category (INLOOP is going to be used as
the UFPA Problem Statement [6] for the 2021 ITU AI/ML
ical layer (PHY) optimizations, network management and in 5G Challenge.
self‑organization [2, 3]. Given the increasing importance
of ML/AI in communications, there are several initiatives Concerning the channel generation, the requirement of
concerning ML/AI architectures, such as the one carried having an associated digital world precludes the adoption
out by ITU [4]. This trend should continue with 6G sys‑ of a class of modern channel models that are not related
tems, which are expected to support augmented real‑ to any virtual world representation, such as the ones pre‑
ity, multisensory communications and high‑ idelity holo‑ sented in [7, 8]. We therefore adopt ray tracing (RT for
grams [5]. One such application is autonomous driving,
MIMO channel generation, which is aligned with other re‑
where digital representations are used to generate sen‑ cent work (see, e. g. [1] and references therein and allows
1
sors for hardware‑in‑the‑loop testing . And because such
the generation of site‑speci ic communication channel re‑
digital representations of the world will low through the sponses with temporal and spatial consistency.
6G network, it is expected that ML/AI can leverage them.
Therefore, a speci ic set of simulation tools for future net‑ Another motivation for this paper is to promote public
works is characterized by the requirement of being able datasets. In many ML application domains, the data is
notonlyofdealingwithcommunicationchannels, butalso abundant or has a relatively low cost. For example, the
the corresponding sensor data, matched to the scene. deep learning‑based text‑to‑speech system presented in
[9], which represents the state‑of‑the‑art, achieves qual‑
This paper focuses on strategies for simulating 6G sys‑ ity close to natural human speech after being trained with
tems that require a representation of the environment, as 24.6 hours of digitized speech. In contrast, the research
captured by cameras and, eventually, additional modal‑ and development of 5G has to deal with a relatively lim‑
ities of sensors. More speci ically, we consider Multiple ited amount of data. Considering the 5G research on
Input / Multiple Output (MIMO) systems and discuss the AI/ML applied to millimeter waves (mmWave MIMO, the
required generation of channels that are consistent with lack of abundant data from measurements or simulations
the scene at each time instant. A simulation that inte‑ hinders some data‑driven lines of investigation. With 6G
gratescommunicationnetworksandarti icialintelligence moving towards the use of even higher (Terahertz fre‑
immersed in virtual or augmented reality can be com‑ quency bands [10], it becomes even more challenging
putationally expensive, especially for time‑varying digital to perform measurement campaigns for this frequency
worlds. We discuss two categories of simulations: the one
range [11], particularly for outdoor environments. Given
1 https://www.ni.com/pt‑br/innovations/white‑papers/17/ that channel measurements for 6G will demand relatively
altran‑and‑ni‑demonstrate‑adas‑hil‑with‑sensor‑fusion.html expensive equipment, the simulation strategies for

© International Telecommunication Union, 2021 113