Page 132 - 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

We irst assume beam selection for a vehicular to infras‑
tructure network, to illustrate an OUTLOOP CAVIAR sim‑
ulation. In this case the communication subsystem is a
downstream MIMO system in which a BS with a Uniform
Linear Array (ULA) of antennas communicates with

cars with ULAs of antennas. ML is used for beam‑

selection.
Discrete Fourier Transform (DFT) codebooks =

̄
̄
{w ̄ , ⋯ , w ̄ } and = {f , ⋯ , f } are used at the trans‑

1
1

mitter and the receiver sides, respectively. The beam pair
[ , ] is converted into a unique index ∈ {1, 2, ⋯ , },
where ≤ . For the ‑th pair, the equivalent chan‑ Fig. 3 – Scene from an INLOOP CAVIAR simulation in which a drone is
served by a BS and RL is used for beam selection and for determining

nel (without considering noise) can be calculated as the drone trajectory.
∗
= w Hf , (1) The scenario depicted in Fig 3 allows us to investigate

several problems that relate communication with drones
̂
and the optimal beam pair index is given by path planning. One important issue is how to obtain the
channels on‑the‑ ly. If the visualization is performed after
̂
= arg max | |. (2) the whole simulation is inished, the time to generate the

∈{1,⋯, } channel (via RT, for instance) may be longer. But in this
The beam selection is then posed as a top‑ classi ication case the scenes need to be visualized along the simula‑
problem. At time , the classi ier inputs are features ob‑ tions (as part of a game, for example), then the minimum
tained from and the output is the beam pair . number of frames per second will impose a limit on the

time to generate the communication channels.
For the scenario presented in this section, the trajec‑
tory of vehicles and all mobile objects do not depend The next section discusses our Raymobtime methodology
on the AI/ML model, hence all the episodes can be pre‑ and the corresponding datasets. Other publicly available
computed. Next, we discuss a simulation in which the RT‑based datasets are listed in Table 1. The ViWi dataset,
trajectories are determined by the AI/ML model and the presented in [13], provides similar output data compared
channels cannot be pre‑computed. to Raymobtime, including visual data. The DeepMIMO
dataset [25] is maintained by the same group as ViWi and
offers only wireless channel information. The dataset de‑
2.2 INLOOP CAVIAR simulation with drones scribed in [1] does not have visual information as well.
and reinforcement learning One of the main differences between these three datasets
and Raymobtime is how mobility is handled. The Ray‑
Unmanned Aerial Vehicles (UAVs) are being used in many mobtime methodology simulates realistic traf ic with sev‑
connected applications, such as surveillance and prod‑ eral moving vehicles using the SUMO software in order
uct delivery. UAVs can also be used as mobile radio base to provide better spatial and temporal consistency, as
stations to extend reach or improve network capacity, well as channel variability due to the moving scatterers.
mainly in situations of disasters and accidents. In order ViWi [13] (in its irst version), DeepMIMO and the map‑
to meet the requirements of all these use cases, the net‑ based channel model in [1] use a ixed grid for Tx‑Rx po‑
work links need to obey particular requirements, ranging
sitions and therefore does not consider varying speeds
fromverylowlatencytohighdatarates[23]. All thismoti‑ for moving transceivers. ViWi version 2 provides one
vates intense research on 5G technologies for supporting new scenario that includes several moving vehicles, each
UAV‑based applications. However, there are currently few equipped with an omnidirectional antenna.
simulation tools for testing and studying telecommunica‑
tion systems that involve UAV solutions and their corre‑
sponding channels. The CAVIAR framework is deeply in‑
tegrated with the Unreal Engine development kit and the 3. IMPROVEMENTS ON RAYMOBTIME
Airsim simulator [24], which bring realism to the physical METHODOLOGY
aspects of the simulations.
The Raymobtime methodology proposed in [12] aims at
As part of the UFPA Problem Statement for the 2021 ITU providing a multimodal dataset, including RT channel in‑
AI/ML in 5G Challenge, we designed an INLOOP CAVIAR formation and data from sensors, such as images, LIDAR
simulation in which RL is executed at the BS and used in and location, as illustrated in Fig. 1. One major challenge
two problems: a) determine the drone trajectory and b) in building the Raymobtime datasets is to provide accu‑
beam selection along the downstream. In the challenge, rate wireless communication channel parameter through
the drones need to deliver pizzas to distinct addresses in the use of RT simulation software. In this work, Remcom’s
a neighborhood. Fig 3 illustrates the scenario. Wireless InSite (WI) RT software [26] was adopted given

116 © International Telecommunication Union, 2021