Page 131 - ITU Journal Future and evolving technologies Volume 2 (2021), Issue 4 – AI and machine learning solutions in 5G and future networks
P. 131
ITU Journal on Future and Evolving Technologies, Volume 2 (2021), Issue 4
As proposed in [16], the CAVIAR framework concerns also relies on the scene description and can extract fea‑
a speci ic category of 6G simulations that rely on vir‑ tures from the raw sensor data to feed its AI/ML algo‑
tual worlds and incorporate two subsystems: wireless rithms.
communications and AI/ML. In the next paragraphs, we
brie ly review the CAVIAR framework, depicted in Fig. 1, Fig. 1 illustrates the INLOOP CAVIAR framework with the
and then focus on the important aspect of generating the AI/ML module within the simulation loop. When the deci‑
communication channel corresponding to a given scene sions of this module do not affect the environment, it can
of the virtual world. We discuss how the Raymobtime be convenient to split the simulation into two stages, with
methodology [12] its well to the demand for communi‑ the irst one being an OUTLOOP CAVIAR simulation that
cation channels imposed by 6G CAVIAR simulations. writes episode iles that will be later used for designing
and assessing AI/ML models. The more evolved INLOOP
A CAVIAR simulation generates multimodal data for each simulation is required in cases such as a drone mission in
discrete time ∈ ℤ, and is able to operate in two which the AI/ML decisions will change the drone trajec‑
modes, the irst mode is focused on online learning, run‑ tory and, consequently, its wireless channel. In general,
ning the simulation and the neural network simulta‑ when the AI/ML model issues commands or actuator sig‑
neously, creating an environment where data is trans‑ nals that effectively change the trajectories of mobile en‑
mitted in real time, or in discrete samples with time tities, alter the environment or the communication sys‑
stamps de ined by the user. The second mode of op‑ tem state (e.g., buffer occupation), the simulations may
eration performs data recording in databases or text need to be INLOOP and communication channels gener‑
iles, working as a tool for creating datasets. Along ated ly. In the simpler OUTLOOP simulation cat‑
the simulation, the machine learning for communications egory, channels can be pre‑computed and the communi‑
(ML4COMM) engine operates on data organized as an cation simulation decoupled from the physical engine, as
episode = [( , ), … , ( , )], with a sequence of often used in AI/ML applied to beam selection [19, 12].
1
1
tuples ( , ), = 1, … , , of paired data, where The next sections provide two examples to distinguish IN‑
and are sets with the input AI/ML parameters and cor‑ LOOP and OUTLOOP CAVIAR simulations.
responding outputs, respectively. In supervised learning,
consists of desired labels for classi ication or regres‑ 2.1 OUTLOOP CAVIAR simulation for beam
sion, while for reinforcement learning consists of re‑ selection
wardsfortheagents. Thetuples( , )denoteevolution
over discrete‑time . In our methodology, the outputs of Beam selection is a classical application of AI/ML to com‑
the simulators are periodically stored as “snapshots” (or munications [20, 21, 22]. The goal is to choose the best
scenes) over time sam , where sam is the sampling period pair of beams for analog beamforming, with both trans‑
and ∈ ℤ.
mitter (Tx) and receiver (Rx) having antenna arrays with
The main steps in Fig. 1 can be summarized as follows. only one Radio Frequency (RF) chain and ixed beam
The environment is composed of a 3D scenery with ixed codebooks. Fig. 2 illustrates beamforming from a Base
and mobile objects. These objects are created and placed Station (BS) to both vehicles and drones.
with specialized tools and data from the Internet, as de‑
scribed in [12] and [17]. The positions and interactions
among mobile objects are determined by a physics engine
(for instance, the Unreal engine or the Simulation of Ur‑
ban MObility (SUMO) traf ic generator [18]).
Once the scene is complete, the environment is repre‑
sented via sensors, such as LIDAR, which is simulated
by Blensor and Blender software, returning point cloud
data (PCD) that maps the shapes of the 3D space around
the sensor. It is possible to adjust the resolution of the
PCD through a quantization process. A ray‑tracing soft‑
ware (Remcom’s Wireless InSite in Fig. 1) also captures
the communication channel for the given scene. The sen‑
sors output constitute the episode input , and the cor‑
responding output is obtained by a signal process‑
ing module. These episodes are actually what is stored
in Raymobtime episodes [12] but in a CAVIAR simula‑
tion they can be created and used on‑the‑ ly, if needed. Fig. 2 – Beamforming from BS to both vehicles and drones.
The CAVIAR 6G virtual world simulator also incorporates
a communication system that has some functionalities
driven by the ML4COMM engine. The ML4COMM engine
© International Telecommunication Union, 2021 115