Page 84 - 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

To carry out the ITU AI for 5G challenge, an open data The remainder of this paper is structured as follows: irst,
set with WLAN measurements obtained from a Section 2 presents the CB problem for next‑generation
network simulator was made available. As later WLANs and describes the data set provided for through‑
discussed, network simulators are gaining importance to put prediction in dense deployments. Then, Section 3
enable future ML‑aware communications by acting as ML gives an overview of the ML‑based solutions proposed by
sandboxes. In the context of the challenge, synthetic data the ITU challenge participants, for which the results are
was used for training ML models. provided in Section 4. Finally, Section 5 concludes the pa‑
per with inal remarks.
In summary, based on the proposed problem statement
and the solutions provided by participants, we discuss
the feasibility of predicting the throughput of complex 2. CHANNEL BONDING IN NEXT‑
WLAN deployments through ML. To the best of our
GENERATION WLANS
knowledge, this is an under‑researched subject with
high potential. Accurate performance predictions may In this section, we irst describe the CB problem in WLANs
open the door to novel real‑time self‑adaptive and underscore the need for ML models for performance
mechanisms able to enhance the performance of wireless prediction. Then, we introduce the data set provided for
networks by leveraging spectrum resources the ITU challenge, which is open to any researcher inter‑
dynamically. For instance, given a change in the network ested in this topic.
con iguration, predictions about future performance
values can be used as a heuristic to guide the choices 2.1 Channel bonding in IEEE 802.11 WLANs
made by algorithms that operate in decentralized
self‑con iguring environments [3]. Next‑generation IEEE 802.11 WLANs are called to face
the challenge of providing high performance under com‑
Table 1 brie ly summarizes all the models proposed by
plex situations, e.g., to provide high throughput in mas‑
the participants of the challenge and the main motivation
sively crowded deployments where multiple devices co‑
behind them. For instance, the model proposed by the
exist within the same area. To ill the strict require‑
ATARI team aims to exploit the graph structure inherent in
ments derived from novel use cases, features such as
WLAN deployments. Alternatively, the solution provided by
Multiple‑Input Multiple‑Output (MIMO), Spatial Reuse
Ramón Vallés is focused on abstracting different categories
(SR), or multi‑Access Point (AP) coordination are be‑
of features (e.g., signal quality, bandwidth usage) and
ing developed and incorporated into the newest amend‑
generate predictions based on them.
ments, namely IEEE 802.11ax (11ax) and IEEE 802.11be
(11be) [5, 6].
Table 1 – Summary of the ML models proposed by the participants of
the challenge. One of the features that are receiving more attention is
Team Proposed Model Motivation Ref. Channel Bonding (CB) [7, 8], whereby multiple frequency
Exploit graph channels can be bonded with the aim of increasing the
Graph Neural
ATARI representation [29] bandwidth of a given transmission, thus potentially im‑
Network
of WLANs proving the throughput. Since its introduction to the
Abstract problem
Ramon Feed‑forward characteristics by [31] 802.11n amendment, where up to two basic channels of
Vallés Neural Network 20 MHz could be bonded to form a single one of 40 MHz,
categories
High performance, the speci ication on CB has evolved and currently allows
STC Gradient Boosting lexibility, and ease [35] for channel widths of 160 MHz (11ac/11ax). Moreover,
of deployment CB is expected to support up to 320 MHz channels in the
UC3M Feed‑forward Learn throughput 11be amendment.
NETCOM Neural Network function exhaustively [36]
Address problem’s
NET Random Forest
INTELS Regression non‑linearity and [37]
reduce dimensionality
The results presented in this paper showcase the feasi‑
bility of applying ML to predict the performance of next‑
generation WLANs. In particular, some of the proposed
models have been shown to achieve high prediction accu‑
racy in a set of test scenarios. Moreover, we have identi‑
ied the main potential and pitfalls of the proposed mod‑
els, thus opening the door to new contributions that im‑
prove the baseline results shown in this paper. The data
set is available online [4], and we expect it can be used for Fig. 1 – U‑NII‑1 and U‑NII‑2 sub‑bands of the 5 GHz Wi‑Fi channels.
benchmarking other ML methods in the future.

68 © International Telecommunication Union, 2021