Page 92 - 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

5. DISCUSSION 4. Finally, we remark the importance of cost‑effectively
predicting the performance in WLANs, which may
5.1 Contributions open the door to novel mechanisms using these pre‑
dictions as heuristics for online optimization. The in‑
With contributions from participants around the globe, corporation of these kinds of models to WLANs is ex‑
the ITU AI for 5G Challenge has unprecedentedly estab‑ pected to be enabled by ML‑aware architectural so‑
lished a platform for addressing important problems in lutions [41].
communications through ML. As for the performance pre‑
diction problem in CB WLANs (referred to as PS‑013), the
challenge has allowed us to glimpse the potential of ML 5.3 The role of network simulators in ML‑
models for addressing it. aware communications

This paper provides a compendium of ML models pro‑
The availability of data for training is key for the success of
posed for throughput prediction in WLANs, including
ML application to future 5G/6G networks. Given the cur‑
popular models such as neural networks, linear regres‑
rent limitations in acquiring data from real networks, sim‑
sion, or random forests. In particular, we have provided
ulators emerge as a practical solution to generate comple‑
an overview of the proposed models and analyzed their
mentary synthetic data for training ML models. The fact
performance in the context of the challenge. By opening
is that data may be scarce because, among other reasons,
the data set used during the competition, we encourage
measurement campaigns are costly, data from networks
the development of mechanisms that improve the base‑
involves privacy concerns, or data tenants are not willing
line performance shown by the ML models presented in
to share their data.
this work.
Through the problem statement discussed in this paper,
5.2 Lessons learned we have contributed to showcase how synthetic data can
be used to train ML models for networks. A notorious ad‑
From the performance evaluation done in this paper, we vantage is that network simulators allow characterizing
have drawn the following conclusions: complex deployments, sometimes representing unknown
1. First, even if the data set was not particularly big, 6 situations, so they help to train and validate ML models.
some of the proposed ML models achieved good re‑ Beyond generating data for training, network simulators
sults. This is quite a positive result since it opens the are envisioned to serve as secure platforms for testing,
door to ML models that can be (re)trained fast, thus training, and evaluating ML models before being applied
becoming suitable for (near)real‑time solutions.
to operative networks [42]. In consequence, we fore‑
2. Second, most of the proposed DL‑based models have see the adoption of simulators into future ML‑aware net‑
works as a key milestone for enhancing both reliability
shown higher accuracy for the denser and more com‑
plex deployments (which more closely match the and trustworthiness in ML mechanisms.
training scenarios) than for the sparser ones. While
capturing complex situations is quite a positive re‑ ACKNOWLEDGEMENT
sult, the pitfalls observed in simpler deployments
also suggest that out‑of‑the‑box DL methods may fail Representative members of each participating team has
at capturing the relationship between interference been invited to co‑author this paper. Nevertheless, the
same credit goes to the rest of the participants of PS‑013
and performance of WLANs. In this regard, well‑
in the ITU AI for 5G Challenge: Miguel Camelo, Natalia
known models characterizing WLANs (e.g., SINR‑
based models [40]) can potentially be incorporated Gaviria, Mohammad Abid, Ayman M. Aloshan, Faisal Alo‑
mar, Khaled M. Sahari, Megha G Kulkarni, and Vishalsagar
into the ML operation for the sake of improving ac‑
Udupi. We would also like to thank enormously every‑
curacy, thus leading to hybrid model‑based and data‑
driven mechanisms. one that made possible the ITU AI for 5G Challenge, with
special mention to Vishnu Ram OV, Reinhard Scholl, and
3. Third, and related to the previous point, GNNs have Thomas Basikolo.
been shown to be particularly useful to capture the This work has been partially supported by grants WIND‑
complex interactions among devices in WLANs, both MAL PGC2018‑099959‑B‑I00 (MCIU/AEI/FEDER,UE),
in terms of interference and neighboring activity. In and 2017‑SGR‑11888.
particular, we have realized the importance of pre‑
processing the data set in order to obtain accurate
prediction results. Deriving information speci ic to REFERENCES
the problem (i.e., signal quality, interference) has [1] ITU‑T, Recommendation Y.3172 (2019). ”Architec‑
turned out to be essential for the sake of generalization. tural framework for machine learning in future net‑
works including IMT‑2020”.
6 Youtube‑8M Data set (http://research.google.com/youtube8m/) con‑
sists of 350.000 hours of video, while only six hundred different ran‑
dom deployments have been used in this paper for training purposes.

76 © International Telecommunication Union, 2021