Page 86 - 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

In view of the complexity of selecting the best con igura‑ The data set generated with Komondor has been used for
tion of channels, the proposed problem statement had the training and validating ML models in the context of the
goal of shedding light on the potential role of ML in DCB. ITU AI for 5G Challenge. The assets provided comprise
In particular, it served to gather participants’ proposals both training and test data sets corresponding to mul‑
of ML models able to predict the performance of differ‑ tiple random WLAN deployments at which different CB
ent CB igurations. This information can be used by con igurations are applied. As for training, two separate
a decision‑making agent to choose the best con iguration enterprise‑like scenarios, namely, training1 and train‑
of channels before initiating a transmission. Through‑ ing2, have been characterized. In each case, a different
put prediction in WLANs has been widely adopted for ixed number of BSSs coexist in the same area, according
performance analysis through mathematical models, in‑ to users’ density.
cluding the well‑known Bianchi model [16], Continuous‑
Time Markov Networks (CTMNs) [17], or stochastic ge‑ In training1, there are 12 APs, each one with 10 to 20 as‑
sociated STAs. Regarding training2, it contains 8 APs with
ometry [18]. However, these well‑known models lack ap‑
5 to 10 STAs associated with each one. For both train‑
plicability for online decision‑making because they fail to
ing scenarios, three different map sizes have been consid‑
capture important phenomena either on the PHY or the
ered (a, b, and c), where STAs are placed randomly. Simi‑
MAC, or they entail a high computational cost. Thus, mod‑
larly to training scenarios, the test data set includes a set
eling high‑density complex deployments through these
models may be highly inaccurate or simply intractable. of random deployments depicting multiple CB con igura‑
tions and network densities. In this case, four different
In this regard, we envision ML models to assist the CB scenarios have been considered according to the number
decision‑making procedure in real time. The fact is that of APs (4, 6, 8, and 10 APs). Note, as well, that, for each
ML can exploit complex characteristics from data, thus type of scenario, 100 and 50 random deployments have
allowing to solve problems that are hard to solve by been generated for training and testing, respectively. In
hand‑programming (see, for instance, its success in im‑ all the cases, downlink UDP traf ic was generated in a full‑
age recognition). Moreover, ML models can be trained of‑ buffer manner (i.e., each transmitter always has packets
line and then improve their accuracy with measurements to be delivered). Table 2 summarizes the entire data set
acquired online. To the best of our knowledge, this is an in terms of the simulated deployments.
under‑researched subject. While ML has been applied for
predicting aspects related to Wi‑Fi networks, such as traf‑
Table 2 – Summary of the simulated deployments used for generating
ic and location prediction [19, 20], it has been barely ap‑ both training and test data sets.
plied for explicitly predicting their performance. In this
Scenario id Map width # APs # STAs
context, the work in [21] provided an ML‑based frame‑
training1a 80 x 60 m
work for Wi‑Fi operation, which includes the application
training1b 70 x 50 m 12 10‑20
of Deep Learning (DL) for waveforms ication, so
training1c 60 x 40 m
that WLAN devices can identify the medium as idle, busy, Training
training2a 60 x 40 m
or jamming. Closer in spirit to our work, [22] proposed
training2b 50 x 30 m 8 5‑10
an ML‑based framework for WLANs’ performance predic‑ training2c 40 x 20 m
tion. test1 4
test2 6
Test 80 x 60 m 2‑10
2.3 Introduction to the data set test3 8
test4 10
To motivate the usage of ML for predicting WLANs’ per‑
2
formance, we provide an open data set obtained with the With respect to input features, these are included in the
3
Komondor simulator. Komondor is an open‑source IEEE ilesusedforsimulatingeachrandomdeployment. Inpar‑
ticular, the most relevant information to be used for train‑
802.11ax‑oriented simulator, whose fundamental opera‑
ing ML models is:
tion has been validated against ns‑3 in [23]. Komondor
was conceived to cost‑effectively simulate complex next‑ 1. Type of node: indicates whether the node is an AP
generation deployments implementing features such as or an STA.
channel bonding or spatial reuse [24]. Furthermore, it in‑
2. BSS id: identi ier of the BSS to which the node be‑
cludes ML agents, which allows simulating the behavior longs.
of online learning mechanisms to optimize the operation
of WLANs during the simulation. 3. Node location: {x,y,z} coordinates indicating the po‑
sition of the node in the map.
2 The data set has been made publicly available at https://zenodo.org/ 4. Primary channel: channel at which carrier sensing
record/4059189, for the sake of openness. is performed.
3 https://github.com/wn‑upf/Komondor, Commit: d330ed9.

70 © International Telecommunication Union, 2021