On the Use of Artificial Neural Networks to Predict the Quality of Wi-Fi Links

Abstract

One of the aspects that mainly characterize wireless networks is their apparent unpredictability. Although several attempts were made in the past years to define for them deterministic medium access techniques, for instance by having data exchanges scheduled by an access point, as a matter of fact they remain a partial solution and are unable to ensure the same behavior as wired infrastructures, since interference may also come from devices outside the network, which obey different rules. A possible way to cope with disturbance on air, both internal and external to the network, is to obtain some knowledge about it by analyzing what happened in the recent past. This information, usually expressed in terms of suitable metrics, is then employed to optimize network operation, for example by prioritizing time-sensitive traffic when needed. In the simplest approaches such metrics coincide with statistical indices evaluated on transmission outcomes, like the failure rate. In this paper we analyze a more sophisticated solution that relies on machine learning, and in particular on artificial neural networks, to predict the behavior of a Wi-Fi link in terms of its frame delivery ratio. Results confirm that more accurate predictions than simpler methods (e.g., moving average) are possible, even when training is partially independent from the specific conditions experienced on the different channels.

Full text

Paper submitted to IEEE ACCESS.
Digital Object Identifier
On the Use of Artificial Neural Networks
to Predict the Quality of Wi-Fi Links
ALBERTO SALVATORE COLLETTO2, STEFANO SCANZIO1, (Senior Member, IEEE),
GABRIELE FORMIS1,2, (Student Member, IEEE), GIANLUCA CENA1, (Senior Member, IEEE)
1National Research Council of Italy (CNR-IEIIT), Corso Duca degli Abruzzi 24, I-10129 Torino, Italy
2Dipartimento di Automatica e Informatica, Politecnico di Torino, Corso Duca degli Abruzzi 24, I-10129 Torino, Italy
Corresponding author: Stefano Scanzio (e-mail: stefano.scanzio@cnr.it)
ABSTRACT One of the aspects that mainly characterize wireless networks is their apparent unpre-
dictability. Although several attempts were made in the past years to define for them deterministic medium
access techniques, for instance by having data exchanges scheduled by an access point, as a matter of fact
they remain a partial solution and are unable to ensure the same behavior as wired infrastructures, since
interference may also come from devices outside the network, which obey different rules.
A possible way to cope with disturbance on air, both internal and external to the network, is to obtain some
knowledge about it by analyzing what happened in the recent past. This information, usually expressed
in terms of suitable metrics, is then employed to optimize network operation, for example by prioritizing
time-sensitive traffic when needed. In the simplest approaches such metrics coincide with statistical indices
evaluated on transmission outcomes, like the failure rate.
In this paper we analyze a more sophisticated solution that relies on machine learning, and in particular on
artificial neural networks, to predict the behavior of a Wi-Fi link in terms of its frame delivery ratio. Results
confirm that more accurate predictions than simpler methods (e.g., moving average) are possible, even when
training is partially independent from the specific conditions experienced on the different channels.
INDEX TERMS Channel quality prediction, Wireless networks, Dependable wireless communication,
IEEE 802.11, Artificial neural networks, Machine learning
I. INTRODUCTION
WIRELESS networks are progressively replacing ca-
bles in a variety of contexts. In application fields like,
e.g., home and car entertainment [1], [2], home automation
[3], building automation [4], and sensing [5], their adoption
to support wire-free communication (and hence, interaction)
among humans and machines, in any combination (H2H,
M2M, and H2M), is a popular ongoing trend and one of the
main enablers of the so-called Internet of Things (IoT) [6].
Besides throughput, also the requirements about latency
and dependability may be sometimes quite demanding. This
is particularly true in contexts like modern industrial environ-
ments (both factory and process automation), which are being
re-shaped by recent paradigms like Industry 4.0 [7], Industry
5.0 [8], and the Industrial Internet of Things (IIoT) [9]. In
particular, upper bounds on response times are customarily
defined in soft/firm real-time systems, which in wireless net-
works are often specified in probabilistic terms. For example,
a given fraction of time-sensitive messages (e.g., 99.99%) is
required to not exceed the intended deadlines, otherwise the
controlled system is not ensured to operate correctly.
To tackle such a wide range of requirements, wireless
networking is currently characterized by a noticeable het-
erogeneity of protocols [10]. Among the most relevant tech-
nologies, which suit different scenarios, we find: 5G/6G [11]
and IEEE 802.11 (Wi-Fi) [12] that apply to contexts where
high throughput is demanded on geographic and local areas,
respectively, LoRaWAN for long-range low-rate data ex-
changes [13], IEEE 802.15.4 (e.g., ZigBee, WirelessHART,
WIA-PA, 6TiSCH, ISA 100.11) for ultra low-power mesh
networks [14], and Bluetooth Low Energy (BLE) to connect
both personal and industrial devices [15].
In the past years, a fair amount of research activities were
spent for improving the quality of communication on wire-
less networks by exploiting techniques like seamless redun-
dancy [16], transmission scheduling [17], software-defined
networking [18], automatic network configuration [19], and
so on. Some solutions rely on enhanced medium access con-
VOLUME 6, 2016 1
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3327523
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/

Alberto S. Colletto, Stefano Scanzio, et al.: Paper submitted to IEEE ACCESS
trol (MAC) mechanisms that enable deterministic behavior.
Besides legacy solutions, like the Point Coordination Func-
tion (PCF) and the Hybrid Coordination Function (HCF)
Controlled Channel Access (HCCA) in IEEE 802.11 (which
knew limited usage), the use of trigger frames in Wi-Fi 6
and multi-link operation (MLO) in Wi-Fi 7 [20] achieve
tangible improvements. The same holds for Time Slotted
Channel Hopping (TSCH) [21], [22] and the Deterministic
and Synchronous Multi-channel Extension (DSME), which
exploit mixed time-frequency diversity in IEEE 802.15.4.
It is important to point out that, although effective, de-
terministic MACs are unable to ensure for wireless net-
works the same dependability as the wired ones. In fact,
besides intra-network interference, transmission on air may
also suffer from extra-network disturbance, caused by nearby
wireless networks (possibly based on different transmission
technologies) and electromagnetic noise (generated by indus-
trial/power equipment). In the latter cases, there are simply
no means to prevent interference.
A different approach to cope with above phenomena,
without necessarily changing the MAC mechanism, consists
in applying machine learning (ML) to wireless communica-
tions. In the recent years, a number of research activities were
started targeted at exploiting intelligence to improve commu-
nication quality. Some proposals rely on suitable algorithms
to predict the future quality of a link in terms of metrics like,
e.g., the frame delivery ratio (FDR). The ability to foresee
variations of the communication quality can be then exploited
by end nodes and intermediate equipment to proactively react
to events that affect the wireless spectrum (either worsening
or improving it), in an attempt to provide some probabilistic
guarantees for, e.g., the typical key performance indica-
tors (KPIs) relevant to industrial communication networks,
namely, end-to-end latency, dependability, determinism, and
even power consumption [23].
For example, if the prediction model foresees that channel
conditions are likely to deteriorate in a while, applications
could react in advance by reducing the amount of generated
best-effort traffic (e.g., by lowering the sampling rate at the
perception layer for non-critical information, like environ-
mental monitoring), so as to privilege higher-priority traf-
fic with soft/firm real-time requirements. As an alternative,
communication could be switched to a different channel
before the current one worsens too much. This solution may
bring tangible advantages also for hand-over procedures of
mobile nodes: ML predictions can be exploited to determine
the instant when a roaming Client Station (STA) needs to
reassociate to an different Access Point (AP) when it moves
away from the AP it is currently associated. In this case,
the quality of every available channel must be predicted,
and this requires that all of them are probed, e.g., using
reinforcement learning (RL) techniques based on exploration
and exploitation. This resembles the operation of the Minstrel
algorithm, customarily implemented in commercial Wi-Fi
equipment, and rate adaptation algorithms [24].
Similar approaches can be adopted when redundancy is
exploited to increase reliability and decrease communication
latency, by allowing frames to be sent on different channels
(MLO). In this case, ML can be exploited to drive channel
selection. They also apply to seamless redundancy, when
the same frame is sent on multiple channels at the same
time. In particular, deferral techniques [25] could use ML to
determine the primary channel on which the first copy of the
frame is sent. Transmission of the second copy is deferred
for a while to achieve the best trade-off between reliability
and resource consumption. The last, more complex example,
among many possible ones, concerns the combined adoption
of seamless redundancy and ML to support node mobility, for
preserving communication quality during hand-over [26].
In this work, artificial neural networks (ANNs) are ex-
ploited to predict the FDR in the near future starting from
the outcomes of transmission attempts logged by a wireless
node in the recent past. Our analysis is based on extensive
experimental campaigns that involved real Wi-Fi devices.
Results show that the quality of a wireless channel in the
near future (a few minutes) can be predicted with satisfactory
accuracy.
The structure of the paper is the following: Section II
summarizes the state-of-the-art about the use of artificial in-
telligence (AI) to improve the behavior of wireless networks;
Section III presents a simple model for a wireless link, while
prediction models are introduced in Section IV; Section V
describes the experimental testbed we used to acquire the
databases that characterize the behavior of real wireless links
and the software implementations related to this work; Sec-
tion VI outlines the results obtained by applying the different
prediction models to real data; finally, Section VIII draws
relevant conclusions and outlines future work.
II. LITERATURE REVIEW
The two surveys [27], [28] identify many challenges intrinsic
to the use of ML and AI techniques to improve key per-
formance indicators related to wireless networks. In [29],
[30] ML techniques are used for the selection of the best
AP, whereas in [31], [32] they are exploited to drive the
handover decision between APs. Other works are based on
traffic prediction [33], [34]. This information can be used
to indirectly infer the quality of the wireless channel: for
example, in [35] it is used to select channel allocation.
Although this is an interesting research direction, there is
not a linear dependency between traffic and FDR. Finally,
reinforcement learning was used to select the best channel
for transmission in [36], where the reduction of packet losses
was assessed through simulation.
Regarding the subset of works in the scientific literature
related explicitly to Wi-Fi and the prediction of the future
behavior of the channel, the use of ANNs was envisaged in
[37] but only on artificial data. In [38], ML techniques were
used for predicting the signal strength. Instead, in [39], [40]
models based on ANNs are described aimed at predicting the
channel gain in specific application contexts. All of the above
works refer to the prediction of aspects related to the physical
2VOLUME 6, 2016
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3327523
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/

Alberto S. Colletto, Stefano Scanzio, et al.: Paper submitted to IEEE ACCESS
layer. Generally speaking, they cannot be directly and easily
employed to assess the KPIs that customarily characterize
communication at the application level, such as the FDR.
In [41], [42], ANNs were applied in a preliminary form
to data acquired on a real Wi-Fi testbed, using the FDR as a
metric. Compared to the research activities presented in [41],
[42], the analysis in this work was done: 1) with a noticeably
more complex software, suitable for efficiently handling big
data; 2) with much larger databases in terms of their size,
especially for what concern testing (93.5 days for training
and 53.8 days for test instead of 38 days for training and
2.5 days for test), hence providing more reliable results; 3)
analyzing four Wi-Fi channels (1, 5, 9, and 13) instead of just
channel 13, hence providing better significance; 4) including
potentially interesting architectures like multi-output ANNs,
which will be presented in Subsection VI-D and are aimed to
lower training time and memory occupation in the end-user
device; and, 5) considering the effects on FDR prediction
accuracy achieved by extending the training database with
data acquired on different channels (to this purpose, relevant
experiments were carried out where training relied on chan-
nels other than the one used for testing).
III. CHANNEL AND PROTOCOL MODEL
All the experimental campaigns reported in this work specif-
ically concern Wi-Fi. Nevertheless, this kind of analysis
can be easily extended to any wireless communication tech-
nologies that support confirmed transmission services. The
simple network model we consider includes two nodes com-
municating over a wireless link. In particular, the sender
node repeatedly sends data frames to a recipient node at
a fixed rate. When a frame is correctly received, the latter
replies by returning an acknowledgment (ACK) frame to
the sender to notify that the transmission succeeded. How
this mechanism was actually implemented on commercial
Wi-Fi devices (e.g., by disabling retransmission, fixing the
transmission speed, etc.) is described in detail in Section V.
Data frame transmission is performed cyclically with pe-
riod Ts= 0.5 s, and the reception of every ACK frame
is logged. Possible outcomes are success (xi= 1) if the
ACK frame associated with the i-th data frame is correctly
received by the sender node; otherwise failure (xi= 0) in
those cases when either the data frame or the ACK frame
are corrupted and the transmission timeout expires. This
behavior is intentional, and mimics the point of view of the
application executing in the sender node, which only indi-
rectly has a way to know about transmission errors. Generally
speaking, this provides an ordered sequence of outcomes
D= (x1, ..., xi, ..., x|D|)as output, we call database, whose
size (in terms of the number of elements) is denoted |D|. The
database is depicted schematically in Fig. 1, along with some
quantities whose definition is provided below.
The goal of this research work is to find a prediction
function f(·)that, at any time, estimates the value of the FDR
evaluated over a given future horizon (one to ten minutes)
given the previous Nptransmission outcomes as input. Let
x1x2x3x4x5x6x7x8x9x10 x11 x12 …
=1
 =+ + +
4



= 5 = 4
 = 7
x||

=  = (,,,,, )
FIGURE 1. Example of past and future sliding windows, on which the
prediction function f(·)is applied and the target tkis computed, respectively.
xkbe the outcome of the most recent transmission attempt.
Then, the input of the prediction function is the sequence
Ik= (xk−Np+1, ..., xk−1, xk), and the predicted FDR can
be expressed as
˜
tk=f(Ik,w) = f(xk−Np+1, ..., xk−1, xk,w),(1)
where w(a vector, in the most general case) describes the
model parameters determined in the training phase (it was
explicitly added to the prediction function to stress the fact
that it directly affects the related output ˜
tk).
During the training phase a specific database Dch
tr is em-
ployed, where ch represents the channel on which it was ac-
quired (e.g., D5
tr represents the training database obtained on
Wi-Fi channel 5). The training phase consists in estimating
the model parameters w∗that minimize the average error
between the predicted FDR value ˜
tkand the target tk, which
characterizes the real failure rate in the future. The mean
squared error is typically used for this purpose, which is the
same as considering the sum J(w)of squared errors:
w∗= arg min
w
J(w),(2)
where
J(w) =
|Dtr|−Nf
X
k=Np
[tk−f(Ik,w)]2.(3)
The target coincides with the FDR observed in a future
interval whose width is Tf=Nf·Ts. At any time, it
is computed as the simple moving average (SMA) of the
sequence (xk+1, xk+2 , ..., xk+Nf), which includes Nffuture
outcomes that belong to database D,
tk=1
Nf
k+Nf
X
j=k+1
xj.(4)
The value tkgiven by (4) provides a statistical estimate of the
probability that a frame transmission attempt on the testbed
succeeds, which is what the prediction function should in
theory seek. It unavoidably suffers from a certain error, as
it is computed using a limited number Nfof outcomes. In the
case of quasi-stationary spectrum conditions, transmission
attempts can be approximately modeled as independent and
identically distributed (iid) random variables, and the mean
squared error affecting the estimate tkis equal to σ2
x/Nf,
where σ2
xrepresents the variance of samples (xk)in the
VOLUME 6, 2016 3
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3327523
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/

Alberto S. Colletto, Stefano Scanzio, et al.: Paper submitted to IEEE ACCESS
database. This makes the problem considered here more
complex than usual time series analysis, where the true target
to be predicted is directly available.
To evaluate the accuracy of the prediction function
f(Ik,w)a test database Dch
te is employed. The same pro-
cedure described above is applied to Dch
te to obtain the target
values tk. Then, the prediction function is applied to every
input interval Ikfor evaluating its ability to estimate the
target. Two kinds of prediction errors were defined in this
work: the absolute error |ek|=|tk−˜
tk|, where | · | is the
modulus operator, and the squared error e2
k= (tk−˜
tk)2.
Starting from them, the mean absolute error (MAE) µ|e|
and the mean squared error (MSE) µe2can be computed as
1
|Dte|−Nf−Np+1 P|Dte |−Nf
k=Npek, where ekstands for |ek|and
e2
k, respectively. Besides averages, other statistical indices
can be also evaluated for prediction errors, like percentiles.
For example, |e|p90 represents the 90-percentile of the abso-
lute error.
IV. PREDICTION MODELS
The prediction models we took into account for estimating
the FDR on a wireless link in the immediate future are the
simple moving average (SMA), which is used as a reference
for the other proposed techniques, and artificial neural net-
works (ANNs).
A. SIMPLE MOVING AVERAGE (SMA)
The main assumption behind SMA prediction is that, the
FDR in the immediate future remains (almost) the same
as in the immediate past. The prediction function mostly
resembles (4), but applies to the past samples
fSMA(Ik, Np) = 1
Np
k
X
j=k−Np+1
xj,(5)
where Npis the number of outcomes exploited for prediction.
The Npvalue is explicitly included in the prediction function
to remark that model accuracy heavily depends on it.
Accuracy of SMA prediction can be optimized by exploit-
ing the training database Dtr. In particular, the optimal value
N∗
pcan be found as
N∗
p= arg min
Np
|Dtr|−Nf
X
k=Np
tk−fSMA(Ik, Np)2.(6)
Practically, Npis varied over a range of values wide enough
to find the minimum of the error function. The N∗
pvalue
determined in the training phase is then used to parameterize
the prediction function (5) when the test database Dte is used.
B. ARTIFICIAL NEURAL NETWORK (ANN)
The ANN model of this work is the multi layer perceptron
(MLP). A second model was also briefly investigated, namely
long short-term memory (LSTM), which is customarily em-
ployed for time series forecasting. Results about LSTM were
not reported because, for the specific task of predicting
x20 x21 x22 x23 x24
x8x9x10 x11 x12 x13 x14 x15 x16 x17 x18 x19
a3a2


…
1·
128units
ReLu
Linear
=1
8 


2·

Target
x5x6x7x25 x26
= 4
3·
= 12 (= 4)
a1
 = 8
= −

Error
FIGURE 2. Front-end module operations and structure of the ANN.
channel quality, a conventional MLP fed with a sufficiently
large number of past samples, as the one presented in this
work, provides a better prediction accuracy.
A front-end module was implemented to carry out some
preliminary elaboration on the input features. Many types of
transformation were analyzed and tested, and the one that
provided the best results consists of a front-end function g(·)
that, given the past samples Ik, produces a sequence
g(Ik) = ai=1
i·Ns
k
X
j=k−i·Ns+1
xji∈1,..., Np
Ns
,(7)
where i·Nsis the number of outcomes used for computing
ANN input aiand Np/Nsrepresents the number of such
inputs. This sequence is then fed to the ANN.
As highlighted in the lower part of Fig. 2, every ANN
feature corresponds to the arithmetic mean of the most re-
cent outcomes of Ik, evaluated on sequences (intervals) of
increasing width. In particular, each such interval includes a
number of samples that is a multiple of the constant Ns. As
a consequence, the function fANN(Ik,w)can be split into
two components
fANN(Ik,w) = hANN (g(Ik),w),(8)
where g(Ik)is the front-end module given by (7) and
hANN(·,w)is a function that implements only the operations
related to ANN (upper part of Fig. 2).
To determine the optimal values for the parameters of the
ANN prediction model, the following optimization process is
used
w∗= arg min
w
|Dtr|−Nf
X
k=Np
tk−fANN(Ik,w)2,(9)
where wis the sets of weights and other quantities that
characterize the model (e.g., biases). Again, model parameter
estimation is assessed on the training database Dtr.
Using an ANN rather than existing statistical approaches
like SMA has the potential to improve prediction accuracy,
since it tries to infer a trend from the past and not just a single
4VOLUME 6, 2016
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3327523
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/

Alberto S. Colletto, Stefano Scanzio, et al.: Paper submitted to IEEE ACCESS
value. Linear and polynomial interpolation techniques were
also preliminarily analyzed, but their performance was found
to be quite poor, often worse than SMA.
By applying a linear transformation (similar to the one
carried out by the ANN input layer) to the input features
computed by the front-end module as per (7), evenly spaced
FDR estimates can be easily obtained that are evaluated on
adjacent intervals whose width is Ns·Ts(one minute in
our case). These values accurately describe the trend of the
channel quality in the recent past. Therefore, the ANN can
be seen as sort of a nonlinear finite input response (FIR) filter
applied to the FDR, which is potentially able to cope with
dynamically changing spectrum conditions.
For the above reasons, an ANN suitably trained on real
data could capture some properties of Wi-Fi transmission that
known theoretical models fail to describe properly. This is
exactly what our work is meant to determine.
V. TESTBED
The first part of our work, which lasted many months, con-
sisted in the experimental characterization of the wireless
spectrum as seen by real Wi-Fi equipment. Starting from the
outcomes of this part, several prediction models were then
created and their accuracy tested.
Investigation relied on several components. In the follow-
ing subsection, the hardware and software we developed for
the acquisition of the relevant databases are described. Then,
the architecture of the ANN and all details about its train-
ing are discussed in Subsection V-B, while Subsection V-C
illustrates the software we implemented to support training
starting from the huge amount of experimental data.
A. DATABASE ACQUISITION
The acquisition of databases relied on an experimental setup
made up of two Linux PCs, each one equipped with two
Wi-Fi adapters of type TP-Link TL-WDN4800 that comply
with IEEE 802.11n. Overall, there were four Wi-Fi STA,
each of which associated with a distinct AP located about
3÷4meters apart. In this experimental analysis we focused
on the four “canonical” channels 1, 5, 9, and 13 in the
2.4 GHz band. They are spaced wide enough to prevent any
adjacent channel interference (ACI) [43] effects, and every
pair STA/AP was configured to operate on one of these
channels.
Frame transmissions (frame size was set to 50 B) were
performed almost simultaneously by the four STAs, and were
triggered by the two Linux operating systems, which were
time-synchronized through the network time protocol (NTP)
and installed with the RT-Preempt Linux patch [44], [45] to
improve soft real-time capabilities.
The main goal of the testbed is to sample the channels’
conditions periodically, causing a negligible perturbation of
the related spectrum. To this purpose the device driver of the
STAs was modified in order to: 1) set transmissions to a fixed
bit rate of 54 Mb/s(consequently disabling the operations
of the Minstrel algorithm, which automatically optimizes the
TABLE 1. Length of training (Dtr) and test (Dte ) databases (days).
Channel Training (Dtr) Test (Dte) Total
ch1 21.2 12.8 34.0
ch5 21.9 10.6 32.5
ch9 24.9 15.2 40.1
ch13 25.5 15.2 40.7
All 93.5 53.8 147.3
transmission speed); 2) disable automatic frame retransmis-
sions (every frame is sent only once); 3) disable the backoff
procedure (frames are sent immediately when the channel is
idle); 4) disable the request to send/clear to send (RTS/CTS)
mechanism; 5) disable some specific features of the IEEE
802.11n version of the standard like frame aggregation, by
downgrading adapters’ operation to IEEE 802.11g.
Enforcing this behavior is possible thanks to the use of
the ath9k device driver along with the SDMAC framework
[18], [46], which allows transferring some relevant infor-
mation related to the transmission of a frame from kernel
space, where the driver executes, to the application devoted to
database acquisition, which runs in user space. Every time the
ACK frame associated with a data frame arrives at the sender
or the relevant timeout expires, the outcome xiis conveyed
by the driver to the application, which logs it.
The testbed was employed to characterize the wireless
spectrum in the 2.4 GHz band in our lab, which is shared
by a few tens of Wi-Fi networks (also including some wire-
less sensor networks) and the related nodes. The substantial
number of active nearby mobiles and notebooks exchanging
data over the air, which varies over time as researchers and
students keep entering and leaving facilities during the day,
make the spectrum conditions not stationary at all.
From the overall logs, the training and test databases were
extracted with no specific criteria (we are seeking for results
of general validity). Fig. 3 includes 8 timing diagrams that
report the FDR evaluated on these databases. As can be seen,
the overall amount of interference differs tangibly among
the considered channels, and the same holds for the related
interference patterns.
Table 1 summarizes the length (in days) of databases. The
training database for any channel spanned over at least three
weeks, while test databases lasted ten days or more. The
total amount of data used for training and test embraced 93.5
and 53.8days, respectively. Overall, all channels included,
databases covered 147.3days, corresponding to about 5
months.
B. ANN ARCHITECTURE
The topology of the ANN model is reported in Fig. 2.
Regarding the front-end module, we set Np= 14400 and
Ns= 120, and hence the input layer of the ANN consisted
of Np/Ns= 120 inputs. A number of additional experiments
were performed, like those for finding N∗
pin SMA, using
VOLUME 6, 2016 5
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3327523
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/

Alberto S. Colletto, Stefano Scanzio, et al.: Paper submitted to IEEE ACCESS
0 5 10 15 20
0.5
1.0
FDR
Ch1
Train
0 2 4 6 8 10 12
Ch1
Test
0 5 10 15 20
0.5
1.0
FDR
Ch5
0246810
Ch5
0 5 10 15 20 25
0.5
1.0
FDR
Ch9
0 2 4 6 8 10 12 14
Ch9
0 5 10 15 20 25
Days
0.5
1.0
FDR
Ch13
0 2 4 6 8 10 12 14
Days
Ch13
FIGURE 3. Timing diagrams about the measured FDR for the training (on the left) and test (on the right) databases on channels 1, 5, 9, and 13 (top to bottom).
1 3 5 7 9 11 13 15
Epochs
0.002
0.003
0.004
0.005
0.006
Loss
1 min
2 min
5 min
10 min
FIGURE 4. Loss vs. training epochs for channel 1 and Tfequal to 1, 2, 5, and
10 min.
larger Npvalues, but we did not observe any improvements
concerning accuracy. Conversely, using smaller values led to
a slight worsening. An important property of ANNs is that,
a specific optimization like (6) is not required, because the
ANN automatically trains the model parameters to weight
less those inputs that have a lower influence on the prediction.
The hidden layer of the neural networks is composed of
128 neurons of type ReLu, and there is a single linear output
neuron that provides the prediction ˜
tk. Training consists of 15
epochs, the momentum was set to 0, and the learning rate was
initialized to 0.01, halving at each epoch (i.e., 0.01,0.005,
0.0025, . . . ). In every epoch the patterns used to train the
model are reordered randomly. The batch size was set to 64,
which means that model weights are updated every 64 input
patterns.
The progress of the training procedure is sketched in
Fig. 4, where the loss (i.e., the MSE) was reported with
respect to the training epoch for channel 1 and with different
values of Tf. As can be seen, the most part of the loss decrease
takes place within the first 5 epochs; then, losses converge
asymptotically to fixed values. We decided to extend the
training up to 15 epochs because, although the improvements
are minimal, the obtained model is slightly better. Moreover,
by halving the learning rate at each epoch more stable results
are achieved, since in the final epoch very small adjustments
are brought to weights (and biases). The behavior for the
other three channels is quite similar. Finally, the loss com-
puted on the test database (by saving the ANN model at the
end of each epoch) showed a rather similar evolution, except
that it converges asymptotically to different loss values.
To obtain more reliable values of the ANN prediction
accuracy, any given configuration in the experiments below
(corresponding, e.g., to single rows in Table 3) was evaluated
five times. Every time, the ANN was trained from scratch and
then tested, leaving databases unchanged. Reported results
have been obtained by evaluating the relevant statistics on
the concatenation of all the prediction errors obtained from
tests, for the five separate repetitions of every specific config-
uration.
C. TRAINING SOFTWARE
The training and test software was written in python and
makes use of the Keras module, which is included in
tensorflow. The ANN model weights are randomly ini-
tialized at the beginning of the training with the Glorot
normal initializer, and the SGD optimizer was used for
weights update.
Due to the huge dimension of the training database, a spe-
cific software was developed to manage such “big data”. The
main problem stems from the fact that every outcome in the
input database involves the generation of a pattern composed
of Np/Ns= 120 inputs and one target. In addition, at each
epoch such patterns must be supplied randomly to the ANN
for training.
Loading all patterns in memory simultaneously is not
feasible due to memory limits, and randomly loading them
from the hard disk is too time-consuming. Therefore, we
specifically developed a suitable software that splits the train-
ing database into groups that include Nmpatterns at most. At
the beginning of every epoch, a random number is assigned
to each group, which permits to identify the order in which
they are selected. Following this order, the program loads
Nggroups for a total of Ng·Nmpatterns into memory,
6VOLUME 6, 2016
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3327523
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/

Alberto S. Colletto, Stefano Scanzio, et al.: Paper submitted to IEEE ACCESS
shuffles them, and executes the training process. Then, the
same process is repeated using another set of Nggroups.
When all groups are trained the epoch finishes, and the
process restarts by assigning new random numbers to each
group. The size Nmis a compromise between speedup and
randomness of the patterns provided to the ANN (a higher
value of Nmmakes operations faster). Instead, Ngdepends
on the available memory of the PC used for training, and must
be maximized. Increasing the number Ngof groups loaded
contextually into memory increases the randomness of the
patterns provided to the ANN, because groups are randomly
selected over the entire training database. The values used in
this paper are Nm= 100000 and Ng= 10.
VI. RESULTS
Starting from the training and test databases (Dch
tr and Dch
te )
we obtained from the above experimental testbed on every
Wi-Fi channel ch ∈ {1,5,9,13}, a number of campaigns
were carried out to analyze how much aspects like the pre-
diction model (plain SMA heuristic vs. ANN), the ANN ar-
chitecture (single output vs. multiple output), and the kind of
training (specific channel-dependent vs. generalized channel-
independent) impact on accuracy. Doing so makes results
comparable.
A. SPECIFICALLY PARAMETERIZED SMA
In the initial campaign a simple moving average was used for
prediction. This is a very basic approach which assumes that,
from a probabilistic viewpoint, behavior in the near future is
the same as the recent past. Since we wished to make a fair
comparison with ANNs, the width of the moving window on
which the average is evaluated was not defined once and for
all, but was instead specifically set for every given channel
and for every given horizon based on a preliminary training
phase according to (6).
The four plots in Fig. 5 show the SMA prediction error
(MSE) on channels 1, 5, 9, and 13 for two different Nf
values (5and 10 min) when Npis varied. As can be seen,
a minimum can be always singled out clearly, whose position
provides the optimal value N∗
p(for the considered conditions,
it lies in the range from 120 to 1000). This can be explained
by considering the two main phenomena that contribute to
the prediction error:
1) The spectrum is non-stationary, hence knowing the
past is not enough to characterize the future precisely.
The farther the horizon, and the deeper the past on
which the SMA is computed, the more the behavior
of the wireless link may change in the meanwhile,
also depending on the specific spectrum dynamics.
This contribution increases as Tpand Tfgrow.
2) The target we used for training (FDR) is obtained by
averaging the outcomes of Nftransmissions occurring
in the future. The lower the number of samples, the
higher the intrinsic variability affecting the estimation
of the related success probability. The very same holds
for the past interval, which includes Npoutcomes.
These contributions decrease as Tpand Tfgrow.
It is worth noting that the above reasoning mostly applies
also to the case where ANNs are used for prediction.
Results are reported in Table 2, which consists of four
parts, each one referring to a specific channel. Every such part
is split in four rows, which refer to different horizons Tfequal
to 1, 2, 5, and 10 min, respectively. For every row several
metrics about prediction accuracy are reported, evaluated on
the related test database Dch
te . Besides the mean squared error
(MSE), the absolute error is also considered, and in particular
its mean value (MAE), as well as its 90 and 95 percentiles.
On the rightmost column, parameter N∗
pis included.
In the following, prediction accuracy refers to the MAE,
unless otherwise stated. For the reasons described above, it
depends on both the considered channel and the horizon Tf
on which the target FDR is evaluated. By looking at the table,
one can see that the best accuracy (µ|e|= 1.91%) is achieved
for channel 5 when the future horizon spans over 5 minutes.
Behavior of channel 9 is the hardest to predict: in this case the
best estimates are obtained for a future interval of 2 minutes,
and when the horizon is enlarged to 10 minutes the absolute
error grows and exceeds 4.4%.
This can be explained by looking at the time diagrams
in Fig. 3. As can be seen, the failure rate on channel 5 is
quite low, whereas it is sensibly worse on channel 9. Non-
negligible interference also affects channels 1 and 13, but in
these cases the failure patterns are more “regular” and hence
they can be predicted to a better extent. Conversely, channel
9 is characterized by faster spectrum dynamics, with rapidly
changing interference.
B. SPECIFICALLY TRAINED ANN
In this campaign a distinct ANN was considered for every
channel and for every horizon. As for the SMA in the
previous section, the databases used for training and test refer
to the same channel (channel-dependent prediction model),
which implies that they are strongly correlated. In fact, al-
though the interference pattern observed by the testbed on
any channel is likely to vary over time, this implying that it
differs in the training and test phases (displaced by several
weeks), it depends on the same set of nearby APs (deployed
in fixed positions) and similar sets of STAs. Intuitively, doing
so should provide the best prediction accuracy, since every
ANN is optimally trained using data acquired in conditions
that mostly resemble those it will encounter when used for
prediction. For this reason the results reported here constitute
sort of a best case for ANNs, against which the outcomes in
the following sections have to be checked.
Results are shown in Table 3, whose structure resembles
the one used for SMA. The rightmost column reports the
winning ratio W, that specifies how many times the ANN
provided a better accuracy than SMA. However, this quantity
is not particularly interesting, because it does not consider
how much the prediction of the loser is actually worse.
From the point of view of the MAE, the best accuracy is
VOLUME 6, 2016 7
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3327523
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/

Alberto S. Colletto, Stefano Scanzio, et al.: Paper submitted to IEEE ACCESS
0 500 1000 1500 2000 2500
Np
0.0020
0.0025
0.0030
0.0035
0.0040
0.0045
0.0050
0.0055
0.0060
MSE
N
*
p
5
min
= 320
N
*
p
10
min
= 370
Plot 4.a: ch1
Nf=5 min
Nf=10 min
0 500 1000 1500 2000 2500
Np
0.0010
0.0015
0.0020
0.0025
0.0030
MSE
N
*
p
5
min
= 310
N
*
p
10
min
= 340
Plot 4.b: ch5
Nf=5 min
Nf=10 min
0 500 1000 1500 2000 2500
Np
0.010
0.015
0.020
0.025
0.030
MSE
N
*
p
5
min
= 120
N
*
p
10
min
= 120
Plot 4.c: ch9
Nf=5 min
Nf=10 min
0 500 1000 1500 2000 2500
Np
0.0010
0.0015
0.0020
0.0025
0.0030
MSE
N
*
p
5
min
= 960
N
*
p
10
min
= 1000
Plot 4.d: ch13
Nf=5 min
Nf=10 min
FIGURE 5. SMA training: prediction error (MSE) vs. past interval width Npfor channels 1, 5, 9, and 13 (optimal N∗
pvalues are highlighted).
…
128 hidden
units
120 inputs …
1 output
…
…
 −  ( )  − 
4 output

= 1
min

= 2
min

= 5
min

= 10
min
FIGURE 6. Single-output and multi-output ANN architectures: predictions are
made for a single vs. multiple (four) time horizons.
obtained for Tf= 2 min on channels 1and 9, whereas on
channel 5it is seemingly found when Tfapproaches 10 min
(µ|e|= 1.70%). Finally, on channel 13 error is minimal when
Tfexceeds 10 min. Since the variance of the target depends
on the number of samples on which FDR is evaluated (and
hence, on Tf), this behavior can be explained by the faster
dynamics of channels 1and 9, which make predictions over
larger horizons worse.
By comparing prediction accuracy on the different chan-
nels, we can observe that the error on channel 5 is generally
small, whereas it is quite large on channel 9. ANN always
managed to outperform SMA. This is a relevant result, and
implies that ML, even in its simplest form, goes beyond
the simple assumption that the future resembles the past. In
particular, it proves to be able to model non-trivial hidden
aspects of the spectrum in the presence of Wi-Fi traffic, which
leads to better predictions.
C. MULTI-TARGET ANN
In this campaign, we used a single ANN with four outputs
for every channel (channel-dependent prediction model) to
perform contextual FDR predictions for all time horizons.
Changes with respect to the previous campaign are purely
architectural. In fact, the same databases as before were used
for training and test. The single-output architecture of the
previous campaign and the multi-output architecture of the
current one are depicted side by side in Fig. 6.
Results are reported in Table 4. As can be seen, accuracy
TABLE 2. Prediction accuracy by using 16 SMAs where the window width is
specifically optimized for channels 1, 5, 9, and 13 (channel-dependent
prediction model) and future horizons 1, 2, 5, and 10 minutes.
Channel Tfµe2µ|e||e|p90 |e|p95 N∗
p
[min] [·10−3][%]
ch1
1 2.064 3.31 7.08 9.16 240
21.789 2.95 6.45 8.83 290
5 2.059 3.03 7.04 10.00 320
10 2.794 3.49 8.85 12.06 370
ch5
1 1.115 2.51 5.21 6.45 310
2 0.842 2.11 4.36 5.43 310
50.793 1.91 3.95 5.02 310
10 0.976 1.97 4.09 5.46 340
ch9
1 2.993 3.83 7.82 9.67 130
22.931 3.52 6.98 8.81 130
5 4.072 3.76 7.16 9.66 120
10 6.291 4.43 8.16 14.99 120
ch13
1 2.379 3.86 7.98 9.60 970
2 1.555 3.09 6.40 7.76 950
5 1.107 2.56 5.33 6.60 960
10 1.029 2.43 5.09 6.43 1000
Note: boldface denotes best cases.
of a multi-output ANN (MAE is taken again into account
as the metric for comparison) somehow resembles what
provided by a plurality of single-output ANNs, but the latter
typically show a lower error. There are some exceptions,
e.g., predictions over the next 5 min for channels 5 and 13,
but they are mostly inessential. Sometimes, the multi-output
ANN behaved worse than SMA. Summing up, a simpler (and
cheaper, in terms of both memory occupation and training
time) ANN implementation is possible at the price of a
diminished accuracy.
D. GENERALIZED ANN TRAINING
A criticism about the use of ANNs to predict the quality
of wireless channels concerns their training. In the above
campaigns, ANNs were trained on a channel-by-channel
basis, to reflect differences between the related spectrum
conditions. To face spectrum non-stationarity in the long
8VOLUME 6, 2016
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3327523
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/

Alberto S. Colletto, Stefano Scanzio, et al.: Paper submitted to IEEE ACCESS
TABLE 3. Prediction accuracy by using 16 single-output ANNs specifically
trained on channels 1, 5, 9, and 13 (channel-dependent prediction model) and
future horizons 1, 2, 5, and 10 minutes.
Channel Tfµe2µ|e||e|p90 |e|p95 W
[min] [·10−3][%] [%]
ch1
1 1.973 3.22 6.96 9.23 54.2
21.670 2.86 6.32 8.65 54.4
5 1.880 2.92 6.95 9.64 55.4
10 2.254 3.19 8.00 10.71 57.7
ch5
1 1.041 2.40 4.94 6.14 55.7
2 0.768 1.97 4.04 5.10 56.8
50.664 1.71 3.51 4.53 57.3
10 0.759 1.70 3.43 4.62 59.1
ch9
1 2.662 3.51 7.03 8.86 58.6
22.514 3.15 6.12 8.01 59.3
5 3.184 3.27 6.03 9.25 58.2
10 4.434 3.95 7.18 12.23 53.3
ch13
1 2.243 3.76 7.77 9.31 53.7
2 1.416 2.96 6.12 7.40 54.3
5 0.941 2.38 4.93 6.04 55.6
10 0.813 2.15 4.54 5.68 59.4
Note: boldface denotes best cases.
TABLE 4. Prediction accuracy by using 4 multiple-output ANNs specifically
trained on channels 1, 5, 9, and 13 (channel-dependent prediction model)
providing contextual outputs for future horizons 1, 2, 5, and 10 minutes.
Channel Tfµe2µ|e||e|p90 |e|p95 W
[min] [·10−3][%] [%]
ch1
1 2.171 3.35 7.36 9.80 51.9
21.833 2.99 6.75 9.23 52.0
5 1.963 2.99 7.17 9.88 53.8
10 2.372 3.27 8.28 11.03 55.8
ch5
11.090 2.43 5.01 6.23 54.6
2 0.854 2.03 4.17 5.30 55.0
50.654 1.69 3.45 4.47 58.5
10 0.767 1.70 3.45 4.65 59.0
ch9
1 2.735 3.55 7.09 8.98 57.6
22.587 3.18 6.15 8.13 58.5
5 3.312 3.37 6.32 9.59 56.7
10 4.726 4.18 7.80 13.22 50.1
ch13
1 2.261 3.77 7.80 9.35 52.9
2 1.437 2.98 6.17 7.46 53.8
5 0.942 2.37 4.92 6.07 56.4
10 0.847 2.20 4.63 5.86 56.3
Note: boldface denotes best cases.
term, training could be reiterated periodically by the involved
devices, e.g., by automatically invoking a suitable procedure.
Clearly, doing so is not trivial at all, and demands for the
permanent availability of computational resources on APs
and STAs.
It could be interesting to determine whether or not a
generic training can be performed for ANNs, in order to char-
acterize Wi-Fi interference independently of the channel, and
hence of the traffic actually found on it (channel-independent
prediction model). To this purpose, we trained a single ANN
for every horizon with a database obtained by merging all the
training data used in the previous campaigns
Dall
tr =[
c∈{1,5,9,13}
Dc
tr,(10)
where the union symbol denotes the ordered concatenation
of sequences. Then, we evaluated the prediction accuracy for
every single test database Dch
te .
This campaign permitted to appreciate how much a
specific, channel-dependent training, improves accuracy.
Channel-independent training demands for a noticeably
lower effort, and also provides a more generic solution. In
fact, the interference patterns observed by the testbed on the
different channels are mostly uncorrelated, since their fre-
quencies do not overlap. This is not completely true because
of the presence of nearby equipment (APs and STAs, whose
placement and traffic was not under our control) tuned on
non-canonical channels (other than 1,5,9, and 13). Luckily,
there were only a few of them. For example, an AP operating
on channel 3interferes with both channels 1and 5of the
testbed, which implies some correlation between the related
databases. Likewise, the contextual presence of traffic on
channel 3and a 40 MHz link obtained by bonding channels 6
and 10 may create some dependency between channels 1and
13. In theory, the farther the channels, the lower correlation.
In the above campaign, the training and test databases are
still correlated, since the former contain data acquired on ev-
ery channel. To make them completely uncorrelated, at least
in theory (truly-channel-independent prediction model) a
further campaign was carried out where the training database
for every channel includes all databases with the exception of
the one related to the channel itself
D¬ch
tr =[
c∈{1,5,9,13}\ch
Dc
tr.(11)
This is expected to represent the worst case for what concerns
ANN training.
Results are reported in Table 5, which is split in 16 parts,
one for every channel ch and future horizon Tf. Every such
part includes in turn three rows. The first row refers to the
channel-dependent prediction model, and shows the same re-
sults as Table 3 (included here to ease comparison), whereas
the second is related to the channel-independent prediction
model. Finally, the third row concerns the above case of bad
training using databases created according to (11).
As can be seen, not necessarily specific training always
bests generalized training. Curiously, excluding the channel
under test from training did not always led to the worst
results. Channels 5 and, especially, 13 are those which ben-
efit more from a specialized training. By looking at time
diagrams in Fig. 3, this likely depends on the fact that the
interference patterns used for test and training are similar.
Conversely, predictions that rely on a generalized training are
typically more accurate on channel 9. By referring again to
the time diagrams about the FDR, this is probably due to the
VOLUME 6, 2016 9
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3327523
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/

Alberto S. Colletto, Stefano Scanzio, et al.: Paper submitted to IEEE ACCESS
TABLE 5. Prediction accuracy on channels 1, 5, 9, and 13 and for future horizons 1, 2, 5, and 10 minutes by using ANNs trained with the (specific) related
databases Dch
tr (channel-dependent prediction model), a single (generic) combined database Dall
tr (channel-independent prediction model), and (specific)
combined databases D¬ch
tr purposely created to be as independent as possible from the channel under test (truly-channel-independent prediction model).
Channel Training Tfµe2µ|e||e|p90 |e|p95 W Tfµe2µ|e||e|p90 |e|p95 W
(test) database [min] [·10−3][%] [%] [min] [·10−3][%] [%]
ch1
Dch
tr
1
1.973 3.22 6.96 9.23 54.2
5
1.880 2.92 6.95 9.64 (55.4)
Dall
tr 1.882 3.19 6.81 8.87 53.8 1.735 2.87 6.52 8.97 53.6
D¬ch
tr 1.887 3.20 6.81 8.88 53.4 1.760 2.91 6.56 9.01 52.0
Dch
tr
2
1.670 2.86 6.32 8.65 54.4
10
2.254 3.19 8.00 10.71 57.7
Dall
tr 1.589 2.83 6.14 8.28 53.5 2.194 3.25 7.71 10.37 52.9
D¬ch
tr 1.608 2.85 6.16 8.31 52.3 2.213 3.31 7.70 10.31 50.4
ch5
Dch
tr
1
1.041 2.40 4.94 6.14 55.7
5
0.664 1.71 3.51 4.53 57.3
Dall
tr 1.045 2.47 5.11 6.28 51.4 0.652 1.82 3.70 4.65 52.4
D¬ch
tr 1.060 2.49 5.15 6.30 50.2 0.703 1.93 3.86 4.80 47.7
Dch
tr
2
0.768 1.97 4.04 5.10 56.8
10
0.759 1.70 3.43 4.62 (59.1)
Dall
tr 0.751 2.04 4.20 5.19 52.6 0.777 1.92 3.89 4.97 49.2
D¬ch
tr 0.773 2.07 4.28 5.26 50.9 0.895 2.18 4.25 5.29 40.5
ch9
Dch
tr
1
2.662 3.51 7.03 8.86 58.6
5
3.184 3.27 6.03 9.25 58.2
Dall
tr 2.649 3.46 6.92 8.77 60.5 3.191 3.11 5.80 9.41 64.2
D¬ch
tr 3.214 3.62 7.11 9.51 58.5 4.132 3.40 6.95 12.59 62.6
Dch
tr
2
2.514 3.15 6.12 8.01 59.3
10
4.434 3.95 7.18 12.23 53.3
Dall
tr 2.511 3.06 5.91 7.90 62.3 4.485 3.63 7.09 13.46 62.9
D¬ch
tr 3.140 3.26 6.26 9.05 60.0 5.684 3.92 9.65 17.1 63.7
ch13
Dch
tr
1
2.243 3.76 7.77 9.31 53.7
5
0.941 2.38 4.93 6.04 55.6
Dall
tr 2.499 3.97 8.21 9.84 47.6 1.156 2.66 5.53 6.68 47.6
D¬ch
tr 2.628 4.06 8.43 10.1 46.2 1.271 2.80 5.81 7.03 45.0
Dch
tr
2
1.416 2.96 6.12 7.40 54.3
10
0.813 2.15 4.54 5.68 59.4
Dall
tr 1.646 3.21 6.62 7.95 47.1 1.053 2.52 5.28 6.44 47.5
D¬ch
tr 1.768 3.33 6.87 8.24 45.3 1.179 2.69 5.62 6.78 44.7
Note: boldface denotes best cases, underline denotes worst cases.
fact that this channel suffers from a severe and irregularly-
shaped interference. Therefore, widening the variety of cases
on which training is performed is beneficial. The same holds
for channel 1, where the patterns used for training and test
look dissimilar (seemingly, something happened that made
the spectrum conditions change tangibly in the course of the
experiment). In this case, performing training on all channels
with the exclusion of the one under test occasionally provided
slightly better accuracy.
From above results one can see that generalized training
often constitutes a valid alternative to channel-specific train-
ing. This is another relevant outcome of this work, since the
former is way simpler to implement than the latter. Likely,
this depends on the fact that ANN behavior is quite complex
and cannot be described easily. Therefore, experimental cam-
paigns like those described in this paper are the only way to
assess their performance in the wild.
VII. PRACTICAL FEASIBILITY
A relevant point when ANNs are exploited to predict channel
behavior in real equipment is the trade-off between prediction
accuracy and computational complexity. Generally speaking,
increasing the overall number Npof past outcomes achieves
better accuracy, because the ANN is provided a larger amount
of information about channel’s conditions. Accuracy is ex-
pected to improve also when Nsis shrunk, as doing so
makes time resolution of FDR statistics more fine-grained.
However, reducing Nstoo much might increase variability
when computing averages aion short time intervals (when i
is low).
A. COMPLEXITY
The number of input features fed to the ANN, as specified by
(7), is Np/Ns. The asymptotic complexity of the ANN input
layer (and hence, its contribution to the time taken to perform
a test) is linear versus this quantity.
A first experiment was run to evaluate the average exe-
cution time ¯
Tpfor a single test operation. Three different
architectures were considered: the first and the second were
based on Intel® CoreTM CPUs, and in particular a desktop
based on an i3-10105 @3.70GHz and a notebook based on an
i7-11800H @2.30GHz, respectively, whereas the third was a
Raspberry PI 2 based on an ARMv7 Processor rev 4 (v7l).
The ANN was tested by letting Np= 14400 and varying Ns
from 30 to 480 in steps of 5. For every Nsvalue the test was
repeated one million times, and the arithmetic mean of the
related execution times was computed. Results for Intel Core
and ARM are shown in plots a and b of Fig. 7, respectively.
As expected, ¯
Tpis inversely proportional to Ns.
10 VOLUME 6, 2016
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3327523
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/

Alberto S. Colletto, Stefano Scanzio, et al.: Paper submitted to IEEE ACCESS
0 100 200 300 400 500
Ns
0
10
20
30
40
50
60
70
Tp
[
s
]
Plot a
Intel i7-11800H
Intel i3-10105
0 100 200 300 400 500
Ns
0
100
200
300
400
500
600
700
800
Tp
[
s
]
Plot b
ARMv7
FIGURE 7. Mean execution time ¯
Tpvs. Ns(Intel Core and ARMv7).
Concerning differences among the different architectures,
when Ns= 120 (as in the experiments described in the
previous section) ¯
Tpis 11.4µsfor the Intel i7, 14.6µsfor
the Intel i3, and 142.6µsfor the ARMv7. These architectures
(particularly the latter) are mostly compatible with high-end
commercial APs software implementations, and execution
times are low enough to enable adoption in real devices.
B. ACCURACY
As said above, the value selected for Nsalso impacts on
prediction accuracy. For this reason, a second experiment was
carried out where Nswas set equal to 30,60,120,240, and
480 (corresponding to basic interval widths from 15 s to 4
minutes). We considered, as an example, the specific case
of a single-output ANN specifically trained (and tested) on
channel 1for a future horizon equal to 2minutes (cf. Table 3).
Again, Np= 14400, i.e., the past interval was kept fixed to
two hours.
Results, reported in Table 6, describe the relative variation
(as a percentage) of accuracy metrics with respect to the
reference case when Ns= 120. They clearly highlight that,
when Nsis lowered down to 60 (or even 30), improvements
on µe2(the performance indicator minimized by the ANN)
are negligible compared to the increase of computation times
(see plots of Fig. 7). Increasing Nsto 240 (or 480) leads, on
the one side, to a sensible decrease of computation times, but
TABLE 6. Relative improvement of the prediction accuracy by using a
single-output ANNs trained and tested on channels 1 (channel-dependent
prediction model) with Tf= 2 min and variable value of
Ns= 30,60,120,240,480.
Channel Nsµe2µ|e||e|p90 |e|p95
[min] [%] [%]
ch1
30 -0.56 +0.21 -3.11 -2.73
60 -0.13 +0.20 -3.11 -2.50
120 0 0 0 0
240 +3.09 +0.31 -3.18 -1.04
480 +12.71 +3.11 +0.15 +5.48
on the other it causes a substantial performance degradation,
especially for what concerns µe2. In conclusion, the value
Ns= 120 we selected appears to be a good compromise
between prediction accuracy and computational complexity.
VIII. CONCLUSIONS
Despite wireless networks are more and more used in many
different context, due to their ability to provide wire-free
connections, they are not able yet to offer applications the
same quality of service as conventional wired solutions like
Ethernet. While throughput has increased steadily over the
past decades, to the point that, in theory, the performance
of currently available technologies like Wi-Fi 6 and 5G is
comparable (and, sometimes, higher) than Gigabit Ethernet,
transmission on air lags behind cables for what concerns
dependability and timeliness.
Deterministic MAC mechanisms, like trigger frames in
Wi-Fi 6, effectively counteract intra-network interference.
Unfortunately, they can do little or nothing against distur-
bance due to external sources. In this case, knowledge about
the spectrum conditions in the recent past could be exploited
by ML to statistically improve network behavior beyond
what deterministic MACs can reasonably do.
This paper is aimed to present our most recent findings in
this field. In particular, we predicted the mean failure rate on
a wireless link over specific future time horizons by means
of ANNs trained on the outcomes of the past transmission
attempts. Results show that prediction accuracy achieved by
ML is better than conventional methods that rely on moving
averages.
A second question we tried to answer is how much training
impacts on accuracy. In particular, we compared the per-
formance of ANNs specifically trained on the channel on
which they will be used and those where training exploits
a generic database that covers all channels. Results show
that accuracy is mostly comparable, and channel-independent
training could be the best option in those cases where spec-
trum conditions are likely to change (as happens over long
time spans). Advantages of channel-independent prediction
are undeniable: in fact, the same ANN could be employed in
a variety of scenarios, e.g., by integrating it in the network
equipment directly.
VOLUME 6, 2016 11
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3327523
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/

Alberto S. Colletto, Stefano Scanzio, et al.: Paper submitted to IEEE ACCESS
Among the activities we plan for our future work, the
most important are the analysis of the proposed methodology
under different conditions, including varying traffic loads and
dynamically changing network configurations, as well as the
study of ANN model scalability, its extendability to protocols
other than Wi-Fi, and the suitability of alternative methods
(for instance, probabilistic ones based on Markov chains).
ACKNOWLEDGEMENT
This work was partially supported by the European Union
under the Italian National Recovery and Resilience Plan
(NRRP) of NextGenerationEU, partnership on “Telecom-
munications of the Future” (PE00000001 - program
“RESTART”).
Computational resources were provided by HPC@POLITO,
a project of Academic Computing within the Department
of Control and Computer Engineering at the Politecnico di
Torino (http://www.hpc.polito.it).
Both Stefano Scanzio, and Alberto Salvatore Colletto,
have to be considered as first authors.
REFERENCES
[1] Q. Huang, H. Chen, and Q. Zhang, “Joint Design of Sensing and Com-
munication Systems for Smart Homes,” IEEE Network, vol. 34, no. 6, pp.
191–197, 2020.
[2] W. Na, N.-N. Dao, and S. Cho, “Mitigating wifi interference to improve
throughput for in-vehicle infotainment networks,” IEEE Wireless Com-
munications, vol. 23, no. 1, pp. 22–28, 2016.
[3] M. B. Attia, K. K. Nguyen, and M. Cheriet, “Dynamic QoS-Aware
Scheduling for Concurrent Traffic in Smart Home,” IEEE Internet of
Things Journal, vol. 7, no. 6, pp. 5412–5425, 2020.
[4] S. Kwankajornkeat and C. Aswakul, “Differential Private Motion Sensor
and Wasted Energy in Building Energy Management System,” IEEE
Access, vol. 10, pp. 486–501, 2022.
[5] S. Scanzio, G. Cena, and A. Valenzano, “Enhanced Energy-Saving Mech-
anisms in TSCH Networks for the IIoT: The PRIL Approach,” IEEE
Transactions on Industrial Informatics, pp. 1–11, 2022.
[6] A. Al-Fuqaha, M. Guizani, M. Mohammadi, M. Aledhari, and M. Ayyash,
“Internet of Things: A Survey on Enabling Technologies, Protocols, and
Applications,” IEEE Communications Surveys Tutorials, vol. 17, no. 4,
pp. 2347–2376, 2015.
[7] H. Cañas, J. Mula, M. Díaz-Madroñero, and F. Campuzano-Bolarín, “Im-
plementing Industry 4.0 principles,” Computers Industrial Engineering,
vol. 158, p. 107379, 2021.
[8] P. K. R. Maddikunta, Q.-V. Pham, P. B, N. Deepa, K. Dev, T. R.
Gadekallu, R. Ruby, and M. Liyanage, “Industry 5.0: A survey on
enabling technologies and potential applications,” Journal of Industrial
Information Integration, vol. 26, p. 100257, 2022. [Online]. Available:
https://www.sciencedirect.com/science/article/pii/S2452414X21000558
[9] E. Sisinni, A. Saifullah, S. Han, U. Jennehag, and M. Gidlund, “Industrial
Internet of Things: Challenges, Opportunities, and Directions,” IEEE
Transactions on Industrial Informatics, vol. 14, no. 11, pp. 4724–4734,
2018.
[10] S. Scanzio, L. Wisniewski, and P. Gaj, “Heterogeneous and
dependable networks in industry – A survey,” Computers
in Industry, vol. 125, p. 103388, 2021. [Online]. Available:
https://www.sciencedirect.com/science/article/pii/S0166361520306229
[11] D. C. Nguyen, M. Ding, P. N. Pathirana, A. Seneviratne, J. Li, D. Niyato,
O. Dobre, and H. V. Poor, “6G Internet of Things: A Comprehensive
Survey,” IEEE Internet of Things Journal, vol. 9, no. 1, pp. 359–383, 2022.
[12] A. Garcia-Rodriguez, D. López-Pérez, L. Galati-Giordano, and G. Geraci,
“IEEE 802.11be: Wi-Fi 7 Strikes Back,” IEEE Communications Maga-
zine, vol. 59, no. 4, pp. 102–108, 2021.
[13] L. Leonardi, L. Lo Bello, and G. Patti, “LoRa support for long-range real-
time inter-cluster communications over Bluetooth Low Energy industrial
networks,” Computer Communications, vol. 192, pp. 57–65, 2022.
[14] G. Cena, C. G. Demartini, M. Ghazi Vakili, S. Scanzio,
A. Valenzano, and C. Zunino, “Evaluating and Modeling IEEE
802.15.4 TSCH Resilience against Wi-Fi Interference in New-
Generation Highly-Dependable Wireless Sensor Networks,” Ad
Hoc Networks, vol. 106, p. 102199, 2020. [Online]. Available:
https://www.sciencedirect.com/science/article/pii/S1570870519310509
[15] K. E. Jeon, J. She, P. Soonsawad, and P. C. Ng, “BLE Beacons for Internet
of Things Applications: Survey, Challenges, and Opportunities,” IEEE
Internet of Things Journal, vol. 5, no. 2, pp. 811–828, 2018.
[16] G. Cena, S. Scanzio, and A. Valenzano, “Seamless Link-Level Redun-
dancy to Improve Reliability of Industrial Wi-Fi Networks,” IEEE Trans-
actions on Industrial Informatics, vol. 12, no. 2, pp. 608–620, 2016.
[17] J. Li, S. K. Bose, and G. Shen, “Cooperative Resource Scheduling for
Time-Sensitive Services in an Integrated XGS-PON and Wi-Fi 6 Net-
work,” IEEE Communications Letters, vol. 26, no. 6, pp. 1338–1342,
2022.
[18] G. Cena, S. Scanzio, and A. Valenzano, “SDMAC: A Software-Defined
MAC for Wi-Fi to Ease Implementation of Soft Real-Time Applications,”
IEEE Transactions on Industrial Informatics, vol. 15, no. 6, pp. 3143–
3154, 2019.
[19] M. Friesen, L. Wisniewski, and J. Jasperneite, “Machine Learning for
Zero- Touch Management in Heterogeneous Industrial Networks - A
Review,” in 2022 IEEE 18th International Conference on Factory Com-
munication Systems (WFCS), 2022, pp. 1–8.
[20] C. Deng, X. Fang, X. Han, X. Wang, L. Yan, R. He, Y. Long, and
Y. Guo, “IEEE 802.11be Wi-Fi 7: New Challenges and Opportunities,”
IEEE Communications Surveys Tutorials, vol. 22, no. 4, pp. 2136–2166,
2020.
[21] G. Cena, S. Scanzio, and A. Valenzano, “Ultra-Low Power Wireless
Sensor Networks Based on Time Slotted Channel Hopping with
Probabilistic Blacklisting,” Electronics, vol. 11, no. 3, 2022. [Online].
Available: https://www.mdpi.com/2079-9292/11/3/304
[22] X. Vilajosana, T. Watteyne, T. Chang, M. Vuˇ
cini´
c, S. Duquennoy, and
P. Thubert, “IETF 6TiSCH: A Tutorial,” IEEE Communications Surveys
Tutorials, vol. 22, no. 1, pp. 595–615, 2020.
[23] S. Scanzio, M. G. Vakili, G. Cena, C. G. Demartini, B. Montrucchio,
A. Valenzano, and C. Zunino, “Wireless Sensor Networks and TSCH:
A Compromise Between Reliability, Power Consumption, and Latency,”
IEEE Access, vol. 8, pp. 167 042–167 058, 2020.
[24] C.-Y. Li, S.-C. Chen, C.-T. Kuo, and C.-H. Chiu, “Practical Machine
Learning-Based Rate Adaptation Solution for Wi-Fi NICs: IEEE 802.11ac
as a Case Study,” IEEE Transactions on Vehicular Technology, vol. 69,
no. 9, pp. 10 264–10277, 2020.
[25] G. Cena, S. Scanzio, and A. Valenzano, “Experimental Evaluation of Tech-
niques to Lower Spectrum Consumption in Wi-Red,” IEEE Transactions
on Wireless Communications, vol. 18, no. 2, pp. 824–837, 2019.
[26] G. Cena, S. Scanzio, D. Cavalcanti, and V. Frascolla, “Seamless Redun-
dancy for High Reliability Wi-Fi,” in 19th IEEE International Conference
on Factory Communication Systems (WFCS 2023), 2023, pp. 1–4.
[27] S. Szott, K. Kosek-Szott, P. Gawłowicz, J. T. Gómez, B. Bellalta,
A. Zubow, and F. Dressler, “Wi-Fi Meets ML: A Survey on Improving
IEEE 802.11 Performance With Machine Learning,” IEEE Communica-
tions Surveys Tutorials, vol. 24, no. 3, pp. 1843–1893, 2022.
[28] I. Ahmad, S. Shahabuddin, T. Sauter, E. Harjula, T. Kumar, M. Meisel,
M. Juntti, and M. Ylianttila, “The Challenges of Artificial Intelligence in
Wireless Networks for the Internet of Things: Exploring Opportunities for
Growth,” IEEE Industrial Electronics Magazine, vol. 15, no. 1, pp. 16–29,
2021.
[29] L. Song and A. Striegel, “Leveraging frame aggregation to improve access
point selection,” in 2017 IEEE Conference on Computer Communications
Workshops (INFOCOM WKSHPS), 2017, pp. 325–330.
[30] C. Pei, Z. Wang, Y. Zhao, Z. Wang, Y. Meng, D. Pei, Y. Peng, W. Tang,
and X. Qu, “Why it takes so long to connect to a WiFi access point,” in
IEEE INFOCOM 2017 - IEEE Conference on Computer Communications,
2017, pp. 1–9.
[31] Z. Han, T. Lei, Z. Lu, X. Wen, W. Zheng, and L. Guo, “Artificial
Intelligence-Based Handoff Management for Dense WLANs: A Deep
Reinforcement Learning Approach,” IEEE Access, vol. 7, pp. 31688–
31 701, 2019.
[32] E. Zeljkovi´
c, N. Slamnik-Kriještorac, S. Latré, and J. M. Marquez-Barja,
“ABRAHAM: Machine Learning Backed Proactive Handover Algorithm
Using SDN,” IEEE Transactions on Network and Service Management,
vol. 16, no. 4, pp. 1522–1536, 2019.
12 VOLUME 6, 2016
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3327523
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/

Alberto S. Colletto, Stefano Scanzio, et al.: Paper submitted to IEEE ACCESS
[33] H. Feng, Y. Shu, S. Wang, and M. Ma, “SVM-Based Models for Predicting
WLAN Traffic,” in 2006 IEEE International Conference on Communica-
tions, vol. 2, 2006, pp. 597–602.
[34] A. Thapaliya, J. Schnebly, and S. Sengupta, “Predicting Congestion Level
in Wireless Networks Using an Integrated Approach of Supervised and
Unsupervised Learning,” in 2018 9th IEEE Annual Ubiquitous Comput-
ing, Electronics Mobile Communication Conference (UEMCON), 2018,
pp. 977–982.
[35] Y. Liu, B. R. Tamma, B. S. Manoj, and R. Rao, “On Cognitive Network
Channel Selection and the Impact on Transport Layer Performance,” in
2010 IEEE Global Telecommunications Conference GLOBECOM 2010,
2010, pp. 1–5.
[36] Q. Cui, Z. Zhang, Y. Shi, W. Ni, M. Zeng, and M. Zhou, “Dynamic Mul-
tichannel Access Based on Deep Reinforcement Learning in Distributed
Wireless Networks,” IEEE Systems Journal, vol. 16, no. 4, pp. 5831–5834,
2022.
[37] A. K. Gizzini, M. Chafii, A. Nimr, and G. Fettweis, “Deep Learning Based
Channel Estimation Schemes for IEEE 802.11p Standard,” IEEE Access,
vol. 8, pp. 113 751–113 765, 2020.
[38] A. Kulkarni, A. Seetharam, A. Ramesh, and J. D. Herath, “DeepChannel:
Wireless Channel Quality Prediction Using Deep Learning,” IEEE Trans.
Veh. Technol., vol. 69, no. 1, pp. 443–456, 2020.
[39] W. Jiang and H. D. Schotten, “Neural Network-Based Fading Channel Pre-
diction: A Comprehensive Overview,” IEEE Access, vol. 7, pp. 118 112–
118 124, 2019.
[40] W. Jiang, H. Dieter Schotten, and J.-y. Xiang, Neural
Network–Based Wireless Channel Prediction. John Wiley &
Sons, Ltd, 2020, ch. 16, pp. 303–325. [Online]. Available:
https://onlinelibrary.wiley.com/doi/abs/10.1002/9781119562306.ch16
[41] S. Scanzio, F. Xia, G. Cena, and A. Valenzano, “Predicting
Wi-Fi link quality through artificial neural networks,” Internet
Technology Letters, vol. 5, no. 2, p. e326, 2022. [Online]. Available:
https://onlinelibrary.wiley.com/doi/abs/10.1002/itl2.326
[42] S. Scanzio, G. Cena, C. Zunino, and A. Valenzano, “Machine Learning to
Support Self-Configuration of Industrial Systems Interconnected over Wi-
Fi,” in IEEE 27th International Conference on Emerging Technologies and
Factory Automation (ETFA 2022), 2022, pp. 1–8.
[43] G. Cena, S. Scanzio, and A. Valenzano, “Improving Effectiveness of
Seamless Redundancy in Real Industrial Wi-Fi Networks,” IEEE Trans-
actions on Industrial Informatics, vol. 14, no. 5, pp. 2095–2107, 2018.
[44] M. Cereia, I. C. Bertolotti, and S. Scanzio, “Performance of a Real-
Time EtherCAT Master Under Linux,” IEEE Transactions on Industrial
Informatics, vol. 7, no. 4, pp. 679–687, 2011.
[45] F. Gosewehr, M. Wermann, and A. W. Colombo, “From RTAI to RT-
Preempt a quantative approach in replacing Linux based dual kernel real-
time operating systems with Linux RT-Preempt in distributed real-time
networks for educational ICT systems,” in IECON 2016 - 42nd Annual
Conference of the IEEE Industrial Electronics Society, 2016, pp. 6596–
6601.
[46] G. Cena, S. Scanzio, and A. Valenzano, “A software-defined MAC archi-
tecture for Wi-Fi operating in user space on conventional PCs,” in 2017
IEEE 13th International Workshop on Factory Communication Systems
(WFCS), 2017, pp. 1–10.
ALBERTO SALVATORE COLLETTO received
the B.Sc. degree in computer engineering and the
M.Sc. degree in cybersecurity oriented computer
engineering from the Politecnico di Torino, Italy,
in 2019 and 2023, respectively. His current re-
search interest is artificial intelligence, with partic-
ular reference to artificial neural networks applied
to wireless networks. In this field, he is collaborat-
ing with the National Research Council of Italy,
with the Institute of Electronics, Computer and
Telecommunication Engineering, National Research Council of Italy (CNR-
IEIIT).
STEFANO SCANZIO (S’06-M’12-SM’22) re-
ceived Laurea and Ph.D. degrees in computer
science from Politecnico di Torino, Turin, Italy,
in 2004 and 2008, respectively. From 2004 to
2009, he was with the Department of Computer
Engineering, Politecnico di Torino, where he was
involved in research on speech recognition and in
classification methods and algorithms. Since 2009,
he has been with the National Research Council
of Italy, where he is currently a Senior researcher
with the institute CNR-IEIIT. He teaches several courses on computer
science at Politecnico di Torino and Università degli Studi di Pavia. He
has authored and co-authored more than 90 papers in international journals
and conferences, in the areas of industrial communication systems, real-time
networks, wireless networks, and artificial intelligence. He took part in the
program and organizing committees of many international conferences of
primary importance in his research areas. He received the 2017 Best Paper
Award of the IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS
and the Best Paper Awards for three papers presented at the IEEE Workshops
on Factory Communication Systems in 2010, 2017, and 2019, and for a
paper presented at the IEEE International Conference on Factory Communi-
cation Systems in 2020. He is an Associate Editor of the IEEE Access, the
Ad Hoc Networks (Elsevier), and the Electronics (MDPI) journals.
GABRIELE FORMIS (S’23) received the B.Sc.
in mechanical engineering and the M.Sc. degree
in automation and control engineering from the
Politecnico di Milano, Italy, in 2018 and 2020,
respectively. He is currently pursuing the National
Ph.D. degree in Artificial Intelligence, Politecnico
di Torino, Italy. In addition, he is Research As-
sociate with the Institute of Electronics, Com-
puter and Telecommunication Engineering of the
National Research Council of Italy (CNR-IEIIT).
His research interests include artificial intelligence, wireless networks, and
autonomous driving.
GIANLUCA CENA (SM’09) received the M.S.
degree in electronic engineering and the Ph.D. de-
gree in information and system engineering from
the Politecnico di Torino, Italy, in 1991 and 1996,
respectively. Since 2005 he has been a Direc-
tor of Research with the Institute of Electronics,
Computer and Telecommunication Engineering,
National Research Council of Italy (CNR-IEIIT).
His research interests include wired and wireless
industrial communication systems, real-time pro-
tocols, and automotive networks. In these areas he has co-authored about
170 technical papers and one international patent. He received the Best Paper
Award of the IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS
in 2017 and of the IEEE Workshop on Factory Communication Systems
in 2004, 2010, 2017, 2019, and 2020. Dr. Cena served as a Program Co-
Chairman of the IEEE Workshop on Factory Communication Systems in
2006 and 2008. Since 2009 he has been an Associate Editor of the IEEE
TRANSACTIONS ON INDUSTRIAL INFORMATICS.
VOLUME 6, 2016 13
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3327523
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/