Page 89 - 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

From all the above-mentioned methods, XGBoost [32] was features, each STA of every deployment is characterized by a
selected for competing in the challenge because it of- feature vector ( , …, 21+3 ), with denoting the

fered the highest performance on both the training and number of wireless channels. Note, as well, that the
validation stages compared with the other models. XG‑ entire data set is considered for training, thus combining
boost is a gradient boosting framework available for mul‑ STAs from different deployments. The rationale is that
tiple platforms, thus providing high portability. features such as the number of neighbors in the primary
channel, the SINR, and the interference should
To train our model, we have analyzed the variance of the
differentiate STAs from different deployments.
variables in the data set, and discarded the features with
low or moderate variability. Some of the considered fea‑ When it comes to the regression problem, Gossip uses a
tures are the position of nodes, the primary channel, the feed‑forward neural network with four layers. The input
distance among nodes, or the power received by neigh‑ layers pass the input features to two fully connected lay‑
boring devices. To preprocess the selected features, we ers of neurons with a ReLU activation unit. Finally, a sin‑
have applied Yeo‑Johnson transformation [33] and nor‑ gle neuron receives the output of the hidden layers and
malization. These steps were done for predictors to im‑ generates the prediction of the throughput. As a remark,
prove their utility in the models. the last neuron has a linear activation. It is important to
remark that the proposed neural network is mostly meant
The training data set was split into training and valida‑
to tackle linear regression problems. Nevertheless, even if
tion, so that we could train the model on the whole train‑
the throughput prediction problem for WLANs is not lin‑
ing data set using a 10‑fold cross‑validation procedure.
ear, we expect our model to properly identify local min‑
Further, we used 100 and 300 combinations, respectively,
imum/maximum points that allow providing reasonable
with 10‑fold cross‑validation.
prediction results.
As for the hyper‑parameter setting (e.g., max depth or
To train the proposed neural network, we have used
minimum child weight), we tuned hyper‑parameters us‑
the RMSprop gradient descend method, considering the
ing a grid search. More speci ically, the hyper‑parameter
Mean Squared Error (MSE) as a loss function. Moreover,
values were set using Latin Hypercube Sampling, while
50 training episodes and a batch size of 50 STAs have been
their maximum and lower range value for each hyper‑
considered. Thanks to Gossip design, the training data set
parameter were mostly predetermined using default val‑
ues from tidymodels [34]. is populated with every STA of every deployment present
among all scenarios.
Finally, signi icant efforts have been put to deploy our
model. In particular, we have used docker to make our
3.5 Net Intels
model easy to (re)train and deploy. All the code and doc‑
umentation has been made publicly available [35]. To address the objective of predicting the throughput of
APs and STAs in typical dense environments, we explore
3.4 UC3M NETCOM a set of popular regression techniques. With the help of
these techniques, we aim to build complex mathematical
We formulate the throughput forecasting in WLANs as a relationships among features and labels from the data set,
linear regression problem, which assumes a linear rela‑ so that performance of WLANs can be predicted at unseen
tionship between the features and label(s) of a given data deployments. In particular, we propose using the follow‑
set { , , , ..., , }, and a set of unknown parameters ing techniques: 5
,1

,2
to be learned (being the bias). 1. Arti icial Neural Network (ANN): The ANN method
0
Our solution (named Gossip) is based on a linear regres‑ is selected ly due to its potential and versatil‑
sion method, and it aims to predict the throughput of an ity to model nonlinear and complex relationships in
STA in a given WLAN deployment where CB is ap‑ OBSS data elegantly. The proposed ANN is built using
plied [36]. Based on STAs’ individual throughput, we de‑ Tensor low and Keras libraries in Python [38]. The
rive the performance of each AP by aggregating the NN model is designed with one input layer, 7 hidden
values of their associated STAs. In particular, Gossip layers, and 1 output layer (see Fig. 5). The ReLU func‑
derives the unknown bias and weight parameters by (i) tion is employed to activate hidden layers. In each of
processing the WLAN data set, and (ii) applying a neural the irst six hidden layers, there are 1024 nodes. For
network to perform regression. the seventh hidden layer, there are 512 nodes. The
model is trained using an Adam optimizer. The batch
As for the processing part, Gossip takes the input fea‑
size and number of epochs for training, after multiple
tures generated by the Komondor simulator, and se‑
trials, were set to 250 and 1000 respectively.
lects/generates the most relevant ones: the position of
the STA, the AP to which the STA is associated, the RSSI, 5 The code used to implement all the proposed methods by Net Intels is
the SINR, the set of nodes using the same primary channel, available in Github [37].
and the set of allowed channels. After processing the

© International Telecommunication Union, 2021 73