Page 88 - 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

1. Signal quality: The irst block is meant to abstract
the interactions among APs and STAs in the RF do‑
main. To achieve this, we consider two separate lay‑
ers, which process the RSSI, SINR, distance among
nodes, and SINR. A Parametric Recti ied Linear Unit
(PReLU) activation function is used together with 1‑
dimensional batch normalization.
2. AP bandwidth: The second block analyzes the avail‑
able bandwidth of the APs, and it consists of a single
linear layer, which receives as input a vector with the
Fig. 4 – GNN model proposed by ATARI.
corresponding airtime for all the available channels.
nodes’ states based on their neighbors. Note that, follow‑ This layer outputs a 3‑dimensional array and is acti‑
ing the formal de inition of a GNB in [27], global param‑ vated with a PReLU function.
eters were not used. Our implementation is available in 3. Output: Finally, the last block takes the output from
GitHub [29].
both the signal quality and AP bandwidth blocks and
computes the inal prediction value. For this inal
To train the model, we have considered splitting the pro‑
layer, we employ a simple ReLU (instead of a PReLU)
vided data set (80% for training and 20% for validation).
to avoid negative throughput predictions.
As each deployment is considered to be represented by
a graph, 480 graphs have been used for training, and
120 for validation purposes. The loss function consid‑ With this structure, we have built a much more ef icient
ered to assess the performance of our model is the Root model than if we had used a fully‑connected NN with all
Mean Squared Error (RMSE). The error obtained across the STA features. Notice that the proposed model needs
all the predictions is used to compare the accuracy of our far fewer neurons (and thus, less computational force)
model’s predictions to the actual results. However, initial to capture the most relevant information of each sce‑
results showed that the model mostly focused on predict‑ nario. In our opinion, an excess of neurons in such a
ing the throughput of the APs, given that the error is min‑ complex scenario would result in over itting, thus making
imized on large values. Therefore, we proposed a masked the model less accurate for predicting the performance of
loss provided that AP’s throughput should be equal to the new deployments.
sum of the associated STAs’ throughput. The RMSE is cal‑
As for training the MLP model, we have considered the
culated using the STAs’ predicted and the APs’ computed
throughput. following key features: the type of device (AP or STA), its
location, the minimum, and maximum channel through
which it is allowed to transmit, the RSSI, the SINR, and
3.2 Ramon Vallés
the airtime. The training was performed using 80% of the
To address the throughput prediction problem in CB‑ data set (keeping the 20% left for validation), following an
compliant WLANs, we propose a deep neural network evaluation criterion based on the RMSE of the predicted
where the information of each BSS is processed inde‑ throughput with respect to the real value. The Adam op‑
pendently, thus following the idea of Multi‑Layer Percep‑ timizer has been used to optimize the training process,
tron (MLP) [30]. More speci ically, the proposed model is which is straightforward to implement, computationally
a feed‑forward deep learning algorithm implemented in icient, and icient. For the training phase,
Python with the support of the PyTorch libraries. 4 several experiments had been made by modulating the
hyper‑parameters. The best results were achieved with
Our model aims to predict the aggregate throughput of a learning rate of 0,025 and a total of 700 epochs.
each BSS, rather than the individual throughput at STAs.
The fact is that predicting the throughput per STA is very
3.3 STC
challenging because of the dynamic channel bonding pol‑
icy used in complex scenarios, which contributes to gen‑ Our proposal includes popular ML regression algorithms
erating multiple interactions among nodes that cannot be such as MLP, Support Vector Machine (SVM), Random
captured at a glance. Accordingly, to derive an overall rep‑ Forest, and eXtreme Gradient Boosting (XGboost). These
resentation of each BSS, the features from individual STAs algorithms are backed by rich research, known to do well
are preprocessed so that we consider only their global dis‑ on regression problems, ease of implementation and de‑
tribution (mean, and standard deviation). To do so, our ployment, which are important characteristics from the
model is divided into three main blocks performing business perspective.
different tasks:
4 The code used to implement the method proposed by Ramon Vallés is
open access [31].

72 © International Telecommunication Union, 2021