IEEE Consumer Electronics Magazine - January/February 2023 - 36

Autonomous UAV Base Stations Through Deep Learning
Figure 2. CNN architecture.
connected network for each state-action pair of
all possible UAV-BS and user deployments can be
impractically long.
One of the most successful DL architectures
is CNN,16 which provides superior performance
benchmarks never before achieved for a variety
of problems.
Proposed CNN-Based Solution Approach
A CNN is an artificial neural network with
multiple hidden layers (more precisely, multiple
convolutional layers, where the data from the
previous layer are convolved with different filters).
Indeed, CNN-based approaches have been
shown to perform well in solving a wide range of
complex problems that could not be otherwise
solved efficiently. Therefore, in this study, we
opt for a CNN-based DL approach to determine
the trajectory of an autonomous UAV-BS.
CNNs are comprised of convolutional layers,
pooling layers, and fully connected layers, and
they are generally used with a softmax function
to perform classification tasks. CNNs excel at
learning spatially localized features by the convolution
of the input data with filters in a sliding
window arrangement. Initial layers learn
generic/simple features, whereas deeper layers
learn more abstract/sophisticated features by
nonlinearly combining the features learned by
the preceding layers.
An activation function generally follows each
convolution operation, which performs a nonlinear
mapping of the convolution output. We adopt
the rectified linear unit as our activation function.
After activation, a max pooling operation is generally
performed, which is a downsampling operation
done to reduce the complexity of activation
36
maps that are further processed in the following/
deeper layers of the network. Finally, at the penultimate
layer (i.e., the last convolutional block), a
fully connected neural network is used to learn
nonlinear combinations of all the learned
features.
We utilize a four-layer CNN architecture, as
illustrated in Figure 2, where max pooling operation
(i.e., resampling) with a grid size of 22is
applied after every convolutional layer. The
number of filters at consecutive convolutional
layers are 4, 4, 8, and 16 from the first layer to
the last layer, respectively. At the end, we use
two fully connected layers with 200 neurons
each, which then connect to two output neurons
with a linear activation that yields the coordinates
for UAV-BS location. We employ Adam,17
which is a widely utilized optimization approach
for stochastic gradient descent. The network
depth as well as the max pooling and convolution
functions are determined through empirical
evaluations (e.g., we experimented with shallower/deeper
networks and empirically determined
the optimal architecture).
The network area is spatially organized into
rectangular cells (e.g., a checkerboard structure)
during data preprocessing.MUs are assigned to the
cells according to their locations. An array, which
consists of the numbers of MUs in each cell, is fed
to the model as the input. Furthermore, each cell
encompasses five consecutive time instants in the
temporal dimension (i.e., the temporal horizon).
Utilizing an input, which is embedded with both
temporal and spatial characteristics, has several
advantages. First, the scalability of themodel is not
affected by the number of MUs (i.e., the size of the
input does not change with the number of MUs).
IEEE Consumer Electronics Magazine
IEEE Consumer Electronics Magazine - January/February 2023

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - January/February 2023
