IEEE Consumer Electronics Magazine - January/February 2023 - 35

Multiple MUs can demand the same content,
and for this reason, caching popular content is
beneficial in reducing bandwidth usage as well
as delay.9 However, caching a high volume of
content is not possible with a resource-constrained
UAV-BS. Therefore, the utility function
for caching, which depends on MUs to be served
and the content demand of the MUs, should be
maximized.
Energy dissipation characteristics of UAV-BS
operations is of utmost importance since the service
duration of a UAV-BS is, generally, determined
by its power usage.1 Indeed, in certain
cases the service duration of the UAV-BS is more
important than the coverage or QoS. Therefore,
the energy dissipation of a UAV-BS, as a function
of its speed and acceleration, should be taken
into account in the objective function.
Coverage Optimization
Coverage (i.e., serving as many MUs as possible),
arguably, is the foremost objective in our
problem, which can be achieved by positioning
the UAV-BS, at a given time instant, at a position
(defined by the vector dt) that results in the maximum
number of users experiencing path loss
values lower than a predetermined threshold
(g),10 which can be expressed as follows:
max
X
k2U;t2T
In the model, subscript t signifies the temporal
characteristics of the problem to account for
the dynamics of the channel and the characteristics
of MUs as well as the mobility of the UAV-BS.
The time horizon and the set of MUs are denoted
by T and U, respectively. The variable uk;t is set
to one if the MU-k is served at time instant t. The
bandwidth available to a UAV-BS is limited,
therefore, the aggregated capacity allocated to
MUs at each time instant t is upper bounded by
the total available backhaul capacity.11
Characteristics of Solution Approaches
The solution of the aforementioned optimization
model to optimality is highly challenging.
Depending on the functions utilized in the objective
and constraints, the problem can turn out to
be nonconvex. Furthermore, the temporal
January/February 2023
uk;t:
(1)
dimension of the problem can hugely inflate the
size of the solution space, which in turn can
greatly increase the computational complexity
of the solution algorithms. Solving a single time
instance of the problem, which is a commonly
employed approach, necessitates resolving the
problem for each time instant. Solutions that
consider only a single time instant can result in
suboptimal solutions (for a finite time interval-
not an instant), which can lead to the unnecessary
movements of the UAV-BS. Moreover, due
to errors in collecting or obtaining data (e.g.,
positions of MUs and path loss values), solution
approaches, which cannot tolerate such nonideal
inputs can potentially end up yielding
inconsistent or inefficient UAV-BS trajectories.
To overcome such challenges efficiently, MLbased
approaches can be employed to estimate
the optimal trajectory of a UAV-BS.
ML-Based Solution Alternatives
In RL,12 there is an agent, which can be found
in one of the predefined states. The agent takes
actions to move among different states and the
actions are rewarded. For each state, the agent
decides the action to maximize the total future
reward. RL guarantees the best actions only if
the number of iterations is large enough, therefore,
convergence can potentially require formidably
long time periods depending on the
solution space size. Moreover, if RL keeps state-
action pairs for all deployment scenarios, then it
will be impractical for a resource-constrained
UAV-BS to adopt an RL-based approach due to
limited memory.
DL5 is an ML approach that has begun to
receive much attention due to the unprecedented
increase in computing capabilities offered by
modern GPUs and advances in algorithms. Applications
of DL-based approaches for communications
and optimization problems have resulted in
improved performance and functionality compared
to existing solutions that do not utilize DL.
DRL13 is a form of RL that estimates state-action
pairs using a deep neural network (DNN) to mitigate
the memory requirements of huge state-
action tables. Hence, DNN-based approaches can
be used to alleviate the memory requirements
for RL.4,7,12,14,15 However, the training time for
the establishment and convergence of a fully
35
IEEE Consumer Electronics Magazine - January/February 2023

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