To Transmit or Not to Transmit: Controlling

Communications in the Mobile IoT Domain

K. PANAGIDI, Department of Informatics & Telecommunications, University of Athens, Greece

C. ANAGNOSTOPOULOS, School of Computing Science, University of Glasgow, Scotland
A. CHALVATZARAS and S. HADJIEFTHYMIADES, Department of Informatics &
Telecommunications, University of Athens, Greece

The Mobile IoT domain has been significantly expanded with the proliferation of drones and unmanned
robotic devices. In this new landscape, the communication between the resource-constrained device and
the fixed infrastructure is similarly expanded to include new messages of varying importance, control, and 22
monitoring. To efficiently and effectively control the exchange of such messages subject to the stochastic
nature of the underlying wireless network, we design a time-optimized, dynamic, and distributed decision-
making mechanism based on the principles of the Optimal Stopping and Change Detection theories. The
findings from our experimentation platform are promising and solidly supportive to a vast spectrum of real-
time and latency-sensitive applications with quality-of-service requirements in mobile IoT environments.
CCS Concepts: • Computer systems organization → Embedded systems; Redundancy; Robotics; • Net-
works → Network reliability;
Additional Key Words and Phrases: Real-time decision-making, mobile IoT, optimal stopping theory, change-
point detection, unmanned vehicles
ACM Reference format:
K. Panagidi, C. Anagnostopoulos, A. Chalvatzaras, and S. Hadjiefthymiades. 2020. To Transmit or Not to
Transmit: Controlling Communications in the Mobile IoT Domain. ACM Trans. Internet Technol. 20, 3, Article
22 (August 2020), 23 pages.
[Link]

1 INTRODUCTION
In the last decade, we have been witnessing significant advancements and evolution of the Inter-
net of Things (IoT). Going a step further to the IoT infrastructure, resource-constrained nodes are
enhanced with mobility capabilities forming the Mobile IoT (MIoT) networks; noticeably, huge
growth has been witnessed in the Unmanned Vehicles research area. We can consider a drone as a

This work has received funding from the European Union’s Horizon 2020 Framework Programme for Research and Inno-
vation under the Grant Agreement No 645220, project RAWFIE (Road-, Air-, and Water-based Future Internet Experimen-
tation).
Authors’ addresses: K. Panagidi, A. Chalvatzaras, and S. Hadjiefthymiades, Department of Informatics & Telecommu-
nications, University of Athens, Panepistimioupolis, Ilissia, Athens, 15784; emails: {kakiap, achalv, shadj}@[Link];
C. Anagnostopoulos, School of Computing Science, University of Glasgow, Clasgow, G12 8QQ; email: christos.
anagnostopoulos@[Link].
Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee
provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and
the full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored.
Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires
prior specific permission and/or a fee. Request permissions from permissions@[Link].
© 2020 Association for Computing Machinery.
1533-5399/2020/08-ART22 $15.00
[Link]

ACM Transactions on Internet Technology, Vol. 20, No. 3, Article 22. Publication date: August 2020.
22:2 K. Panagidi et al.

mobile computing and sensing node deployed to different locations tailored to specific tasks. The
fundamental features that “transform” Unmanned Vehicles to popular mobile IoT nodes are the
ability to autonomously make decisions (i.e., without human intervention), the capability of carry-
ing additional application-specific payloads, the endurance, the capability of re-programmability,
and the capacity to stream locally sensed/captured multimedia content. As Unmanned Vehicles be-
come more advanced in terms of computational capabilities, they are expected to present greater
value in application cases of, e.g., environmental surveillance and monitoring, and supporting cri-
sis management activities. For instance, consider the use case where drones equipped with video
camera and various sensors, like air-quality, humidity, and temperature, are programmed to cruise
over forests and spot fires at an early stage.
The ultimate target of an Unmanned Vehicle, also coined as UxV, where “x” can stand for either
“A” aerial, “G” ground, or “S” surface vehicle, is the successful execution of a pre-programmed mis-
sion. A mission is often described as a trajectory with specific waypoints that the UxV is tasked
to approach and collect various measurements, e.g., from on-board sensors, or capture images or
video, e.g., from on-board cameras. The waypoints along with the various commands are deter-
mined from a control unit, i.e., a Ground Control Station (GCS). A GCS is a remote coordinator
(master) node responsible for contextual data acquisition and real-time control and monitoring of
the progress of the UxVs missions. The communication between UxV and GCS is realized in a wire-
less manner. The UxVs themselves can be either involved in a mission as single/individual units or
as groups, i.e., swarm of UxVs. A swarm of UxVs forms a remote sensing system and can be treated
as Mobile Wireless Sensor Network (MWSN) of highly dynamic topology. More importantly, the
on-board computing and sensing elements of the UxVs enhance the in-network embedded intelli-
gence of the swarm. This allows complex local computational and analytics tasks to be realized in a
highly distributed fashion, thus, balancing computational load across the infrastructure and render
communications much more energy efficient. In this MIoT environment of UxV-driven distributed
computing, we are facing the following research and technical challenges:
Challenge 1: Real-time Monitoring. Real-time surveillance and monitoring applications, e.g.,
detection of forest fires, require control messages to be delivered from a swarm of UxVs to the GCS
with the minimal delay and high accuracy. These missions typically involve rural areas, where
the network connectivity is expected to be poor [12]. Moreover, radio paths between the UxVs
and GCS are anticipated to be obstructed, overloaded, or to suffer from high packet loss rate. It is
challenging to predict these network variations in these environments. Hence, it is deemed crucial,
during a mission, an UxV to autonomously decide when to pause telemetry/control measurements
that are not currently prioritized as ‘important’ and save network resources.
Challenge 2: Secure UxV Control and Actuation. The connectivity among UxVs and GCS
needs to take into consideration the mobility factor. This factor adds up a new degree of freedom to
their operation, since the GCS sends control commands to UxVs while UxVs are moving for further
local actuation. The control messages and their acknowledgments must be securely delivered in
order to guarantee safe and successful missions. The usual approach to emergency cases, when
a UxV loses the connection to GCS, is that the UxV returns to its initial position abandoning the
mission. This means that the mission is cancelled, even if the UxV could be really close to the
mission’s end or objective leading to significant waste of time and resources.
In this work, we cope with the above-mentioned challenges by proposing an online stochastic-
driven decision-making scheme that leverages the transmission functionality of UxVs and GCS by
being adaptive to changes in network quality. This is designed and developed by our novel sup-
pression control of telemetry and control messages model based on the principles of the Optimal
Stopping Theory (OST). Our time-optimized control mechanism achieves the optimal delivery of
critical information from UxVs to GCS and vice-versa. Our rationale is that should the network

ACM Transactions on Internet Technology, Vol. 20, No. 3, Article 22. Publication date: August 2020.
To Transmit or Not to Transmit: Controlling Communications in the Mobile IoT Domain 22:3

be performing properly, then the transmission control can be “relaxed” to exploit the available re-
sources in the resource-constrained UxV. Our model introduces two sequential optimal stopping
time decision-making mechanisms based on the Change Detection theory and an application-
specific discounted reward process.
We consider the case where a UxV operator desires to execute a mission and consider the setting
where two main components are provided: a GCS and an Unmanned Ground Vehicle (UGV). The
mission instructions could be consolidated in a domain-specific script, e.g., the mission scripts
compiled through our experimentation platform for UxVs Road-, Air-, and Water-based Future
Internet Experimentation (RAWFIE) [17]. The RAWFIE1 platform is briefly presented in Section 4.3.
The mission script defined by the operator includes (among others) the UxV trajectory waypoints
in the field area to control the device in space and time and the sensing components involved
(sensors) to collect in-field measurements. The main goal of the two components is the monitoring
of an area to detect fire based on camera stream and on-board environmental sensors. This use case
was also conducted during the RAWFIE project lifetime.
The baseline solution/establishment for the UGV’s mission is as follows: The GCS sends specific
commands (directives) to the UGV as indicated in an experimentation script, e.g., “Go-to-Point,”
“Pause” on a specific point, or “Abort” the mission and return home (RTL). The UGV sends sensor
measurements streams, e.g., temperature, humidity, video, and its geo-spatial position (GPS) to
GCS with a predefined frequency. Both GCS and UGV have as a goal the successful completion of
the monitoring of the area. Both UGV and GCS monitor the quality of the network. The quality of
the network can be classified as proposed in Reference [14] and is crucial for the mission because
significant commands (down-link from GCS to UGV) or sensor values/measurements (up-link from
UGV to GCS) can be occasionally lost due to the stochastic network behavior.
We propose a real-time control mechanism to adapt to changes in network quality by dynami-
cally pausing control telemetry and control messages based on optimal sequential decision -making
rules. This is expected to ensure the trouble-free delivery of critical information subject to the
dynamic network status that UxVs encounter while dispatching a certain mission.
Remark 1. Overall, our scheme can be applied in all cases where connections are competing for
stochastically varying network resource and optimally manage their relative priorities.
This article is organized as follows: In Section 2, we present the related work, while in Section 3,
we present the preliminaries for our problem formulation, the proposed optimized information
flow model, and our two optimal stopping problem solutions. Section 4 presents our comprehen-
sive experiments with real UxV settings, where our mechanisms’ performances are followed by
the conclusions in Section 5. A list of abbreviations is presented in Table 1.

2 RELATED WORK & CONTRIBUTION

2.1 Related Work
The challenge of optimizing contextual information flow delivery among UxVs is non-trivial given
the network circumstances and status. To our knowledge, there is no prior holistic work addressing
the problem of time-optimized information flow. In the literature, research has been extensively
focused on message-routing protocol employed on UxVs. Opportunistic networks have been pro-
posed as long as they are capable of maintaining efficient operation in a wide range of network den-
sity and mobility conditions [19, 26]. By classifying the diversity of topological conditions in net-
working environments, one end of the spectrum corresponds to almost static dense topologies. In
this case, conventional topology-based protocols [20] function best by using node labels/identities.

1 [Link]

ACM Transactions on Internet Technology, Vol. 20, No. 3, Article 22. Publication date: August 2020.
22:4 K. Panagidi et al.

Table 1. Nomenclature

Notation Description
DMP Decision-Making Process
TOCP Time-Optimized Change Point DMP
DRP Optimal Discounted Reward DMP
QNI Quality Network Indicator
p(x n , f ) Probability Density Function with parametric density f
fi Normal distribution N (μ i , σi )
H0 No-Change-point Hypothesis
H1 Change-point Hypothesis
FAR False Alarm Rate
Nd Detection Change time
α Detection threshold: 0 ≤ α ≤ FAR
γ ∈ [0, 1] Discounted factor at DRP policy
r (γ , N ) ∈ {1, . . . , N } Stopping time in DRP policy up to time N
t ∗, τ ∗, r ∗ Optimal stopping times in generic OST, Change-point Detection,
and DRP policy, respectively
Th Maximum horizon where a UxV can be paused
L x (n) Log-likelihood ratio at time n for random variable X

As the nodal density decreases and/or the mobility increases, and up to a point where the connec-
tivity status between pairs of nodes remains stable, position-based families of protocols [11, 19]
become more suitable.
Additionally, in networks of low nodal density, intense mobility becomes a prerequisite for the
creation of contact opportunities. For such topologies, protocols based on the “carry” action [15,
26], i.e., the spatial transposition of the message due to the physical movement of the carrier node,
perform efficiently. The aforementioned routing protocols have been designed to accommodate a
restricted set of possible network conditions, corresponding to a particular sub-range, and yield
satisfactory performance only under these conditions.
Opportunistic Networking is also an open and an active field of research where OST can be ap-
plied at routing delivery protocols. A proposal for opportunistic networks (OppNet) [7] is studied
in which the authors present Softwarecast as a general delivery scheme for group communica-
tions based on mobile code. This software code and a delivery state is the main input to persist
refined delivery-decision-making methods based on OST to implement complex decisions. In Ref-
erence [10], the authors present the Relcast, a composite routing-delivery scheme that used OST-
based delivery strategies to route messages to profiles, which are defined by delivery functions
such as best maximum and over-the-average. If we go a step further, we define a routing deliv-
ery protocol to social OppNet like influencers’ networks. The authors in Reference [9] refer to
an OST-based solution to deliver messages in highly connected networks. However, the proposed
solutions are based on metrics like low latency, while the authors in Reference [8] proposed a so-
lution of broadcast protocols for OppNet based on efficiency, preventing unrestrained propagation
of messages.
All the proposed delivery routing protocols are based on variations of the Secretary Problem
[13] like the called rank-based selection and cardinal payoffs variation of the secretary problem [6].
However, a unique strategy cannot be applied to sequences with abrupt changes where each state
shall be treated differently. Other research efforts are focused on delay-tolerant methodologies,
where mobile sinks (e.g., data aggregation nodes) “patrol” a number of static sensor nodes and

ACM Transactions on Internet Technology, Vol. 20, No. 3, Article 22. Publication date: August 2020.
To Transmit or Not to Transmit: Controlling Communications in the Mobile IoT Domain 22:5

collect data [18, 29]. Nonetheless, due to their delay-tolerant principle for data delivery, they cannot
be directly applied to real-time applications like disaster management.
Methods based on the principles of dynamic stochastic optimization frameworks, like Optimal
Stopping Theory, have been successfully applied to information dissemination in ad hoc networks.
The authors in Reference [12] add mobility into wireless network infrastructure, i.e., WiFi access
points (AP) on wheels, which move to optimize user performance. The Roomba devices equipped
with network interfaces move independently around areas in order to maximize the wireless
capacity in this area. However, the mobile devices are moving based on a grid at the floor to
predefined paths. In Reference [25], researchers apply Optimal Stopping Theory based on Change
Detection only in in-network statistics. This method is only applied to pause the generation of
telemetry messages. Pausing period stops when a time threshold is reached and for this period
framework is agnostic to network state. Researchers’ method is compared with our proposed
model in Section 4.3 applied on real UxVs. Contextual data delivery mechanisms have been
studied in the literature though from a different “perspective” in mobile ad hoc networks. The
contextual data delivery mechanisms in [4, 2, 24], and [3] deal with the delivery of quality infor-
mation to context-aware applications in static and mobile ad hoc networks, respectively, assuming
epidemic-based information dissemination schemes. In Reference [4], the authors propose optimal
decision-making approaches on the collection of contextual data from WSNs. The mechanism
in Reference [2] is based on the probabilistic extension of the well-known Secretary Problem
introduced in Reference [13] merged with an optimal online stochastic optimization problem.
The authors in Reference [1] tackle the task offloading decision-making problem by adopting the
principles of OST to minimize the execution delay in a sequential decision manner. Their approach
significantly minimizes the execution delay for task execution, and the results are closer to the
optimal solution than other deterministic offloading methods. The authors in Reference [24] study
a dynamic video encoder that detects scene changes and tunes the synthesis of Groups-of-Pictures
(GOP) accordingly based on a “Black-Jack” like application of OST. The proposed MPEG encoder
tracks the error between the sequential frames in a Group-of-Pictures (GOPs) and optimally
creates GOP sizes that are content-based with the minimum waste of the resources.

2.2 Contribution
Our problem deals with poor network performance during a UxV predefined mission. The online
control of UxVs mission is highly connected with two types of paths: geo-spatial and network. The
union of localization and network factors concludes to a safe mission with accurate data reports.
It is apparent that the mobility factor adds up new complexity to the aforementioned solutions in
literature that handle message forwarding or routing topologies for stationary sensor networks.
Furthermore, our framework is independent of the UxVs technologies and can be applied to dif-
ferent kinds of UxVs (aerial, sea, ground) and to their on-board software like ROS [28] or Ardupilot
[5]. Mostly in literature, the UxV solutions are targeted to problems with a specific type of UxVs.
However, our work in this article does not depend on the type of UxV. Our decision-making pro-
cess handles the control of contextual flow in a mission based on the quality network statistics
with no-prior knowledge of the environment and the category of the device, i.e., aerial, ground,
or surface vehicles. This real-time decision-making framework is based on two Optimal Stopping
Time Policies that optimally schedule context delivery (control messages and values) and deliver
messages with minimum loss of packets in poor or saturated networks.
Our specific technical contribution of this work is:
(1) A stochastic optimization mechanism for online network quality change detection;
(2) A hybrid sequential decision -making mechanism for optimal control commands from the
GCS to UxV based on the Optimal Stopping Theory;
ACM Transactions on Internet Technology, Vol. 20, No. 3, Article 22. Publication date: August 2020.
22:6 K. Panagidi et al.

Fig. 1. UxV state transition model.

Table 2. Rules of State Transitions

Component Network State “Good” Network State “Bad”

UxV—High-priority sensors ON ON
GCS—High-priority messages ON OFF
UxV—Low-priority sensors ON OFF

(3) Proof of optimality of the two proposed mechanisms in UxV MIoT environments;
(4) Comprehensive performance evaluation, sensitivity analysis of the major parameters, and
comparative assessment of the proposed mechanisms in a real-testbed UxVs platform.

3 TIME-OPTIMIZED DECISION-MAKING MODEL FOR UNMANNED VEHICLES

3.1 Rationale
The main contribution of this article is to establish an in-network/on-device lightweight sequential
Decision-Making Process (DMP) that leverages the online derived network statistics to efficiently
control the progress of a UxV mission. Each UxV is equipped with a number of sensors and at least
one network interface. Our DMP is capturing network related information, e.g., packet error rate,
and controls the transmission of messages on both UxV to GCS and GCS to UxV, dynamically.
Fundamentally, based on the real-time captured network statistics, our DMP makes transitions
during the UxV mission between two states: active and passive state as shown in Figure 1. The
time duration for staying in each state and the transition from one state to another are optimally
determined by two real-time decision-making mechanisms as will be discussed in the following
paragraphs.
All messages exchanged between UxV and GCS are categorized in “high” and “low” priority. A
high-priority message is considered (i) the minimum necessary systemic instructions to carry out a
mission and (ii) sensor data defined by the UxV operator as highly important. When the UxV/GCS
is in active state, then the DMP sends messages constantly for telemetry and control. In the pas-
sive state, the DMP sends only high-priority messages. For instance, the position reporting from
the UxV is a prerequisite for the safe execution of the mission. In this case, high-priority com-
mands are being sent constantly. Low-priority messages, e.g., temperature values captured locally
from the UxV sensors, can be delayed until the network exhibits better performance. The message
priority at the GCS is the inverse, i.e., significant messages are to be delayed in order to safely
reach the UxV. The described rules of state transitions based on the network state, the UxV and
GCS, are shown in Table 2.
The DMP runs locally on the UxV and on the GCS, enriched with a Time-Optimized Change-
Point Decision-Making Process (TOCP). The TOCP is triggered when a change on network

ACM Transactions on Internet Technology, Vol. 20, No. 3, Article 22. Publication date: August 2020.
To Transmit or Not to Transmit: Controlling Communications in the Mobile IoT Domain 22:7

performance occurs; the TOCP is discussed extensively in Section 3.3. This will enable the UxV
and the GCS to transit from the active state to the passive state. When the DMP concludes on the
“passive” state, then a Discounted Reward Decision-Making Process (DRP) is activated, as will be
discussed in Section 3.4. The rationale is that the DRP sequentially ranks the network quality mea-
surements from the relatively worst to the relatively best and then, optimally, it delays its pause
interval for the (stochastically) globally best network observation to resume from the pausing pe-
riod as dictated by the TOCP. The pausing period has a maximum deadline, hereinafter referred to
as the pausing horizon Thmax . This indicates the maximum time interval the UxV waits without
receiving any command and acknowledgment (ACK) messages from the GCS. To sum up, we pro-
pose a mechanism for temporal control of the transmission of the messages to and from the UxV.
This mechanism is based on a network condition model that transits from good to bad and vice
versa. All these transitions are monitored and validated through our system using the principles
of the change detection and optimal stopping theory.

3.2 Preliminaries in Optimal Stopping Theory and Change Detection Theory

Before elaborating on our problem formulation and the proposed time-optimized mechanisms, we
provide the fundamentals and principles adopted from the OST and the change-detection theory.
3.2.1 Optimal Stopping Theory. The first studied optimal stopping problem is related with the
problem of choosing a time to take a given action based on a sequentially observed random vari-
ables in order to maximize an expected payoff. In addition, our stopping time problem has a finite
horizon, i.e., there is an upper bound on the number of stages at which we may stop.
Let Fn be defined as the σ -algebra generated by the random variables Y1 , Y2 , . . . , Yn in a proba-
bility space (Ω, F, P). We envisage Fn as the filtration (information) observed up to (discrete) time
instance n by collecting the realization values of the random variables up to n. For instance, in
our context Y1 , Y2 , . . . , Yn are considered the observed Quality of Network Indicator (QNI) values
in discrete timesteps t = 1, . . . n. A stopping rule or stopping time is defined as the random vari-
able τ with realization values in a set of natural numbers such that {τ = n ∈ Fn } for n = 1, 2, . . .
and probability P (τ < ∞) = 1. We denote with M(n, N ) the class of all stopping rules τ in which
P (n ≤ τ ≤ N ) = 1 for any n = 1, 2, . . . and N > 0. The real-valued payoff function is then defined
as the mapping W : R → R being a Borel measurable function, while values W (y) interpret the
payoff of a decision maker when it stops the Markov chain (Yn , Fn ) at the state y ∈ R. In our case,
the reward can be defined as the selection of the best network metric (QNI value) reached so far.
Assume now that for a given state y and for a given stopping rule τ , the expectation of the
reward (payoff) function exists as E[W (Yτ )|Y1 = y]. Then, the expected payoff E[W (Yτ )|Y1 = y]
corresponding to a chosen stopping rule τ exists for all states y ∈ R, which refers to the reward
value of the stopping problem. Based on the principles of optimality, the reward value VN (y) is the
supremum of the expected payoff of all the stopping rules belonging to M(1, N ), i.e.,
VN (y) = sup E[W (Yτ )|Y1 = y], (1)
τ ∈M(1, N )

where the supremum is taken for all stopping rules τ ∈ M(1, N ) for which the expectation
E[W (Yτ )|Y1 = y] exists for all y ∈ R. Based on the optimal value VN (y), where the supremum
in Equation (1) is attained, the optimal stopping rule t ∗ ∈ M(1, N ) should satisfy the condition:
VN (y) = E[W (Yt∗ )|Y1 = y], ∀y ∈ R. (2)
It is then clear that the optimal value VN (y) is the maximum possible accepted reward to be
obtained observing the random variables Y1 , . . . , YN up to the N -th observation.

ACM Transactions on Internet Technology, Vol. 20, No. 3, Article 22. Publication date: August 2020.
22:8 K. Panagidi et al.

Consider also that the expectations E[W (Yτ )|Y1 = y] exist for all y ∈ R and are based on the
principles of optimality. Let us then introduce the operator Q over the reward function W ∈ R
such that:
QW (y) = max{W (y), E[W (Yt∗ )|Y1 = y]}. (3)
Then, the optimal stopping rule t ∗ , which attains the optimal value in Equation (2), is estimated
by the Theorem 3.1:
Theorem 3.1. Assume that W ∈ R. Then:
—Vn (y) = Q nW (y), n = 1.2, . . .;
—Vn (y) = max{W (y), E[Vn−1 (Y1 )]}, where V0 (y) = W (y);
— the optimal stopping time tn∗ is evaluated as:
tn∗ = min{0 ≤ k ≤ n : Vn−k (y) = W (y)} (4)
This refers to an optimal stopping rule in M(1, n). If E[|W (Yk )|] < ∞, for k = 1, . . . , n, then
the stopping rule tn∗ in Equation (4) is optimal in the class M(1, n).
Proof. Please refer to Reference [13]. 
3.2.2 Change Point Detection Theory. The second category of the optimal stopping problem is
the detection of a change point. Consider that we are monitoring a sequence of random variables,
like values of the QNI, {Y1 , Y2 , . . . Yn } with a known distribution f 0 . At some point m in time,
unknown to us, the distribution changes to another known distribution f 1 . Our goal is to detect
the change as soon as it occurs. Let Fn , n ≥ 1 be the σ -algebra generated by the random variables
{Y1 , Y2 , . . . Yn }. A sequential change point detection rule is then derived by the stopping time τ
of the observed values. The stopping time τ for the change point detection has the following
characteristics:
— Average Run Length (ARL): ARL, proposed in Reference [22], is defined as the expected
number of observed values before a change decision is taken, where Nd is the detection
time and f is assumed to be constant, i.e., ARL = E[Nd ].
— The Detection Delay D n is the average detection delay corresponding to the observed
{Y1 , Y2 , . . . Yn } needed before a detection change occurs. Therefore, this quantity has to be
as small as possible to minimize the reaction time of the algorithm.
— The False Alarm Rate (FAR) [16] is calculated as the ratio between the number of negative
events wrongly categorized as changes.
In the following, we describe the two in-network/on-device optimal stopping rule mechanisms
running on the UxV; the same mechanisms also run on the GCS.

3.3 Time-Optimized Change-Point Decision-Making Process

3.3.1 Problem Formulation. In this section, we introduce the TOCP, which reflects the be-
havior of the UxV being in the active state. Specifically, consider the network quality readings
x 1 , x 2 , . . . , x n as a discrete random signal with independent and identically distributed (i.i.d.) ran-
dom variables observed sequentially in real time. Consider also that the network readings follow a
probability density function p(x n , fi ). In our case, fi expresses the normal distribution with mean
value μ i and variance σi . To estimate p(x n , fi ), a probability density function comparison method
has been adopted to derive the closest distribution to our QNI values.
The QNI derives from the normalization of the basic network metrics: Packet Error Rate (PER),
Signal-to-Noise Ratio (SNR), and the interference quality indicator (Q). The SNR is defined as the

ACM Transactions on Internet Technology, Vol. 20, No. 3, Article 22. Publication date: August 2020.
To Transmit or Not to Transmit: Controlling Communications in the Mobile IoT Domain 22:9

Fig. 2. (Upper) (a) The probability density function f 0 and (lower) (b) the f 1 model fitting for good and bad
quality of QNI values, respectively.

ratio of signal power to the noise power. The PER is calculated as the rate between the lost packets
and the total packets sent through the network. The interference quality indicator Q is exported
by an access point in the scale [0, 100] and depends on the level of contention or interference,
like the bit or frame error rate, or other hardware metric. The holistic QNI at time n indicates
the quality of the current network connectivity defined as the weighted sum of the (normalized)
quality indicators:
QN In = ˆ n + a 2SNˆ R n + a 3Q̂ n ,
a 1 PER (5)
3
where the QNI is the affine combination of PER, SNQ, and Q in [0, 100] such that i=1 ai = 1,
ai ∈ [0, 1], ∀i.
We consider the incoming QNI values as an adapted strong Markov process (X n )n←0 defined by
the filtered probability space p(x n , f 0 ). The estimation of the p(x n , fi ) is based on model fitting of
all the parametric probability distributions to the QNI. The output of this model fitting is shown
in Figure 2(a) for p(x n , f 0 ) and Figure 2(b) for p(x n , f 1 ). The list of examined probability distri-
butions is extensive. We based our decisions and reasoning on the fundamental Negative of the
Log Likelihood (NLogL) and the Bayesian Information Criterion (BIC) metrics. For each distribu-
tion examined, we derived the corresponding NLogL and BIC values provided in Table 3. As it is
shown in Figure 2(a) and (b), the best distribution fitting to our experimental data is the Normal
Distribution.
We further studied an abrupt change from good to bad network conditions. In this case, we
performed experiments in which the network conditions changed at time m. As shown in Figure 3,
before time m, the QNI follows the distribution p(x n , f 0 ), and after time m, the QNI follows p(x n , f 1 ).

ACM Transactions on Internet Technology, Vol. 20, No. 3, Article 22. Publication date: August 2020.
22:10 K. Panagidi et al.

Table 3. NLogL and BIC Metrics for the Probability Distributions

Examined Distribution NLogL BIC

Normal (N ) 1,876.5 3,765.5
Gamma (Γ) 1,904.2 3,820.8
Log-logistic 1,909.3 3,831
Inverse Gaussian (IG) 1,919.4 3,851.2
Rayleigh 2,366.6 4,739.5
Exponential (Exp) 2,700.7 54,076

Fig. 3. The behavior of the QNI and the cumulative log-likelihood ratio corresponding to a change from a
“good” network state to a “bad” network state.

Under these experimental observations, the QNI distribution observed between the first sample x 0
and the current x k sample takes two forms, where H 0 represents No-Change-Point Hypothesis and
H 1 represents the Change-Point Hypothesis:

⎧ k
⎪
⎨ n=0 p(x n , f 0 ), No-Change-Point Hypothesis H 0 ;
p(x ) = ⎪ m−1 k (6)
⎩ n=0 p(x n , f 0 ) n=m p(x n , f 1 ), Change-Point Hypothesis H 1
The challenge is to decide between the two hypotheses H 0 , H 1 w.r.t. QNI, and to approximate
efficiently and timely the potential change point time m. A feasible solution derived by the change-
point detection theory adopts the minmax approach in Reference [23].
Let us define the conditional expected detection delay by

EH 1 [(Nd − m + 1) + |n = 0, 1 . . . , m − 1], (7)

as defined in Reference [23], where the expectation is taken under one change hypothesis H 1 . The
minimax performance criterion is given by its supremum taken over. Specifically, the worst-case
detection delay is estimated as:

D n (τ ) = sup ess sup Ek [(τ − k + 1) + |Fk−1 ], (8)

n ≥1

with x + = max {x, 0}. Based on this objective, we formulate the change-point detection Problem 1:

Problem 1. The UxV should determine an optimal change-point detection time τ that minimizes
the worst-case detection delay in Equation (8).

ACM Transactions on Internet Technology, Vol. 20, No. 3, Article 22. Publication date: August 2020.
To Transmit or Not to Transmit: Controlling Communications in the Mobile IoT Domain 22:11

3.3.2 Solution for TOCP. Let us first denote the FAR defined as Reference [16]:
1
FAR(τ ) = .
E∞ [τ ]
Based on our examined distribution fitting, we introduce the instantaneous log-likelihood ratio at
time n by:
p(x (n), f 0 ) σ 2 (x − μ 1 ) 2 (x − μ 0 ) 2
L x (n) = ln = ln 12 + − , (9)
p(x (n), f 1 ) σ0 2σ12 2σ02
and its cumulative summation of the ratios from 0 to n:

n
S (n) = L x (k ). (10)
k=0
The expectation E∞ [τ ] defines the expected time between false alarms. A false alarm in our
case is defined when the DMP mechanism detects a change for state transition to passive, while
the network quality is characterized as good. Under the Lorden criterion, our objective is to find
the stopping rule that minimizes the worst-case delay subject to an upper bound on the FAR. The
decision function in our problem in a change between good and severe network conditions is
shown in Figure 3(b).
The optimal solution to Equation (8) was determined in Reference [21], which is provided by
the Cumulative Sum (CUSUM) test [22]. A presentation of the CUSUM approach applied to our
problem can be found in Appendix C and its description is shown in Algorithm 1. The optimal
stopping time for detecting the change point is given by:
n
τ ∗ = min{n ≥ 1, max L x (i) ≥ α } (11)
1≤k ≤n
i=k
Let the detection threshold α be chosen such that the ARL to false alarm derives FAR ≥ α > 0.
Clearly, this condition is equivalent to limit the rate of false detection by a given maximum value.
When α → ∞, the CUSUM algorithm minimizes the worst-case detection delay EH 1 [Nd ]. The
value of this delay can be approximated by using Kullback-Leibler (KL) divergence. The KL cap-
tures the discrimination between the post and pre-change hypotheses and measures the detectabil-
ity of the change, which is proved to be:
ln α
EH 1 [Nd ] = . (12)
σ02 +(μ 0 −μ 1 ) 2
ln( σσ01 ) + 2σ12
− 1
2

See Appendix B for the derivation of the expectation EH 1 [Nd ].

3.4 Discounted Reward Decision-Making Process

3.4.1 Problem Formulation. We propose a hybrid solution based on the change-point detection
and a DRP with Linear Discount Function (LDF). The reason is that the UxV cannot pause forever
to send commands or to send telemetry messages. The UxV has a hard limit for sending message
to GCS in order to report that it is alive and active. The same stands for GCS, i.e., the GCS cannot
leave a UxV with no control messages. Therefore, the pausing period when a UxV decides whether
to start again the streaming of commands can be treated as a finite horizon problem as will be
described here.
It is assumed that when the pausing period starts, the UxV receives a QNI value x k at a time
instance k. The objective is to seek a stopping rule that will maximize the probability of choosing
the best (maximum) QNI value x k indicating the best possible network condition.

ACM Transactions on Internet Technology, Vol. 20, No. 3, Article 22. Publication date: August 2020.
22:12 K. Panagidi et al.

Let us define a random variable uk , which represents the LDF reward if the kth QNI observation
is chosen, that is:
 γ
1 − N k if x k = max{x l , l = 1, . . . , k − 1},
uk = (13)
0 otherwise.
The parameter γ ∈ [0, 1] denotes the discount factor. The discount factor γ represents the mod-
eling abstraction where UxV focuses on selecting the best QNI value of the N received QNI values.
The LDF in Equation (13) indicates that the UxV has to report at least one QNI value observing at
most N QNI values. The higher the discount factor γ is, the higher the penalty gets until a recep-
tion of a better QNI value. The UxV receives the reward uk if the kth observation is chosen and
refers to the highest QNI value among all N QNI values; otherwise, uk is zero.
Problem 2. Given a fixed time horizon N , the UxV has to determine an optimal stopping rule
r , 1 ≤ r ≤ N , which maximizes the expectation E[ur ].
3.4.2 Solution for DRP. For solving Problem 2, consider first receiving the kth observation of
the QNI value x k . We can then define the random variable zk = j (1 ≤ j ≤ k ), which denotes the
relative ranking of the QNI value x k among the first k observations of the UxV. The assignment
zk = 1 means that the kth QNI value refers to the highest QNI value among the first k QNI values
seen. We then state the optimal policy for a UxV w.r.t. LDF in Equation (13) as follows in our
optimal policy.
Remark 2. Optimal Policy: There exists a time r ∗ (1 ≤ r ∗ ≤ N ) such that the UxV observes the
QNI values of the first r ∗ − 1 QNI values without accepting any of them. Then for r ∗ ≤ k ≤ N
the UxV accepts x k if zk = 1. In case of zk > 1, ∀r ∗ ≤ k < N , or r ∗ = N , then the UxV accepts x N ,
which is the last observed QNI value, with u N = 1 − γ .
Let ωk (j) be the conditional expected reward of the kth observation given that zk = j, that is,
ωk (j) = E[uk |zk = j]. The probability of finding the maximum QNI value x k , i.e., (j = 1), at the
kth observation is:
k
P (uk = 1, zk = j) if j = 1,
P (uk = 1|zk = j) = = N
P (zk = j) 0 otherwise.
Hence, we have for the ωk (j) that:
 γ
k
(1 − N k) if j = 1,
ωk (j) = N (14)
0 otherwise.
The value ωk (j) = 0 for j  1 indicates that there is no reward if the best quality network state is
not chosen.
For each r = 1, . . . , N , let ξ (r ) denote a stopping rule; that is, the first r − 1 QNI values are
observed and the next QNI value, which exceeds all of its predecessors, is accepted. If none of the
first N − 1 QNI values is reported, then the last one is reported. Then, we obtain that:
r −1
P (ξ (r ) = k ) = ; (15)
k (k − 1)
thus, the corresponding expected payoff ϕ (r ; γ , N ) w.r.t. to the reward function in Equation (13) is

N
r − 1  1 − N k 
N γ
ϕ (r ; γ , N ) = E[u ξ (r ) ] = ωk (1)P (ξ (r ) = k ) = (16)
k=r
N  k −1
k=r

ACM Transactions on Internet Technology, Vol. 20, No. 3, Article 22. Publication date: August 2020.
To Transmit or Not to Transmit: Controlling Communications in the Mobile IoT Domain 22:13

Fig. 4. Analysis of LDF based on different values of discount factor γ .

It follows that the r ∗ of the proposed optimal policy that maximizes the expected payoff
ϕ (r ; γ , N ) in Equation (16) is the optimal stopping rule. The ϕ (r ∗ ; γ , N ) is the maximum proba-
bility of finding the best QNI value on the UxV.
Theorem 3.2. There exists a r ∗ (1 ≤ r ∗ ≤ N ), which maximizes ϕ (r ; γ , N ) over 1, 2, . . . , N . Then,
the optimal stopping rule r ∗ satisfies the following:
⎧
⎪ 
N −1
1
γ
2N 1 +γ ⎫
⎪
r ∗ = r ∗ (γ , N ) = min ⎨
⎪ r ≥ 1|λ(r ; γ , N ) = + r γ −
⎬
γ ≤ 0⎪ (17)
⎩ k=r
k 1 − N 1 − N ⎭
Proof. See Appendix D. 
The implementation of the optimal stopping time r ∗ is shown in Algorithm 2. For γ = 0 and
a large N , we obtain the classical optimal stopping rule r ∗ = Ne . Figure 4(b) depicts the value
λ(r ; γ , N ) and the optimal stopping rules r ∗ for which λ(r ∗ ; γ , N ) ≤ 0 for different values of γ and
N = 200. As γ → 0, then r ∗ → Ne as illustrated in Figure 4(b) (for γ = 10−5 , r ∗ = 74 ≈ N /e). The
UxV reports y at observation k ≥ r ∗ for which x k > max{xl : l = 1, . . . , r ∗ }.
In Figure 4(a), we illustrate the value of the maximum probability of choosing the best QNI value
ϕ (r ∗ ; γ , N ). For γ = 0 we obtain the classical secretary problem, i.e., ϕ (r ∗ ; 0, N ) ≈ 1/e = 0.3678 for
large N . As the discount factor increases, the maximum expected payoff decreases for large N . This
indicates that we obtain a low likelihood (close to 0.161 for N = 200) in accepting the best QNI value
once γ = 1, and this is the highest probability of achieving this. Moreover, in Figure 4(b), we show
the optimal stopping rules for different values of discount factor γ and N = 200. The arrows depict
the earliest (optimal) stopping times r ∗ such that λ(r ∗ ; γ , N ) ≤ 0.
Remark 3. For 0 < γ 1 < γ 2 ≤ 1, the corresponding optimal stopping rule r 1∗ > r 2∗ . This indicates
that the UxV finds a QNI value earlier (stops the process earlier) when the discount factor is higher.
Furthermore, as the discount factor is low, then the UxV accepts a QNI value later in N ; note also
that the initial value of r ∗ → Ne as γ → 0 for all N .

4 PERFORMANCE EVALUATION
We evaluate a complete functional ground UxV that operates on two different missions, i.e., scan-
ning search for a specific value and exhaustive scan of a certain location. We focus on the latency

ACM Transactions on Internet Technology, Vol. 20, No. 3, Article 22. Publication date: August 2020.
22:14 K. Panagidi et al.

ALGORITHM 1: TOCP-DRP Algorithm

1: n ← 0
2: Th ← maximum threshold interval
3: r ← number of observations
4: α ← Change point detection threshold
5: active ← T RU E
6: counter ← 0
7: (x ∗ , r ∗ ) = LDS OST (r , γ )
8: while the algorithm is not stopped do
9: if active then /* CUSUM Algorithm described in Appendix C */
10: measure the current QNI x n
p (x (n), f )
11: sn = ln p (x (n), f01 )
n
12: Sn = k=0 sk
13: G n = Sn − min1≤k ≤n {Sk−1 }
14: if G n > α then /*A change point is detected; DRP is activated*/
15: Nd ← n
16: n̂ ← arg min1≤k ≤n Sk−1
17: Change occurs
18: active ← FALSE
19: Reset
20: n =n+1
21: else
22: if n == Th then /* maximum pausing time is reached Th */
23: active ← T RU E;
24: break
25: else[x ∗ , stopped, m] = LDSF (n, r ∗ , x n , x ∗ ) /* invocation of DRP*/
26: if stopped == T RU E then
27: active ← T RU E;
28: break
29: n =n+1

and the quality of the network during the mission and the impact of various parameters like mo-
bility. We begin with a brief description of our experiment methodology.

4.1 Experimental Platform and Methodology

The open source TurtleBot device was used as ground UxV, i.e., UGV, in our experiments as
shown in Figure 6. The TurtleBot uses a camera with depth sensor, i.e., XBOX Kinect for mapping
purposes. ROS (Robotic Operating System) is the main operating system, which is an open
source, meta-operating system executing on a Raspberry Pi, as shown in Figure 6. UGV receives
movement commands from the GCS in order to approach the given trajectory’s waypoints
and finally reaches the objective waypoint. The UGV creates a map of the environment and,
simultaneously, localizes itself in it, which is commonly known as the SLAM (Simultaneous
Localization and Mapping) technology. This is also required to safely navigate within open spaces
and proceed with informed decisions about the exploration targets. The Rviz [27] software was
used to illustrate the mapping instance created by the UGV in Figure 8.
The communication spine between GCS and UxV is a message bus platform based on Apache
Kafka, as shown in Figure 5. The ROS publish-subscribe message pattern facilitates the inter-
operability with Apache Kafka, which is basically a messaging system where clients publish
ACM Transactions on Internet Technology, Vol. 20, No. 3, Article 22. Publication date: August 2020.
To Transmit or Not to Transmit: Controlling Communications in the Mobile IoT Domain 22:15

Fig. 5. The TOCP-DRP proposed architecture for the UxV Management.

ALGORITHM 2: DRP Procedures

1: function LDS OST (r , γ )
2: for 1 < n < r do
3: y(n) = LDS (n, γ , r )
4: return y
5:
6: function LDSF (k, r ∗ , x, x ∗ )
7: stopped ← FALSE
8: position ← −1
9: if k < r ∗ then
10: if x > x ∗ then
11: x∗ = x
12: else
13: if x > x ∗ then
14: x∗ = x
15: stopped ← T RU E
16: position = k
17: return x ∗ , stopped, position
18:
19: function LDS(x, r , γ )
20: s←0
21: for x < i < γ do
22: s = s + x1
γ
2 1+γ
y = y + r 1−Nγ − γ
1− N
N
23: return y

messages and from where consumers “consume” them. The main advantages of the Apache Kafka
are (i) the high performance in delivering messages and (ii) the ability to scale out by distributing
the workload among different servers, therefore, supporting a cluster-based architecture. As such,
it can be used for transmitting UGV measurements that will be routed from producers, i.e., UxVs,
to the consumers, i.e., the GCS for monitoring, control, and so on.

ACM Transactions on Internet Technology, Vol. 20, No. 3, Article 22. Publication date: August 2020.
22:16 K. Panagidi et al.

Fig. 6. The Turtlebot UGV with XBOX Kinect and Raspberry Pi computing modules.

4.2 Model Parameters and Real Datasets

Prior to the real experimentation of our DMP and TOCP mechanisms, we consider a large-scale
experiment generated randomly as a combination of real-life datasets. The real-life datasets were
generated after multiple runs of different network conditions. We can categorize the scenarios as
follows:
(1) Good dataset, experiencing no disconnections, i.e., QNI values range in (60, 100];
(2) Medium dataset, indicating a saturated network where the QNI values range in [40, 70];
(3) Bad dataset with several disconnections experienced, i.e., the QNI values range in [20, 50].
The randomly selected blocks of all the three datasets are producing a dynamic QNI for each run of
the experiment. Based on the produced dynamic QNI, we run a number of experiments in order to
study the three design parameters of the TOCP and DRP optimal model, i.e. α, γ , and r number of
observations. We consider equal weights in Equation (5) for all network parameters, i.e., ai = 13 . We
run 100 experiments with specific threshold Thmax and the maximum number of Cmax . Figure 7(a)
shows the detection delay function D n against different α values. The D n is more adaptive to QNI
changes while α values are decreasing. The detected changes in the interval [0,0.02] are 50% more
than that of α ≥ 0.1. For the DRP model, γ is a discount factor, i.e., D n stops earlier with higher γ
values as shown in Figure 7(b). DRP adopts the LDF function in the “ passive” state. Therefore, we
can observe frequent changes from passive to the active state as expected.
We further investigate the behavior of the detection delay function D n as r approaches infinity.
As shown in Figure 7(c), for small values of r , the D n is sensitive even to small changes in net-
work. While r is working in higher intervals, D n is more reluctant to the DRP phase. However, the
probability of waiting for a large number of observations r to report a change in network quality
tends to be zero as shown in Figure 7(c).

4.3 Experiments: Performance and Comparative Assessment

We report on the experimental evaluation of our framework and mechanisms to examine their
performance based on model parameters presented in Table 4. We also provide a comparative

ACM Transactions on Internet Technology, Vol. 20, No. 3, Article 22. Publication date: August 2020.
To Transmit or Not to Transmit: Controlling Communications in the Mobile IoT Domain 22:17

Fig. 7. The detection delay function D n vs. (a) different α values; (b) different γ values; and (c) different
observations r .

Table 4. Model Parameters for the Experiments

Parameter Names Values

Change point detection threshold α [0,1]
DRP discount factor γ [1 10]
Maximum pause horizon Th 60

assessment with models found in the literature. The UxV and GCS are part of the RAWFIE platform,
which offers an experimentation framework for interconnecting numerous testbeds over which
remote experimentation can be realized.
The RAWFIE platform has been developed in the context of H2020 EU-funded (FIRE+ initiative)
project, which focuses on the MIoT paradigm and provides research and experimentation facilities
through the ever growing domain of UxVs. The RAWFIE platform is device agnostic, promoting
the experimentation under different technologies of UxVs that are equipped with different sensors,
cameras, and network interfaces. Any UxV is managed by a central controlling entity, which is

ACM Transactions on Internet Technology, Vol. 20, No. 3, Article 22. Publication date: August 2020.
22:18 K. Panagidi et al.

Fig. 8. Real-time monitoring of the robot while executing “Mission 1-Path exploration” and “Mission
2-Scanning of an area.”

programmed per case and fully overview/drive the operation of the respective mechanisms (e.g.,
auto-pilots, remote controlled ground vehicles), as shown in Figure 5. The basic requirement is
that each UxV shall be able to receive/send and decode/encode the incoming/outgoing messages
from the testbed and deliver them to the relevant on-board component.
Our TOCP-DRP optimal mechanisms extended the functionalities of RAWFIE and can be ap-
plied to any MIoT device, i.e., UAV, UGV, and USV. The used UGV in our experiments offers the
convenience to make multiple repetitions of the same experiment in the campus of the University
of Athens, Greece, unaffected from weather conditions and with real users.
The UGV was used in two real case applications: (1) scanning search for a specific sensor value
or a detection of an event designed by a user (mission 1-M1) and (2) exhaustive scan of a room
(mission 2-M2). In both missions, the user creates a path as shown in Figure 8 and the UGV should
follow the waypoints in order to reach the final destination. The depicted area is an amphitheater of
the Department of Informatics and Telecommunications of the University of Athens and a corridor
outside. During the execution of the experiments, the area is used from students and staff members
that are moving around and their mobile devices are connected to the same WiFi network.
We performed 100 runs of 10 mins duration each, where each run involves sampling for more
than N = 100 observations for every sensor integrated on UGV. The comparative assessment is
based on four different policies of decision-making: (i) the no-policy model; (ii) the heuristic
threshold-based model, in which the transmission of messages is paused when QNI falls below
a threshold; (iii) TOCP model based on Reference [25], which applies a change detection policy
triggering the “pause” mode operation (the passive mode lasts forTh, and then it is activated again);
and (iv) the hybrid TOCP-DRP model applied on both UGV and GCS. The performance metrics are
QNI measured, PER, based on packets sent and packets lost, and the end-to-end message latency.

4.3.1 Expected Performance in Mission M1. Figure 9 plots the QNI performance of the four
policies. We can observe that in mission M1, two areas of poor connectivity exist in timesteps [35–
45] and [75–90]. The no-policy, the threshold-based policy, and TOCP policy reach QNI values less
than 50%, while our TOCP-DRP policy has a mean value close to 68%. In addition, for N > 60, the
TOCP-DRP is more intolerant to network changes with mean values around [70–85].

ACM Transactions on Internet Technology, Vol. 20, No. 3, Article 22. Publication date: August 2020.
To Transmit or Not to Transmit: Controlling Communications in the Mobile IoT Domain 22:19

Fig. 9. The QNI for all the compared policies regarding the mission M1: Exploration of a Path.

Fig. 10. The QNI for all the compared policies regarding the mission M2: Scanning of an unknown Area.

The PER maximum values are for all the policies: {no − policy, threshold − based policy, TOCP,
and TOCP − DRP } are {25, 45, 15, 10}, respectively, with TOCP-DRP achieving the minimum PER,
i.e., we obtain up to 20% less PER compared with the other policies. The TOCP-DRP has better
performance than the TOCP policy because TOCP overviews network data only in active mode
and TOCP-DRP monitors QNI in both active and passive modes. The deactivation of passive mode
in TOCP happens when the threshold is reached and this means that the algorithm is triggered in
random timesteps independently of the network status. This is the reason for observing relatively
small PER values every 50 steps when the algorithm recognizes a change detection.
4.3.2 Expected Performance in Mission M2. Figure 10 shows the QNI performance of the four
comparison policies for scanning missions. The M2 mission is performed indoors where areas of
low connectivity and objects exist as obstacles to the UGV. The QNI has greater fluctuation in
this mission relative to the M1 mission. Our TOCP-DRP mechanisms from the early beginning of
mission M2, where UGV is positioned in one random corner of an amphitheater, outperforms the

ACM Transactions on Internet Technology, Vol. 20, No. 3, Article 22. Publication date: August 2020.
22:20 K. Panagidi et al.

Fig. 11. The latency (ms) measured during the no-policy and the TOCP-DRP policy in (a) mission M1 and
(b) mission M2.

other policies. The average values of QNI for all policies: {no − policy, threshold − based policy,
TOCP, and TOCP − DRP } are {68.4446, 70.8197, 65.8525, 76.3498}, respectively.
The performance of the PER is similar to the M1 mission. The PER is minimized in our TOCP-
DRP policy, where the maximum value is 10% in observations. In the remaining policies, the PER
achieve values between 20% and 30%.
4.3.3 Expected Latency in Missions M1 and M2. We plot the latency of the no-policy and our
TOCP-DRP policy in Figure 11(a) and (b) for the missions M1 and M2, respectively. The TOCP-DRP
policy is considered more efficient than the no-policy for all the observations in both missions. In
particular, in M1, we can measure 24% less end-to-end message latency compared to the original
no-policy decision-making of UGV. Moreover, the TOCP-DRP policy achieves systematically a
message latency value, which is close to 9% less of the original message latency. We can conclude
that the double hybrid optimal stopping model in the two phases of the network, i.e., active and
passive, based on the network assessment monitoring, results to missions with low end-to-end
latency and low expected PER.

5 CONCLUSIONS
We propose an in-network/on-device time-optimized decision-making model of real-time control
adaptive to changes of the network quality. This adaptive model dynamically pauses telemetry and
control messages based on derived optimal stopping rules in order to assess in real time the tradeoff
between the delivery of the messages and the network quality statistics. Our DMP policy optimally
schedules critical information delivery to a back-end system. This policy uses two optimal stop-
ping theory mechanisms based on change-detection theory and the linear discounted secretary
problem. When the quality of the network significantly changes, the UxV and the GCS can decide
in real time to pause/start the transmission of telemetry in order not to overload a saturated net-
work, or to risk completely losing the messages. Our experimental performance evaluation and
comparison assessment showed the successful delivery of messages in poor network conditions
and the moderate production of messages so as not to burden an already saturated network.
Our future research agenda includes the adoption of our TOCP-DRP model in a swarm of UxVs
in order to handle the offloading of the services/tasks, e.g., generation of telemetry, between the
swarm entities. We also plan to apply our TOCP-DRP mechanism in other types of UxVs to assess in
more detail the challenges, in the air and/or the sea, dealing with higher network quality variability.

ACM Transactions on Internet Technology, Vol. 20, No. 3, Article 22. Publication date: August 2020.
To Transmit or Not to Transmit: Controlling Communications in the Mobile IoT Domain 22:21

APPENDICES
A APPENDIX
Proof. The function L∗ (·) of the log-likelihood ratio between f¯0 and f¯1 is continuous over the
support of f¯1 and has an extremum. The proof is based on the first derivative test as shown in
Figure 12:

Fig. 12. Monotony analysis of L.

 2 
dL 2(x − μ 1 ) 2(x − μ 0 ) σ0 − σ12 μ 0σ12 − μ 1σ02
= − = x + . (18)
dx 2σ12 2σ02 σ02σ12 σ02σ12
μ 1 σ02 −μ 0 σ12
For μ 0 > μ 1 and σ1 > σ0 , we obtain that x ∗ = σ02 −σ12
. 

B APPENDIX
Proof. The KL divergence captures the discrimination between the post- and pre-change hy-
potheses and is a measure of the tractability of the change:
lnα lnα
EH 1 [Nd ] = = , (19)
I (pf 0 , pf 1 ) E f 0 [ln( p (xn , f 0) )]
p (x n , f 1)
where
   
p(x n , f 0)
I (pf 0 , pf 1 ) = E f 0 ln = [(ln( f 0) − ln( f 1))]f 0dx
p(x n , f 1)
 
−1 −(x − μ 0 ) 2 1 (x − μ 1 ) 2
= ln(2π ) − ln(σ0 ) − + ln(2π ) + ln(σ1 ) +
2 2σ02 2 2σ12

1 (x − μ 0 ) 2
× exp − dx
2πσ02 2σ02
     
σ1 1 (x − μ 1 ) 2 (x − μ 0 ) 2 1 (x − μ 0 ) 2
= ln + − × exp − dx
σ0 2 σ12 σ02 2πσ02 2σ02
   
σ1 1 (x − μ 1 ) 2 (x − μ 0 ) 2
= E 0 ln + −
σ0 2 σ12 σ02
     
σ1 1 1
= ln + 2 E 0 (X − μ 1 ) 2 − 2 E 0 (X − μ 0 ) 2
σ0 2σ1 2σ0
      1
σ1 1
= ln + 2 E 0 (X − μ 1 ) 2 + 2(μ 0 − μ 1 )E 0 (X − μ 0 ) + (μ 02 − μ 12 ) −
σ0 2σ1 2
 
σ1 σ + (μ 0 − μ 1 )
2 2
1
= ln + 0 −
σ0 2σ1 2 2 

ACM Transactions on Internet Technology, Vol. 20, No. 3, Article 22. Publication date: August 2020.
22:22 K. Panagidi et al.

C APPENDIX
In the CUSUM algorithm, we further define the generalized log-likelihood ratio G x

k
p(x (n), f 0 )
G x [k] = max 1≤m ≤k L x [k, m] = max 1≤m ≤k ln ,
n=m
p(x (n), f 1 )
= S[k] − min 1≤m ≤k S[m − 1],
where m̂ is defined as
m̂ = arg min S[m − 1] (20)
1≤m ≤k
Equation (20) shows that the decision function G[k] is the current value of the cumulative sum
S[k] minus its current minimum value. Equation (20) shows that the change time estimate is the
time following the current minimum of the cumulative sum. Therefore, each step composing the
whole algorithm relies on the same quantity: the cumulative sum S[k]. This explains the name of
the cumulative sum or CUSUM algorithm.

D APPENDIX
Proof. The expected payoff of the stopping rule r is ϕ (; γ , N ). Hence, we find the first opti-
mal stopping rule r ∗ for which it holds true that ϕ (r ; γ , N ) − ϕ (r + 1; γ , N ) ≥ 0, to stop at r given
the conditional expectation of the reward at r + 1 after observing the relative rankings up to r .
Specifically, since the conditional expectation at r + 1 is

r  1− Nk
N γ
ϕ (r + 1; γ , N ) = , (21)
N k −1
k=r +1

we can derive that:

r −1  1− Nk 1  1− Nk
γ γ γ
N
1− Nr N
ϕ (r ; γ , N ) = = + ϕ (r + 1; γ , N ) − .
N k −1 N N k −1
k=r k=r +1

Hence, in order to stop at the first r , which satisfies that:

ϕ (r ; γ , N ) − ϕ (r + 1; γ , N ) ≥ 0, (22)
we obtain that:
γ 1
N −1
γ γ
1− r + (N − r ) − 1 − ≥ 0, (23)
N N N k
k=r
which concludes

N −1
1 2
γ
1 +γ
+ r Nγ − γ ≤ 0. (24)
k=r
k 1− N 1− N
Hence, the optimal stopping time r ∗ is obtained at the first r ≥ 1, where the above equation turns
non-positive. 

REFERENCES
[1] Ibrahim Ahmed I. Alghamdi, Christos Anagnostopoulos, and Dimitrios P. Pezaros. 2019. On the optimality of task
offloading in mobile edge computing environments. In IEEE Global Communications Conference 2019. Hawaii.
[2] C. Anagnostopoulos and S. Hadjefthymiades. 2011. Delay-tolerant delivery quality information in ad hoc networks.
J. Parallel Distrib. Comput. 71(7) (2011), 974–987.
[3] C. Anagnostopoulos and S. Hadjefthymiades. 2012. Optimal quality-aware scheduling of data consumption in mobile
ad hoc networks. J. Parallel Distrib. Comput. 72(10) (Oct. 2012), 1269–1279.

ACM Transactions on Internet Technology, Vol. 20, No. 3, Article 22. Publication date: August 2020.
To Transmit or Not to Transmit: Controlling Communications in the Mobile IoT Domain 22:23

[4] C. Anagnostopoulos and S. Hadjiefthymiades. 2014. Advanced principal component-based compression schemes for
wireless sensor networks. ACM Trans. Sen. Netw. 11, 1, Article 7 (July 2014), 34 pages. DOI:[Link]
2629330
[5] ArduPilot Open Source Autopilot. Retrieved June 29, 2019 from [Link]
[6] Neil Bearden. 2006. A new secretary problem with rank-based selection and cardinal payoffs. J. Math. Psychol. 50 (02
2006), 58–59. DOI:[Link]
[7] Carlos Borrego, Gerard Garcia-Vandellós, and Sergi Robles. 2017. Softwarecast: A code-based delivery manycast
scheme in heterogeneous and opportunistic ad hoc networks. Ad Hoc Networks 55 (2017), 72–86.
[8] Carlos Borrego Iglesias, Joan Borrell, and S. Robles. 2019. Efficient broadcast in opportunistic networks using optimal
stopping theory. Ad Hoc Networks 88 (05 2019). DOI:[Link]
[9] Carlos Borrego Iglesias, Joan Borrell, and S. Robles. 2019. Hey, influencer! Message delivery to social central nodes in
social opportunistic networks. Comput. Commun. 137 (02 2019). DOI:[Link]
[10] Carlos Borrego Iglesias, Adrián Sánchez-Carmona, Zhiyuan Li, and S Robles. 2017. Explore and wait: A compos-
ite routing-delivery scheme for relative profile-casting in opportunistic networks. Comput. Networks 123 (05 2017).
DOI:[Link]
[11] Y. Cao and Z. Sun. 2003. Position based routing algorithms for ad hoc networks: A taxonomy. Ad Hoc Wireless Net-
working, Kluwer (2003).
[12] A. Dhekne, M. Gowda, R. R. Choudhury, and S. Nelakuditi. 2018. If WiFi APs could move: A measurement study. IEEE
Transactions on Mobile Computing 17, 10 (Oct. 2018), 2293–2306. DOI:[Link]
[13] Thomas S. Ferguson. Accessed May 2015. Optimal Stopping and Applications. Mathematics Department, UCLA. https:
//[Link]/∼tom/Stopping/[Link].
[14] F. Fu and M. van der Schaar. 2010. Dependant optimal stopping framework for wireless multimedia transmission. In
Accoustics Speech and Signal Processing (ICASSP). IEEE, 1–6.
[15] Ai-Chun Pang Hao-Min Lin, Yu Ge and J. S. Pathmasuntharam. 2010. Performance study on delay tolerant networks
in maritime communication environments. In OCEANS 2010 IEEE-Sydney.
[16] V. V. Veeravalli J. Unnikrishnan and S. Meyn. 2009. Least favorable distributions for robust quickest change detection.
In 2009 IEEE International Symposium on Information Theory, IEEE.
[17] Kostas Kolomvatsos, Michael Tsiroukis, and Stathes Hadjiefthymiades. 2017. An experiment description language
for supporting mobile IoT applications. In Building the Future Internet through FIRE, Martin Serrano, Niklaos Isaris,
Hans Schaffers, John Domingue, Michael Boniface, and Thanasis Korakis (Eds.). River Publishers, Gistrup, Denmark,
461–460. [Link]
[18] C. Konstantopoulos, G. Pantziou, D. Gavalas, A. Mpitziopoulos, and B. Mamalis. 2012. A Rendezvous-based approach
enabling energy-efficient sensory data collection with mobile sinks. IEEE Transactions on Parallel and Distributed
Systems 23, 5 (May 2012), 809–817.
[19] S. Mao-Y. Xiao I. Chlamtac M. Chen, V. Leung. 2009. Hybrid geographical routing for flexible energy-delay trade-offs.
IEEE Trans. Veh. Technol 58, 9 (2009), 4976–4988.
[20] A. B. McDonald. 1997. Survey of adaptive shortest-path routing in dynamic packet-switched networks. Technical
Report at the Dept of Information Science and Telecommunications (April 1997).
[21] G. V. Moustakides. 1986. Optimal stopping time for detecting changes in distributions. Ann. Statist. 14, 4 (1986), 1379–
1387.
[22] E. S. Page. 1954. Continuous inspection schemes. Biometrica 41 (1954), 100–115.
[23] E. S. Page. 1971. Procedures for reacting to a change in distribution. Ann. Math. Statist. 42, 6 (1971), 1897–1908.
[24] K. Panagidi, C. Anagnostopoulos, and S. Hadjiefthymiades. 2017. Optimal grouping-of-pictures in IoT video streams.
Computer Communications—In press (2017).
[25] K. Panagidi, I. Galanis, C. Anagnostopoulos, and S. Hadjiefthymiades. 2018. Time-optimized contextual information
flow on unmanned vehicles. In 2018 14th International Conference on Wireless and Mobile Computing, Networking and
Communications (WiMob). 185–191.
[26] L. Blazevic S. Giordano, I. Stojmenovic. 2013. Routing in delay/disruption tolerant networks: A taxonomy, survey and
challenges. IEEE Communications Surveys and Tutorials 15, 2 (2013), 654–677.
[27] Rviz ROS Software. Retrieved June 29, 2019 from [Link]
[28] Robot Operating System. Retrieved June 29, 2019 from [Link]
[29] Y. Wang, W. Peng, and Y. Tseng. 2010. Energy-balanced dispatch of mobile sensors in a hybrid wireless sensor net-
work. IEEE Transactions on Parallel and Distributed Systems 21, 12 (Dec. 2010), 1836–1850. DOI:[Link]
1109/TPDS.2010.56

Received July 2019; revised October 2019; accepted October 2019

ACM Transactions on Internet Technology, Vol. 20, No. 3, Article 22. Publication date: August 2020.