Received February 22, 2021, accepted March 19, 2021, date of publication April 12, 2021, date of current

version April 23, 2021.

Digital Object Identifier 10.1109/ACCESS.2021.3072625

An Enhanced CoAP Scheme Using Fuzzy

Logic With Adaptive Timeout for
IoT Congestion Control
PHET AIMTONGKHAM1 , PARAMATE HORKAEW 2, AND
CHAKCHAI SO-IN 1 , (Senior Member, IEEE)
1 Applied Network Technology (ANT) Laboratory, Department of Computer Science, Faculty of Science, Khon Kaen University, Khon Kaen 40002, Thailand
2 School of Computer Engineering, Institute of Engineering, Suranaree University of Technology, Nakhon Ratchasima 30000, Thailand
Corresponding author: Chakchai So-In (chakso@kku.ac.th)
This work was supported in part by a grant through the Post-Doctoral Training Program under the Thailand Science Research and
Innovation (TSRI), and in part by the National Research Council of Thailand (NRCT) through the International Research Network
Program under Grant IRN61W0006 and Khon Kaen University.

ABSTRACT Congestion management in the Internet of Things (IoT) is one of the most challenging
tasks in improving the quality of service (QoS) of a network. This is largely because modern wireless
networks can consist of an immense number of connections. Consequently, limited network resources can be
consumed simultaneously. This eventually causes congestion that has adverse impacts on both throughput
and transmission delay. This is particularly true in a network whose transmissions are regulated by the
Constrained Application Protocol (CoAP), which has been widely adopted in the IoT network. CoAP has a
mechanism that allows connection-oriented communication by means of acknowledgment messages (ACKs)
and retransmission timeouts (RTOs). However, during congestion, a client node is unable to efficiently
specify the RTO, resulting in unnecessary retransmission. This overhead in turn causes even more extensive
congestion in the network. Therefore, this research proposes a novel scheme for optimally setting the initial
RTO and adjusting the RTO backoff that considers current network utilization. The scheme consists of
three main components: 1) a multidimensional congestion estimator that determines congestion conditions
in various aspects, 2) precise initial RTO estimation by means of a relative strength indicator and trend
analysis, and 3) a flexible and congestion-aware backoff strategy based on an adaptive-boundary backoff
factor evaluated by using a fuzzy logic system (FLS). The simulation results presented here reveal that the
proposed scheme outperforms state-of-the-art methods in terms of the carried load, delay and percentage of
retransmission.

INDEX TERMS Adaptive timeout, congestion control, constrained application protocol, fuzzy logic
systems, Internet of Things.

I. INTRODUCTION will be more than 75 billion IoT devices connected to the

The Internet is no longer limited to human-to-human Internet. This vast number of devices has thus far greatly
or human-to-computer communication [1] but extends to impacted Internet architectures in various areas, e.g., signal
electronic devices that are connected to the network for ubiq- encoding techniques, media sharing, address references, and
uitously controlling, monitoring, and accessing other enti- resource allocation [5].
ties [2]. Moreover, communications can be sent among these One of the most important aspects of communication char-
devices for collaborative operations. Such a network is also acteristics in IoTs is their topology. IoT devices are topo-
known as an Internet of Things (IoT) [3]. logically categorized into three node types according to their
According to a report made by experts in industrial data primary functions [6]. The first type is a server node, whose
analysis, Statista [4], it has been predicted that by 2022, there function is to receive data from a client node or border router.
Then, it passes these data on to the network for storage or
The associate editor coordinating the review of this manuscript and processing via cloud computing. The second node type is
approving it for publication was Xiaolong Li . called a client node. Its functions are to receive and transmit

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
VOLUME 9, 2021 58967
 P. Aimtongkham et al.: Enhanced CoAP Scheme Using Fuzzy Logic

The diagram in Fig. 1 explains the congestion control

process between client and server nodes by means of CON
and ACK messages. The round-trip time (RTT) is the time
period between a client node transmitting a CON message
and receiving an ACK. During normal situations (noncon-
gestion), the RTT periods are relatively regular and thus can
usually be assumed to be an initial RTO. In contrast, during
congestion, a client node may not receive an ACK within a
specified RTO. This triggers the client node to retransmit and
to adjust the RTO with a backoff mechanism [16].
FIGURE 1. Message communication between a client and a server. However, a major concern of RTO adjustment in CoCoA
is that it cannot verify the condition of the destination node.
Therefore, the newly adjusted RTO may not be appropriate
data from and to input devices, such as a variety of sensors, for the congestion condition currently experienced by the
or to await commands to be executed, e.g., turning electrical network. Furthermore, if the RTO is updated to the maximum
equipment on and off and controlling motion [7]. To this end, backoff threshold and the client node is unable to receive
a client node can communicate with the server node, provided an ACK, then the client node will abort the retransmission,
that it is within the reachable range of the signal. However, effectively causing that packet to be undeliverable. In addi-
if the active radius of a client node does not reach the server tion to having adverse effects on the overall performance
node, a border router is required. This third node type serves of the network in terms of both throughput and delay, the
as a relaying agent that receives and transmits data from and congestion problem can result in excessive consumption of
to a client node and then passes them onto the server node [8]. energy resources in constraint devices. For instance, because
Generally, IoT communication relies on the constrained most devices in IoTs depend on battery power, network
application protocol (CoAP), which is a standard protocol architectural design in all layers, especially in congestion
specified by the Internet engineering task force (IETF) [9], management, must take into account energy consumption
in transmitting data in the application layer. CoAP was concerns [17].
designed to resemble a datagram (as in the User Datagram Recent advances in machine learning (ML) algorithms
Protocol (UDP)) to minimize its frame size and to suit trans- have currently enabled ML adoption in various areas of
mission in a constraint network. Nonetheless, CoAP was also improvements to the IoT. These include strengthening net-
designed to support connection-oriented operations by means work security [18] and authentication [19]. In particular,
of acknowledgment (ACK) messages, similar to Transmis- ML features decision support and reinforcement based on
sion Control Protocol (TCP) [10]. Thus, in cases where data both supervised learning model and information gathering
do not reach their destination, i.e., the ACK message is not from historical events. In addition, ML has been employed
received, retransmission is triggered. in congestion control schemes. For instance, ML was used
Although CoAP supports connection-oriented transmis- in optimally and automatically adjusting the RTO given cur-
sion, with modern communication standards, there remains a rent network utilization. This enhancement is beneficial to
problem when a considerable number of IoTs are connected the retransmission mechanism and hence efficient energy
to a network. For instance, on a Massive Machine-Type consumption. Thus far, ML-controlled congestion is still
Communications (mMTC) 5G network, the numbers can impeded by considering only the timing variable in adjusting
often reach 1 million connections per square meter [11], the RTO and discarding other equally dominant parameters,
[12], while with standard ZigBee, as many as 65,536 devices such as congestion conditions, concurrency, and delay due to
are feasible [13]. This overwhelming number of connections retransmission.
leads to congestion issues when transmissions from the client A fuzzy logic system (FLS) is a multivariate decision
node to the server are demanded concurrently. Consequently, support system that does not require any supervised learning.
with a large proportion of failed transactions, a number of An FLS employs a logical method in determining its input
retransmissions due to connection-oriented operations are variables, whose dependencies are fuzzy. With this method,
inevitable, imposing a large burden on the network beyond decisions are made by weighting these input variables by
its available capacity [8]. specified fuzzy rules to produce a determinate output. In fact,
The congestion issue undermines the quality of the services the FLS was developed from a rule-based scheme, which
that are otherwise normally provided by a network. To alle- has very low complexity [20], [21]. As such, the FLS has
viate congestion, CoAP employs a basic congestion control been widely adopted in research on small devices and espe-
mechanism called CoAP simple congestion control/advanced cially IoTs. For instance, the FLS has been applied to node
(CoCoA) [14]. With this mechanism, congestion is controlled selection [22] and the development of IoTs in irrigation
by specifying the retransmission timeout (RTO) for detecting systems [23].
ACK signaling after having transmitted a confirmation mes- This paper proposes a novel development in congestion
sage (CON) [15], as depicted in Fig. 1. management and control in an IoT network by incorporating

58968 VOLUME 9, 2021

P. Aimtongkham et al.: Enhanced CoAP Scheme Using Fuzzy Logic

an FLS into the backoff strategy and RTO adjustment cor- in cases where the RTT does not change over time. Accord-
responding to dynamic congestion conditions on CoAP. The ing to their experiments, it was found that CoCoA+ was
proposed scheme could significantly reduce retransmission able to improve the efficiency over CoCoA in terms of
and hence energy consumption while increasing the network both the packet delivery ratio (PDR) and delay, especially in
efficiency in terms of both throughput and transmission delay. cases where the number of connections increased. However,
The scheme evaluates the RTT and accordingly specifies the CoCoA+ remains impeded by limited thresholds for backoff
optimal RTO by means of the following key components: adjustment to only three levels. This may cause retransmis-
1) The multidimensional congestion estimator consid- sion overhead in the fourth backoff.
ers the RTT from various perspectives to create auxil- Thus far, both CoCoA and CoCoA+ share a shortcoming,
iary parameters in adjusting the initial RTO and backoff that is, that the RTO is specified as a constant. Therefore,
ratio when congestion is detected. J. J. Lee et al. [15] later proposed a novel RTO specification.
2) The accurate initial RTO configuration specifies the In their work, the retransmission count (RC) value obtained
initial RTO based on changing trends in each round and from request/response attempts was used to set an adaptive
the RTT with and without retransmission, weighted by RTO that corresponds to the current congestion condition in
statistical indicators. the network while maintaining other processes as designated
3) The adaptive-boundary backoff adjusts the RTO by in the original CoCoA. The reported experiments indicated
using a flexible backoff mechanism based on an FLS, that an adaptive RTO could help improve both throughput and
whose inputs are congestion related, to reduce retrans- PDR. However, this scheme remains limited by a relatively
mission overhead. short network lifetime. Insight into the reason behind conges-
The remainder of this paper is organized as follows: tion in a network with a large number of clients also revealed
Section II reviews the related literature and the state-of-the- that it was caused by bursty traffic, which can generally be
art congestion-aware routing methods. Based on this survey, remedied by an accurate initial RTO.
section III discusses the motivation and problem statement. Later, in 2018, S. Bolettieri et al. proposed a precise con-
Subsequently, section IV describes the proposed congestion gestion control algorithm for CoAP (pCoCoA) [25]. With this
control method for CoAP. The scheme consists of three major algorithm, the RTO was initialized precisely. First, a transmis-
components. First, a multidimensional congestion estima- sion counter (TC) was incorporated into the header of CON
tor determines the current conditions based on multivariate and ACK messages and then used to estimate the RTT. Then,
congestion parameters. Second, a novel initial RTO estima- the mean deviance (MDEV) method was adopted in RTO
tor based on a relative strength indicator and trend anal- initialization by adjusting the backoff level through statistical
ysis is proposed. Finally, an FLS is employed to evaluate smoothing. Although it is able to reduce the retransmission
the adaptive-boundary backoff factor. The simulations and rate and hence congestion, compared to CoCoA+, its main
experimental results are explained, analyzed, and discussed limitation is RTT estimation by the MDEV method. In partic-
in section V and are compared to those obtained by state-of- ular, with constant smoothing weights, pCoCoA lacks flexi-
the-art methods. Finally, concluding remarks on the outcome bility to abrupt changes in the RTT.
of the proposed scheme and future works are presented in The concept of enhancing the precision of RTT and RTO
section VI. adjustments by means of time-series forecasting has since
continuously propelled developments in the field. In 2019,
II. RELATED WORKS V. Rathod et al. proposed CoCoA++ in an article on delay
Research on congestion control schemes based on CoAP gradient-based congestion control for the IoT [26]. In their
was first introduced in 2014 by A. Betzler et al., in which proposal, the precision of RTT estimation was enhanced
CoCoA [14] was proposed. CoCoA is thus considered the by using a statistical prediction method called the CAIA
first technique that aimed to manage congestion by estimating delay gradient (CDG) in predicting RTT levels during con-
the RTT from the strong and weak RTT. The former type is gestion. Additionally, the VBF was also replaced by the
received when there is no retransmission, while the latter is probabilistic backoff factor (PBF). The reported experimental
received when there are at least two retransmissions. These results demonstrated that CoCoA++ was able to reduce
two RTT values are then used to adjust the backoff rate. With the transmission delay while maintaining the package deliv-
this technique, a variable backoff factor (VBF) is used to ery ratio and lowering retransmission compared to CoCoA.
determine an appropriate RTO. The major drawback is that However, since CoCoA++ employs a probabilistic method
CoCoA is inflexible across various network environments. and assumes a uniform random distribution in specifying
In other words, defining the RTO based primarily on the RTT the backoff sequence, in some PBF rounds, there might be
might not give the optimal results. retransmission overhead. Moreover, it is unclear whether
Later, in 2015, this same research team improved the CoCoA++ could maintain its performance in the case of a
CoCoA scheme and proposed CoCoA+ [24]. With this higher data transmission rate.
improvement, the RTO was weighted between the RTO Nonetheless, the prevailing determinants of congestion
obtained from retransmission and the accumulated average control on CoAP are not limited to optimal RTT estimation
RTO. They also introduced RTO aging to update the RTO and RTO adjustment. Because the characteristics of network

VOLUME 9, 2021 58969

 P. Aimtongkham et al.: Enhanced CoAP Scheme Using Fuzzy Logic

utilization in the IoT change constantly and greatly over TABLE 1. Nomenclature used in this study.
its lifecycle, a congestion control scheme must be designed
so that it can spontaneously accommodate this dynamism.
Accordingly, in the same year, S. Gheisari et al. proposed a
technique that integrated learning automata and game theory
to determine and adjust network parameters to suit current
congestion conditions. This technique is called the cognitive
approach for congestion control in IoTs using a game of
learning automata (CCCLA) [27]. In that work, a random
environment network model was created. Given the network
conditions, the best solution was determined by seven optimal
congestion parameters, i.e., the congestion windows, packet
size, duty cycle, retransmission timer, maximum retransmis-
sion, and contention window. The simulation results indi-
cated that the CCCLA could improve network efficiency
by automatically configuring the client node parameters to
better control congestion occurring in the network. However,
by creating a random environment of dynamic conditions,
the CCCLA may not be able to define a set of optimal network
parameters in some cases.
In the latest development by G. A. Akpakwu et al.,
the context-aware congestion control (CACC) approach for
lightweight CoAP/UDP-based IoT traffic was proposed [28].
This technique considers failed RTT in considering conges-
tion conditions in a network. They proposed three parts of issues, i.e., the initial mismatch in bursty traffic and the
RTO estimation, i.e., weak, strong and failed RTO. Their lack of a mechanism in a constraint network. The details are
experimental results revealed that CACC performed much addressed in the next subsections. Table 1 lists the definitions
more efficiently than CoCoA+. However, failed RTT can be of the notation used in this study.
evaluated only when retransmission reaches the maximum
number of attempts. This implies that CACC can improve
A. INITIAL RTO MISMATCH IN BURSTY TRAFFIC
the network efficiency only when retransmission overhead
For estimating the initial RTO in CoAP, statistical methods
occurs.
involving time-series models [14], [24], [25] are the most
In conclusion, awareness of network utilization is a key
prevalent. This is because congestion is caused by an accumu-
factor in devising congestion control methods. Typically,
lated load at a server node. The server node is thus subject to
congestion control on CoAP implies a congestion condition
overload when there are overwhelming concurrent demands
by observing the RTT during confirmation of CON and
from client nodes in the network. This congestion scenario
ACK messages. Most existing methods therefore adopt this
occurs frequently during bursty traffic.
mechanism in specifying an RTO that suits current network
Nevertheless, the changes in network traffic still conform
conditions. However, such implications are based solely on a
to cyclical trends. According to previous studies [14], [25],
timing parameter (i.e., RTT). Hence, this type of method is
[26], the initial RTO is estimated by the expressions given in
unable to predict transient congestion, such as bursty traffic
Eqs. (1) and (2).
phenomena [25]. This is because a client node must wait
to detect an abnormal RTT, which manifests only when the
RTTVARx = (1 − β) × RTTVARx
server node is already congested. ∣ ∣
Accordingly, this paper presents an improvement on con- +β × ∣RTT x − RTT xnew ∣ (1)
gestion control by proposing RTT estimation and initial RTT x = (1 − α) × RTT x + α × RTT xnew (2)
RTO adjustment on the basis of indicative parameters. These
parameters imply various aspects of network conditions, In the above equations, RTTVARx and RTTx for an RTT
including the stability and node burden of the network due to a state x are determined by either strong (RTTstrong ) or weak
fluctuating RTT. In addition, an FLS is employed to calculate (RTTweak ) RTT for a round trip without or with retransmis-
the adaptive-boundary backoff factor (ABF) to make the sion, and xnew are determined by the RTT state for a newly
retransmission backoff much more flexible. RTT measured, when β and α are fixed at 0.25 and 0.125 [24],
respectively. The resulting RTTVARx and RTTx are then used
III. MOTIVATION AND PROBLEM STATEMENT to estimate the RTO following Eq. (3).
This section presents the motivation for studying congestion
control in CoAP. The emphasis is placed on two important RTOx = RTT x + Rmax × RTTVARx (3)

58970 VOLUME 9, 2021

P. Aimtongkham et al.: Enhanced CoAP Scheme Using Fuzzy Logic

Eq. (5) expresses the VBF steps for backoff adjustment,

in which the chosen step depends on RTOinit . However,
this method is limited by fixed increasing and decreasing
steps. This results in the VBF being insufficiently accurate
for RTO adjustment. Sometimes, an unsuitable RTO, deter-
mined by providing more or less backoff than necessary, can
cause retransmission overhead and hence further delay. Later,
another study [26] proposed flexible backoff adjustment by
using a PBF, as given in Eq. (6).

FIGURE 2. An example of backoff mechanisms.
142,
 Probability_backoff > RandNo
PBF = and g > 0 (6)
where Rmax is the maximum number of retransmissions and 

Rmax is assigned to 4 [14], [24]. 07, otherwise
Used as the initial RTO (RTOinit ), RTOoverall then is cal-
culated from the weighted sum of RTOweak and RTOstrong In the above equation, the backoff sequence is adjusted by
(i.e., RTOs before retransmission occurs and when there is means of the PBF, where Probability_backoff is the proba-
no retransmission) as given in Eq. (4). bility obtained by the delay gradient (CDG) [29] process,
RandNo is a uniform random variable between 0 and 1 and
RTOoverall = γx × RTOx + (1 − γx ) × RTOoverall (4) g is the coefficient value of delay gradient [26], [29]. With
In this equation, the weighing factor γ is assigned a value this formula, the PBF adjusts the RTO in two steps to 1.42
of 0.5 or 0.25 for CoCoA [14] or CoCoA+ [24], respectively. or 0.7 when the RTT increases or decreases, respectively.
However, note that the initial RTO (RTOoverall ) following Although backoff adjustment by PBF can minimize the delay
Eq. (4) lacks flexibility due to the use of constants in adjusting better than that by VBF, because of the probability calculation
the ratios between the relevant parameters. per RTT change, it remains limited by the probability pre-
To resolve this issue, this research proposes a more flexible dicted from a uniform distribution, the assumption of which
approach, by which the RS indicator is exploited to enhance may not be faithful to the actual congestion in the network
the initial RTO configuration. In addition, employing trend during bursty traffic.
analyses is suggested in order to make the initial RTO adjust- Examples of backoff sequences created by the above meth-
ments much more accurate. ods are depicted in Fig. 2. In this figure, a sequence of five
backoff obtained by using the BEB, PBF, and VBF in our
B. LACK OF A BACKOFF MECHANISM IN A CONSTRAINT preliminary experiments are shown. These methods resulted
NETWORK in different backoff boundaries.
It can be concluded from the above hypotheses that the
During congestion control, RTO backoff is designated after
backoff RTO proposed in [14], [24], [25], and [27] was pri-
retransmission, which occurs after a client node does not
marily determined from the RTT. However, in a constrained
receive an ACK within a previously specified RTO. This
network, especially one with CoAP, the RTT varies with the
implies that the initial RTO does not correspond to the
load borne by the server node. Thus, frequent changes in
congestion condition in the network. This results in hav-
the server load [14], [24], [25], [27], [28] in turn result in
ing to increase the RTO via a backoff mechanism. To this
a fluctuating RTT. Therefore, this research proposes a highly
end, existing studies [14], [24]–[26] propose various backoff
efficient backoff RTO estimated from various parameters by
approaches, examples of which are illustrated in Fig. 2. Tradi-
using an FLS.
tionally, RTO backoff in CoAP was based on binary exponen-
tial backoff (BEB), where the RTO increases exponentially.
However, the drawback of BEB adjustment is that the timeout IV. FUZZY LOGIC-BASED COAP CONGESTION
sequence is subject to strictly exponential growth. CONTROL SCHEME
To address this problem, other backoff strategies have been This section provides detailed descriptions of the three major
proposed in the literature, e.g., CoCoA+ [24], where backoff FLCoCoA components, an overview of which is presented in
is adjusted based on a VBF, whose sequence is more flexible. the diagram shown in Fig. 3. They are explained as follows:
Specifically, the RTOs are divided into three levels based • Multidimensional Congestion Estimator: This com-
on the RTO in an initial round. In addition, RTO aging is ponent produces an RTT estimate that better corre-
incorporated in setting the threshold during backoff sequence sponds to early congestion based on multidimensional
adjustment, as given in Eq. (5). parameters.
 • RTO Estimation with Trend Adjustment: This com-
3,
 RTOinit < 1seconds
ponent incorporates predictive trends in adjusting the
VBF = 2, 1 ≤ RTOinit ≤ 3seconds (5) RTO to spontaneously respond to actual congestion at


13, RTOinit > 3seconds a given time.

VOLUME 9, 2021 58971

 P. Aimtongkham et al.: Enhanced CoAP Scheme Using Fuzzy Logic

as shown in Eq. (8).

( )
100
RSIndicator = 100 − (8)
1 − RS
where RSIndicator is within the range [0, 1] and RS is a
temporal average that represents the relationship between
RTTstrong and RTTweak . The resulting RSIndicator is used as an
auxiliary variable during the subsequent RTO estimation (see
Section IV-B).
• RTT Interval: This variable measures the stability of
the RTT on the destination network (server node) within
a specific duration. This measurement implies the net-
work utilization when determining the backoff time for
the adaptive-boundary backoff factor (see Section IV-C).
Generally, the RTT in CoAP congestion control is deter-
FIGURE 3. The overview of FLCoCoA.
mined in the time domain [14], [24], [25]. However, that even
during the no-congestion period, the number of received ACK
• Fuzzy Logic Adaptive-boundary Backoff: This com- at the client varies (in other words, no retransmission yet) can
ponent uses a novel flexible backoff strategy that is be used as the extra information to estimate the server load.
adaptive to changes in the network environment during In particular, when the server node is in a near-congestion
congestion. state or experiences a higher load, the number of ACKs
received by a client node decreases accordingly. This change
A. MULTIDIMENSIONAL CONGESTION ESTIMATOR motivated the design of a threshold for detecting such con-
The first part of FLCoCoA involves constructing multidimen- gestion via the RTT Interval. The RTT Interval is evaluated on
sional auxiliary variables. They consist of an RS indicator, the basis of the accumulated amount of ACK packets received
RTT Interval, and RTT Jitter. These variables are utilized in from the beginning of data transmission until retransmission
realistically estimating congestion conditions, both when set- occurs in the RTO round.
ting the initial RTO (as described in Section IV-B) and deter- The resulting RTT Interval is used to determine the
mining the adaptive-boundary backoff factor (Section IV-C). congestion for the adaptive-boundary backoff factor (see
A detailed description and analysis of each variable are pro- Section IV-C).
vided as follows: • RTT Jitter: This parameter measures instability or fluc-
• RS Indicator: Statistical techniques are adopted in tuation in the transmission delay. As such, jitter loosely
assigning weights to two sets of data when they are represents the chance of congestion occurring at the
similar. In this paper, weighting formulas derived from destination node due to the concurrent network demand.
statistical indicators for RTO estimation are proposed During transmission on a constraint network, the data
(Section IV-B). Previously, with congestion analyses in are divided into several small pieces, called packets,
CoAP, as suggested in [14], [24], and [25], the RTT sent out simultaneously. In normal network utilization,
was estimated from RTTstrong and RTTweak . During mod- a client node receives an acknowledgment of each packet
erate congestion, both RTTs are very similar, making from the server node within a regular RTT. In con-
it difficult to determine whether future congestion will trast, during congestion, an overloaded destination node
increase. In contrast to those works, the calculation of requires extra time to process the number of incoming
the RS indicator was adapted from the idea of relative packets, causing an irregular delay in sending out an
strength [30], which is normally used to determine the acknowledgment of each packet and thus in its RTT. This
periodic relationship between a large and a small value irregularity due to node congestion is called jitter.
from their average within a specific duration. The pro- RTT Jitter is thus defined as a fluctuating time span needed
posed RS indicator enables an analysis of time-series to transmit and receive data in a network. This fluctua-
data that indicates the divergence of two datasets. First, tion, characterized by the difference delay in RTTs between
the relative strength (RS) is computed, following Eq. (7). 2 consecutive delays, is caused by irregular delay due to
the overutilization of the network. While the RTT should
avarage(RTT strong ) be almost invariable in an uncongested network, one may
RS = (7)
avarage(RTT weak ) attribute its jitter to, for example, high concurrent loads that
cause delays in data exchange or congestion due to channel
According to Eq. (7), the relative strength of the RTT is sharing [31]. The RTT Jitter is characterized by Eq. (9).
given by average of the ratio between RTTstrong and RTTweak .
The resulting RS is then used to calculate the RS indicator, RTTJitter = |RTT t−1 − RTT t | (9)

58972 VOLUME 9, 2021

P. Aimtongkham et al.: Enhanced CoAP Scheme Using Fuzzy Logic

In Eq. (9), the RTT Jitter is expressed as absolute dif- Accordingly, the future trend of the RTT is predicted from
ferences delay between the RTTs (i.e., RTTt and RTTt−1 ). the value in the previous interval (Ft−1 ), the relative weight
The resulting RTTJitter parameter is then used to determine obtained from the RS indicator (RSIndicator ), and the current
the congestion for the adaptive-boundary backoff factor (see trend (RTTTrend ), given by Eq. (11). The resulting prediction
Section IV-C). is used in adjusting the initial RTO (RTOinit ), as shown in Eq.
(13), to make its configuration more precise.
B. INITIAL RTO RELATIVE WEIGHT ESTIMATION WITH RTOinit = RTOest + FTrend (13)
TREND ADJUSTMENT
This section presents methods of estimating the initial RTO where RTOest are the adjusted initial RTO (RTOinit ) and that
and optimally adjusting the RTO by using relative weights estimated by means of the relative weight method. The FTrend
and RTT Trend evaluation, respectively. To this end, the initial is the trend predicted by Eq. (12).
RTO estimation is based on previous work on forecasting
C. ADAPTIVE-BOUNDARY BACKOFF FACTOR BASED
trends [32]. Estimations of both the initial RTO and its trend
ON FUZZY LOGIC
are explained below.
This process involves enhancing the backoff efficiency by
1) RTO RELATIVE WEIGHT ESTIMATION
adjusting the RTO in response to the present congestion in
the network. To this end, both the flexibility and accuracy of
This section describes the improvement in RTO estimation
the RTO setting are considered while reducing the number of
compared with [14] and [24]. In those works, the RTO
retransmissions due to its overhead. The process consists of 2
was estimated based on constant ratios. Note, however, that
main stages. First, an FLS is utilized to set the weight of the
applying different RTO stages is also feasible to enhance the
backoff time boundary. Second, the efficiency of the backoff
performance for RTO estimation [28], and thus, we consider
strategy is enhanced by means of the FLS adaptive boundary
both the weak and strong stages in this research. With the
after retransmission.
present improvement, however, the ratio is adjusted opti-
mally based on the relative strength indicator, as defined in
1) FUZZY LOGIC SYSTEM
Section IV-A. Therefore, the initial RTO can be estimated
This step applies an FLS to compute the weight considering
by Eq. (10).
multidimensional congestion-related variables, i.e., the net-
work stability (RTT Interval), RTT fluctuation (RTT Jitter),
RTOest = RS Indicator × RTOweak + (1 − RS Indicator )
and statistically analyzed RTT trend (RTT Trend). The FLS
×RTOstrong (10) consists of four key modules: the fuzzifier, fuzzy rule, fuzzy
inference engine, and defuzzifier. Their detailed analyses are
where RSIndicator is an indicator obtained from the RS indi- presented below.
cator (see Section IV-A) and RTOweak and RTOstrong are the
• Fuzzifier: This module converts inputs in crisp sets
approximately RTO when retransmission occurs and when
into fuzzy sets. In this paper, three input variables are
there is no retransmission, respectively. The resulting RTOest
considered. They are RTT Trend, RTT Interval, and RTT
is employed in the calculation as described in the next
Jitter, whose units and ranges differ, each computed
subsection.
during the individual RTO round. Therefore, the former
is normalized to within the range [–1, 1], i.e., the two
2) RTO TREND ADJUSTMENT RTTs during a particular interval over the maximum of
This section extends RTO estimation by considering the pre- the RTT difference. The remaining two are restricted to
diction of the RTT Trend in each round. Insights into changes within the range [0, 1] and evaluated as follows: for RTT
in the RTT between intervals reveal that they also affect the Interval, the number of received ACKs over the number
prediction of the RTO. The RTT Trend is defined simply as of transmitted packets, while for RTT Jitter, the differ-
the difference between RTTs in consecutive intervals, i.e., t ence delay of the pair of RTTs over the maximum of the
and t – 1, as expressed in Eq. (11). difference delay within the RTO round. Subsequently,
membership functions (MFs) are defined. According to
RTTTrend = RTT t−1 − RTT t (11) a preliminary experiment, among different types of MFs,
i.e., Gaussian, triangular, trapezoidal, generalized bell,
This trend represents the RTT change in each interval t, and sigmoid functions, the triangular function was the
where RTTt−1 and RTTt are the RTTs of the transmission in most suitable for partitioning the inputs, whose bands
the previous and current intervals, respectively. The resulting were generally narrow. Note that based on our experi-
RTTTrend is used both to predict the future trend by smooth- ment, Gaussian, trapezoidal, and generalized bell MFs
ing, as in Eq. (12), and as a parameter in the FLS (see were suitable for a wide range of variables when con-
Section IV-C.1). sidering the crisp input variables. Similarly, the sigmoid
function is appropriate for constructing a single-level
FTrend = Ft−1 + RSIndicator × (RTT Trend − Ft−1 ) (12) MF [20], [21].

VOLUME 9, 2021 58973

 P. Aimtongkham et al.: Enhanced CoAP Scheme Using Fuzzy Logic

Thus, the triangular MF expressed in Eq. (14) was selected

in the analyses below. Here, xi is a real number (x ∈ R) in
a crisp set of input parameters, µ (xi ) is the triangular MF
of the input xi and ã is the membership level of the MF,
whose members are low (L), medium (M), or high (H); that is,
ã ∈ {aL , aM , aH }.
   M 
L L L M
x − a  a − a , a ≤ x ≤ a

µ (x)i = aH − x  aH − aM , aL ≤ x ≤ aH (14)


0, otherwise

• Fuzzy Rule: This module defines the rules for determin-

ing the MF level of an output MF. These rules were spec-
ified for all crisp inputs. The condition of each rule was
determined by hypothesis testing on all possible values
drawn from a probability distribution model. Therefore,
the number of rules depended on the crisp input variables
and MF stages. As noted earlier, this study used 3 crisp
input types, each of which was assignable to 3 MF
stages, i.e., L, M, and H.
This resulted in a total of 27 (33 ) fuzzy rules. These rules
are presented in Table 2. There, RTT Trend represents the
trend of RTT adaptation obtained by a statistical method, RTT
Interval represents the stability of the destination (server)
node, and RTT Jitter represents the congestion condition
occurring in the shared network.
• Fuzzy Inference Engine: This module computes the
output weights for the adaptive backoff weight (ABW)
by means of fuzzy inference. To this end, the weights of
all involved inputs were aggregated by intersection, with
W being the MF of the ABW and the MF containing
three membership levels, i.e., L, M, and H. Fig. 4 illus-
trates the process of ABW fuzzy inference. The process
consists of 27 fuzzy rules, the details of which are listed
in Table 2.
Fig. 5 shows an inference when the input values of the FIGURE 4. ABW membership functions, i.e., RTT Trend (a), RTT Interval (b),
normalized RTT Trend, RTT Interval, and RTT Jitter are RTT Jitter (c), and Output Weight (ABW) (d).

0.341, 0.440, and 0.256, respectively. The output is inferred

from the fuzzy set by the AND operator. That is, the least
cross-sectional area of the MF from these three inputs was The resulting CoG, which ranges between 0 and 1, is inter-
crossed over to produce the output MF of each rule. Aggre- preted as the final ABW. According to the defuzzifier exam-
gation was then performed to combine the areas of the output ple illustrated in Fig. 5, the final output CoG is 0.534.
MFs from all 27 rules.
• Defuzzifier: This final FLS module interprets the output 2) ADAPTIVE-BOUNDARY BACKOFF FACTOR (ABF)
in fuzzy set form as a weight. To this end, the center of This process involves an RTO backoff strategy. Generally,
gravity (CoG) of the area in the output ABW aggregated backoff is activated when a client node has not received
in the previous module is evaluated. This method finds an ACK within the preconfigured initial RTO. To make the
the center of the fuzzy set’s area and returns the symmet- strategy more flexible, this paper proposes adaptive backoff
rical corresponding crisp value [20]. Eq. (15) expresses level setting in each round, considering both the initial RTO
the CoG as the normalized MF of the weight (ABW), and the RTO in the previous round based on the FLS weights.
µ(W ), over the membership level of the output (ABW), This new scheme is called the ABF. The boundary term in
ω(W ). the ABF ensures that the backoff level corresponds to the
∫ current congestion based on the ABW obtained by the FLS.
ω (W ) µ (W ) dW In Eq. (16), the boundary backoff is calculated from the RTO
CoG = ∫ (15)
ω (W ) dW in the previous round (RTOn−1 ) and that initialized by the

58974 VOLUME 9, 2021

P. Aimtongkham et al.: Enhanced CoAP Scheme Using Fuzzy Logic

TABLE 2. Examples of 27 ABW fuzzy rules.

FIGURE 5. An example of ABW fuzzy inference and the corresponding

final output.

in the first round with RTOinit . While the backoff count has
not reached the backoff threshold (backoffMax), the backoff
boundary is calculated from either of the RTO conditions as
per Eq. (16). Subsequently, the FLS is employed to adjust the
backoff rate in each round, as described in Section IV-C.1.
The resulting ABW is then used to adjust the backoff RTO
for the next retransmission, as given in Eq. (17).
method in Section IV-B (RTOinit ). In addition, RTOGain is
the threshold for incremental rate of RTO per round such Algorithm 1 ABF Backoff Adjustment
that RTOGain ∈ {RTOGainmin , RTOGainmax } [24]–[26]. Input: RTOn−1, ABW
ABF Output: RTO_backoff

RTOn−1 × RTOGainmax , RTOn−1 < RTOGainmin 1 INITIALIZE RTOinit ,backoffCount = 0



 < RTOn−1< RTOGainmax 2 WHILE (backoffCount < backoffMax) DO

 ( )
 RTOGainmax 3 Calculate backoff boundary based on Eq. (16)
= RTOn−1 × , RTOGainmin ≤ RTOn−1 4 IF (RTOn−1 < RTOGainmin ) THEN

 2

 5 ABF = RTOn−1 × RTOGainmax


 ≤ RTOGainmax 6 ELSE IF (RTOGainmin ≤ RTOn−1 ≤

RTOn−1 × RTOGainmin , RTOn−1 > RTOGainmax RTOGainmax ) THEN
(16) 7 ABF = RTOn−1 × RTOGainmax × 0.5
8 ELSE IF (RTOn−1 > RTOGainmax ) THEN
Once the boundary factor has been evaluated, the backoff
9 ABF = RTOn−1 × RTOGainmin
boundary of the RTO (RTOn or RTO_backoff) is adapted from
10 END IF
that in the previous round (RTOn−1 ) and weighted by the ABF
11 Calculate RTO_backoff based on Eq. (17)
and ABW, as given in Eq. (17).
12 RTO_backoff = (RTOn−1 × ABF)
RTO_backoff = (RTOn−1 × ABF) × ABW (17) × ABW
13 RETURN RTO_backoff
In summary, Algorithm 1 explains the steps involved in
14 END WHILE
ABF backoff adjustment. It starts by initializing the RTO

VOLUME 9, 2021 58975

 P. Aimtongkham et al.: Enhanced CoAP Scheme Using Fuzzy Logic

V. PERFORMANCE EVALUATION
To demonstrate the performance and efficiency of FLCoCoA,
simulations were performed, and the results obtained by
the proposed method were compared to those of the state-
of-the-art methods reviewed in the related studies section.
They were pCoCoA [25], CoCoA+ [24], CCCLA [27], and
CoCoA++ [26].
A. SIMULATION SETTINGS
In the following experiments, this study employed CoAP on
the IoT simulation software package Cooja [33], which was
run on the Contiki operating system version 3.0 [34], and the
FLS model simulation software Octave version 4.2 [35]. Sim-
ilar to the experiments conducted in [14], [24], [25], and [27],
the entire process was executed on a virtual machine installed
on an x86 computer system running Ubuntu Linux 14.04 LTS.
Likewise, the parameter settings defined in Table 3 followed
those prescribed in [14], [24], [25], and [27].
To compare the results with related works [14], [24], [25],
and [27], this study evaluated three indicative metrics:
• Carried load (average throughput per node), defined
as the average transmission per node given different
simulated data rates in the range of 1–10 kbps,
• Delay (average end-to-end delay), defined as the aver-
age receiving delay between the client and server
nodes given different simulated data rates in the range FIGURE 6. Four network topologies considered in the simulations.
of 1–10 kbps, and
TABLE 3. Simulation parameters of [14], [24], [25], and [27].
• Retransmission (average percentage of retransmis-
sion), defined as the average retransmission percentage,
which indicates the efficiency and precision in determin-
ing the RTO.
To elucidate the proposed mechanisms on a variety of net-
works, simulations on four different network topologies were
performed and compared with the related works. Following
the state of the art [14], [24]–[26], there are 4 main topologies
with 36 nodes in 100 square meters that are 1) uniformly
distributed over 6 × 6 grids, 2) randomly deployed, and
3) distributed in a three-leaf group forming a flower pattern
that include 6, 13, and 15 nodes with a server node and a
border router located in the central and in the large group,
respectively, and 4) a linear distribution of 20 nodes in a
chain pattern. In all cases, the server, border router and client
node are illustrated in magenta, yellow and green circles
respectively. The communication range was specified to be
10 meters (interference 20 meters), while the buffer size, B. SIMULATION RESULTS AND DISCUSSION
backoff threshold, and RTOGain were set to 8 packets, 5, This section presents and discusses the experimental results.
and {1 s, 3 s}, respectively [14], [24]–[27]. In this setup, Those obtained by the FLCoCoA methods were bench-
we use accumulated ACKs as RTT Interval and the last pair marked in terms of performance and efficiency against the
of RTTs in each RTO for RTT Trend and RTT Jitter in the state-of-the-art methods, i.e., pCoCoA [25], CoCoA+ [24],
fuzzification process. In addition, to validate the reliability CCCLA [27], and CoCoA++ [26].
of the simulation model, the experiments were carried out The discussion is divided into three areas according to the
10 times in each scenario. Then, for each experimental setup, indicative metrics. They are the 1) carried load, 2) delay,
the average and standard deviation were evaluated at a confi- and 3) retransmission ratio. The first metric measures the
dence interval (CI) of 95%. transmission efficiency from a client to a server, while the sec-
Fig. 6 depicts example node distributions in the simulated ond measures the average delay time in transmission, which
topologies, in which server nodes, border routers, and client determines the efficiency of specifying the RTO timeframe.
nodes are denoted in red, yellow, and blue, respectively. Last, the retransmission ratio represents the efficiency of the

58976 VOLUME 9, 2021

P. Aimtongkham et al.: Enhanced CoAP Scheme Using Fuzzy Logic

Generally, it was revealed that when the offered load reached

network congestion levels, the carried load level similarly fell
below that of the offered load in all scenarios. However, they
differed in the severity of congestion. In particular, the grid
topology exhibited the highest congestion level, followed by
the random, flower, and chain distributions.
Fig. 7a plots the carried load on the grid topology. It is
evident that congestion started at the offered load level
of 4 kbps and became intense when the offered load was
7–10 kbps. Similarly, for all topologies, FLCoCoA performed
most efficiently at offered load levels of 3.9–5.41 kbps dur-
ing congestion, followed by pCoCoA, CoCoA+, CCCLA,
and CoCoA++ at levels of 3.81–5.2 kbps, 3.72–4.95 kbps,
3.61–4.43 kbps, and 3.47–3.94 kbps, respectively.
Likewise, the simulations on the random topology indi-
cated much more severe congestion than those on the grid
topology, as displayed in Fig. 7b. These results were clearly
noted at offered load levels of 4–10 kbps. However, FLCo-
CoA reached maximum efficiency at offered load levels
of 3.93–5.76 kbps, followed by pCoCoA, CoCoA+, CCCLA,
and CoCoA++ at levels of 3.66–4.98 kbps, 3.55–4.8 kbps,
3.34–4.17 kbps, and 3.12–3.78 kbps, respectively.
For the flower and chain topologies, as shown
in Fig. 7c and 7d, respectively, congestion was lower
than that in the grid and random topologies. In these
topologies, FLCoCoA performed best at transmission rates
of 2.94–4.71 kbps and 2.73–3.96 kbps, respectively, fol-
lowed by pCoCoA, CoCoA+, CCCLA, and CoCoA++,
at transmission rates of 2.95–4.52 kbps, 2.76–4.29 kbps,
2.56–4.02 kbps, and 2.33–3.65 kbps, respectively, for the
flower topology and at rates of 2.7–3.76 kbps, 2.6–3.57 kbps,
2.6–3.25 kbps, and 2.68–3.14 kbps, respectively, for the chain
topology.
By relative analyses, in which percentage improvements
were calculated for each scenario, it was found that FLCo-
CoA had a higher transmission efficiency than did pCoCoA,
CoCoA+, CCCLA, and CoCoA++ in the grid topology, with
improvement ratios of 5.4%, 11.33%, 17.5%, and 24.75%,
respectively, and in the random topology, with improvement
ratios of 5.84%, 13.06%, 24.39%, and 37.69%, respectively.
The reason for the increased efficiency of FLCoCoA in
the latter case was that in the random topology, the client
nodes are very heterogeneous, but the FLS managed to
calculate a flexible backoff that adapted to diverse traffic
conditions. For the flower and chain topologies, FLCoCoA
also had a higher percentage improvement in transmission
FIGURE 7. Simulation results for the average carried load versus the
than pCoCoA, CoCoA+, CCCLA, and CoCoA++, at rates
offered load in four network topology scenarios. of 6.44%, 12.24%, 21.76%, and 31.79%, respectively, and
of 12.7%, 20.69%, 33.72%, and 48.76%, respectively, in each
topology.
prescribed backoff. In turn, the number of retransmissions Further analyses of the flower and chain topologies
determines the future congestion. revealed that FLCoCoA could greatly increase its efficiency
Each experiment was conducted on four different physical in the chain topology due to the differing distances from the
topologies, i.e., grid, random, flower, and chain topologies. clients to the server node. This characteristic, in turn, clearly
Fig. 7 shows the carried load of FLCoCoA. In this exper- strengthened the capability of FLCoCoA in specifying the
iment, different offered loads of 1–10 kbps were simulated. initial RTO.

VOLUME 9, 2021 58977

 P. Aimtongkham et al.: Enhanced CoAP Scheme Using Fuzzy Logic

grid topology, followed by the random, flower, and chain

topologies. This trend conforms to that of the carried load
results.
Further detailed analysis revealed that in the grid
topology, FLCoCoA yielded the lowest average delays
of 293–2,699 ms, as shown in Fig. 8a, followed by pCo-
CoA, CoCoA+, CCCLA, and CoCoA++, with delays
of 352–3,123 ms, 408–3,292 ms, 459–3,576 ms, and
551–3,714 ms, respectively.
The results for the random topology are shown in Fig. 8b.
It is evident that, on average, the delays are higher than
those in the other topologies. In particular, FLCoCoA gave
the lowest delays of 266–2,286 ms, followed by pCoCoA,
CoCoA+, CCCLA, and CoCoA++, whose delays were
304–2,736 ms, 404–3,262 ms, 459–3,829 ms, and
560–4,351 ms, respectively.
It is similarly revealed by Fig. 8c that the delays in the
flower topology were on average lower than those of the
grid and random topologies. FLCoCoA gave the lowest
delays of 219–1,619 ms, followed by pCoCoA, CoCoA+,
CCCLA, and CoCoA++, whose delays were 224–1,717 ms,
230–1,874 ms, 204–2,045 ms, and 313–2,229 ms,
respectively.
Similar observations can be made for the chain topol-
ogy, as shown in Fig. 8d. In this figure, FLCoCoA
gave average delays of 150–1,045 ms, followed by pCo-
CoA, CoCoA+, CCCLA, and CoCoA++, giving delays
of 198–1,105 ms, 250–1,203 ms, 290–1,411 ms, and
372–1,600 ms, respectively.
By analyzing the percentage improvement over the com-
pared methods, it was found that FLCoCoA had lower delay
ratios than pCoCoA, CoCoA+, CCCLA, and CoCoA++ in
the grid topology, at 15.13%, 22.69%, 30.01%, and 38.36%,
respectively, and in the random topology, at 12.89%, 24.14%,
31.21%, and 41.32%, respectively. In the other topolo-
gies, FLCoCoA had better percentage delays than pCoCoA,
CoCoA+, CCCLA, and CoCoA++, at 16.24%, 30.47%,
41.14%, and 48.44%, respectively, in the flower topology and
16.89%, 30.25%, 41.32%, and 48.24%, respectively, in the
chain topology.
Further analysis of these delays revealed that RTO backoff
adjustment can help reduce retransmission overhead due to
the smaller number of adjustments required, leading to less
delay.
The third simulation measured the retransmission ratio.
FIGURE 8. Simulation results for the average delay versus offered load in It assessed the efficiency of the proposed backoff strategy.
four network topology scenarios. The lower the retransmission is, the more accurate the initial
RTO, the more suitable the backoff, and the better the con-
gestion reduction. Following the corresponding experiments,
Fig. 8 shows measurements of the transmission delay it may be concluded that when increasing the offered load
between client and server nodes. Similar to the carried load from 1 to 10 kbps, the retransmission established not only an
experiment, simulations were carried out for offered loads in increasing trend but also similar percentage readings in all
the range of 1–10 kbps. It was generally revealed that the topologies.
delays had a tendency to increase as congestion increased Regarding the grid topology, Fig. 9a shows that the
in all topologies. However, they differed in terms of the percentages of retransmission obtained by using FLCo-
specific delay levels. That is, the delay was highest in the CoA were the lowest on average, 2.89–60.4%, followed

58978 VOLUME 9, 2021

P. Aimtongkham et al.: Enhanced CoAP Scheme Using Fuzzy Logic

For the random topology, it is notable in Fig. 9b that

among all topologies, FLCoCoA resulted in the lowest per-
centages of retransmission, 2.92–53.14% on average, fol-
lowed by those resulting from pCoCoA, CoCoA+, CCCLA,
and CoCoA++, which were 3.07–54.8%, 3.15–57.1%,
4.21–59.8%, and 5.1–63.5%, respectively.
Similar characteristics were observed in the flower topol-
ogy, as shown in Fig. 9c. On average, FLCoCoA gave the
lowest percentages of retransmission, 2.51–55.11%, followed
by pCoCoA, CoCoA+, CCCLA, and CoCoA++, giving
percentages of retransmission of 2.67–62.4%, 2.83–64.32%,
3.11–67.56%, and 3.28–68.62%, respectively.
Likewise, in the chain topology, as shown in Fig. 9d,
FLCoCoA gave the lowest percentages of retransmission,
2.32–42.4% on average, followed by pCoCoA, CoCoA+,
CCCLA, and CoCoA++, at 1.85–50.2%, 2.17–52.35%,
2.39–58.32%, and 2.46–62.2%, respectively.
For all topologies, the improvement in these percentages
made by FLCoCoA was the greatest compared to pCoCoA,
CoCoA +, CCCLA, and CoCoA++. The improvements
were on average 5.21%, 9.69%, 12.40%, and 16.71%, respec-
tively. For FLCoCoA, these reduced retransmissions are
attributed to the initial RTO and ABF being adaptive to the
actual congestion. These resulted in, for each transmission, a
client node receiving the ACK within the specified RTO and
hence the least possible retransmission.
In summary, by assessing the overall measurements in
all four topologies, it is evident that the great efficiencies
of FLCoCoA in various aspects were due to the enhanced
initial RTO and the flexible backoff strategy by using ABF.
Consequently, these features resulted in the most efficient
backoff, with which the client nodes received an ACK within
a specified RTO and hence retransmissions were significantly
reduced.

VI. CONCLUSION
This research presents a novel congestion control scheme
for CoAP. In particular, it improves the RTOinit estimation
mechanism and RTO backoff method. Its main contributions
are 1) determining the multidimensional RTT based on the
RS indicator, RTT Interval, and RTT Jitter, 2) initial RTO
estimation by the RS indicator during weight adjustment from
both RTT values in conjunction with trend prediction, and
3) a flexible backoff method based on an FLS to determine
congestion-driven factors and optimally adjust the backoff
ratio according to the network condition.
It was demonstrated in the experiments that all these con-
tributions jointly enhanced the congestion control efficiency.
In particular, compared to pCoCoA, CoCoA+, CCCLA
FIGURE 9. Simulation results for the average delay versus offered load in and CoCoA++, they increased the data transmission rate
four network topology scenarios. by 7.59%, 14.33%, 24.34%, and 35.74% and reduced
retransmissions by 5.21%, 9.69%, 12.40%, and 16.71%,
respectively.
by those obtained by using pCoCoA, CoCoA+, CCCLA, Thus far, FLCoCoA has not been specifically designed to
and CoCoA++, which were 3.12–62.38%, 3.14–64.3%, support congestion control on CoAP in the nonconfirmable
3.2–67.5%, and 3.25–68.6%, respectively. mode. Furthermore, for practicality, intensive simulation

VOLUME 9, 2021 58979

 P. Aimtongkham et al.: Enhanced CoAP Scheme Using Fuzzy Logic

with a real testbed should be investigated, including various [16] C. Vallati, F. Righetti, G. Tanganelli, E. Mingozzi, and G. Anastasi, ‘‘Anal-
topologies, traffic heterogeneity, and diverse parameters and ysis of the interplay between RPL and the congestion control strategies
for CoAP,’’ Ad Hoc Netw., vol. 109, Dec. 2020, Art. no. 102290, doi:
setups. In addition, the diversity and density of the node dis- 10.1016/j.adhoc.2020.102290.
tribution merit consideration for further study, i.e., variations [17] L. P. Verma and M. Kumar, ‘‘An IoT based congestion control
of the number of nodes and scaling areas in the experimental algorithm,’’ Internet Things, vol. 9, Mar. 2020, Art. no. 100157, doi:
10.1016/j.iot.2019.100157.
field, energy consumption in the fuzzy logic system, and [18] F. Hussain, R. Hussain, S. A. Hassan, and E. Hossain, ‘‘Machine learning
applications to real devices, which were not considered dur- in IoT security: Current solutions and future challenges,’’ IEEE Com-
ing network simulation. In future investigations, we intend mun. Surveys Tuts., vol. 22, no. 3, pp. 1686–1721, 3rd Quart., 2020, doi:
10.1109/COMST.2020.2986444.
to include experiments, protocols, and energy consumption [19] P. Punithavathi, S. Geetha, M. Karuppiah, S. H. Islam, M. M. Hassan,
measurements using real-world IoT devices. In addition, and K.-K.-R. Choo, ‘‘A lightweight machine learning-based authentica-
tion framework for smart IoT devices,’’ Inf. Sci., vol. 484, pp. 255–268,
we are working on the energy usage from the computing May 2019, doi: 10.1016/j.ins.2019.01.073.
perspective in MicaZ and Tmote Sky. [20] P. Hilletofth, M. Sequeira, and A. Adlemo, ‘‘Three novel fuzzy logic con-
cepts applied to reshoring decision-making,’’ Expert Syst. Appl., vol. 126,
pp. 133–143, Jul. 2019, doi: 10.1016/j.eswa.2019.02.018.
REFERENCES [21] Y. H. Kim, S. C. Ahn, and W. H. Kwon, ‘‘Computational complexity
[1] P. P. Ray, ‘‘A survey on Internet of Things architectures,’’ J. King Saud of general fuzzy logic control and its simplification for a loop con-
Univ.-Comput. Inf. Sci., vol. 30, no. 3, pp. 291–319, Jul. 2018, doi: troller,’’ Fuzzy Sets Syst., vol. 111, no. 2, pp. 215–224, Apr. 2000, doi:
10.1016/j.jksuci.2016.10.003. 10.1016/S0165-0114(97)00409-0.
[2] E. Sisinni, A. Saifullah, S. Han, U. Jennehag, and M. Gidlund, ‘‘Indus- [22] M. Cuka, D. Elmazi, M. Ikeda, K. Matsuo, and L. Barolli, ‘‘IoT node selec-
trial Internet of Things: Challenges, opportunities, and directions,’’ IEEE tion in opportunistic networks: Implementation of fuzzy-based simulation
Trans. Ind. Informat., vol. 14, no. 11, pp. 4724–4734, Nov. 2018, doi: systems and testbed,’’ Internet Things, vol. 8, Dec. 2019, Art. no. 100105,
10.1109/TII.2018.2852491. doi: 10.1016/j.iot.2019.100105.
[23] R. S. Krishnan, E. G. Julie, Y. H. Robinson, S. Raja, R. Kumar, P. H. Thong,
[3] A. Zanella, N. Bui, A. Castellani, L. Vangelista, and M. Zorzi, ‘‘Internet of
and L. H. Son, ‘‘Fuzzy logic based smart irrigation system using Internet
Things for smart cities,’’ IEEE Internet Things J., vol. 1, no. 1, pp. 22–32,
of Things,’’ J. Cleaner Prod., vol. 252, Apr. 2020, Art. no. 119902, doi:
Feb. 2014, doi: 10.1109/JIOT.2014.2306328.
10.1016/j.jclepro.2019.119902.
[4] Statista.com. (2018). IoT: Number of Connected Devices Worldwide [24] A. Betzler, C. Gomez, I. Demirkol, and J. Paradells, ‘‘CoCoA+: An
2012-2025 | Statista Web. Accessed: Jul. 7, 2020. [Online]. Available: advanced congestion control mechanism for CoAP,’’ Ad Hoc Netw., vol. 33,
https://www.statista.com/statistics/471264/iot-number-of-connected- pp. 126–139, Oct. 2015, doi: 10.1016/j.adhoc.2015.04.007.
devices-worldwide/ [25] S. Bolettieri, G. Tanganelli, C. Vallati, and E. Mingozzi, ‘‘PCoCoA: A
[5] Intel.com. (2016). A Guide to the Internet of Things Infographic, Intel. precise congestion control algorithm for CoAP,’’ Ad Hoc Netw., vol. 80,
Accessed: Mar. 08, 2019. [Online]. Available: https://www.intel.com/ pp. 116–129, Nov. 2018, doi: 10.1016/j.adhoc.2018.06.015.
content/www/us/en/internet-of-things/infographics/guide-to-iot.html [26] V. Rathod, N. Jeppu, S. Sastry, S. Singala, and M. P. Tahiliani,
[6] J. Ren, H. Guo, C. Xu, and Y. Zhang, ‘‘Serving at the edge: A scalable IoT ‘‘CoCoA++: Delay gradient based congestion control for Internet
architecture based on transparent computing,’’ IEEE Netw., vol. 31, no. 5, of Things,’’ Future Gener. Comput. Syst., vol. 100, pp. 1053–1072,
pp. 96–105, Aug. 2017, doi: 10.1109/MNET.2017.1700030. Nov. 2019, doi: 10.1016/j.future.2019.04.054.
[7] C. Bormann, A. P. Castellani, and Z. Shelby, ‘‘CoAP: An application pro- [27] S. Gheisari and E. Tahavori, ‘‘CCCLA: A cognitive approach for
tocol for billions of tiny Internet nodes,’’ IEEE Internet Comput., vol. 16, congestion control in Internet of Things using a game of learning
no. 2, pp. 62–67, Mar. 2012, doi: 10.1109/MIC.2012.29. automata,’’ Comput. Commun., vol. 147, pp. 40–49, Nov. 2019, doi:
[8] J. Misic, M. Z. Ali, and V. B. Misic, ‘‘Architecture for IoT domain 10.1016/j.comcom.2019.08.017.
with CoAP observe feature,’’ IEEE Internet Things J., vol. 5, no. 2, [28] G. A. Akpakwu, G. P. Hancke, and A. M. Abu-Mahfouz, ‘‘CACC: Context-
pp. 1196–1205, Apr. 2018, doi: 10.1109/JIOT.2018.2800691. aware congestion control approach for lightweight CoAP/UDP-based
[9] RFC 7252—The Constrained Application Protocol (CoAP). Accessed: Internet of Things traffic,’’ Trans. Emerg. Telecommun. Technol., vol. 31,
Jul. 7, 2020. [Online]. Available: https://tools.ietf.org/html/rfc7252 no. 2, p. e3822, Feb. 2020, doi: 10.1002/ett.3822.
[10] M. Gohar, J.-G. Choi, and S.-J. Koh, ‘‘CoAP-based group mobility man- [29] D. A. Hayes and G. Armitage, Revisiting TCP congestion control using
agement protocol for the Internet-of-Things in WBAN environment,’’ delay gradients (Lecture Notes in Computer Science: Lecture Notes in
Future Gener. Comput. Syst., vol. 88, pp. 309–318, Nov. 2018, doi: Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 6641,
10.1016/j.future.2018.06.003. no. 2. Berlin, Germany: Springer, 2011, pp. 328–341, doi: 10.1007/978-
3-642-20798-3_25.
[11] C. Bockelmann, N. K. Pratas, G. Wunder, S. Saur, M. Navarro,
[30] B. W. Crowell, Y. Bock, and Z. Liu, ‘‘Single-station automated detection of
D. Gregoratti, G. Vivier, E. De Carvalho, Y. Ji, C. Stefanovic,
transient deformation in GPS time series with the relative strength index: A
P. Popovski, Q. Wang, M. Schellmann, E. Kosmatos, P. Demestichas,
case study of cascadian slow slip,’’ J. Geophys. Res., Solid Earth, vol. 121,
M. Raceala-Motoc, P. Jung, S. Stanczak, and A. Dekorsy, ‘‘Towards
no. 12, pp. 9077–9094, Dec. 2016, doi: 10.1002/2016JB013542.
massive connectivity support for scalable mMTC communications in
[31] Q. Zhang, H. Gong, X. Zhang, C. Liang, and Y.-A. Tan, ‘‘A sensitive
5G networks,’’ IEEE Access, vol. 6, pp. 28969–28992, May 2018, doi:
network jitter measurement for covert timing channels over interactive
10.1109/ACCESS.2018.2837382.
traffic,’’ Multimedia Tools Appl., vol. 78, no. 3, pp. 3493–3509, Feb. 2019,
[12] S. R. Pokhrel, J. Ding, J. Park, O.-S. Park, and J. Choi, ‘‘Towards enabling doi: 10.1007/s11042-018-6281-1.
critical mMTC: A review of URLLC within mMTC,’’ IEEE Access, vol. 8, [32] G. Cui, J. Guo, Y. Fan, Y. Lan, and X. Cheng, ‘‘Trend-smooth:
pp. 131796–131813, 2020, doi: 10.1109/ACCESS.2020.3010271. Accelerate asynchronous SGD by smoothing parameters using param-
[13] F. L. Coman, K. M. Malarski, M. N. Petersen, and S. Ruepp, ‘‘Security eter trends,’’ IEEE Access, vol. 7, pp. 156848–156859, 2019, doi:
issues in Internet of Things: Vulnerability analysis of LoRaWAN, sigfox 10.1109/ACCESS.2019.2949611.
and NB-IoT,’’ in Proc. Global IoT Summit (GIoTS), Aarhus, Denmark, [33] F. Osterlind, A. Dunkels, J. Eriksson, N. Finne, and T. Voigt, ‘‘Cross-
Jun. 2019, pp. 1–6, doi: 10.1109/GIOTS.2019.8766430. level sensor network simulation with COOJA,’’ in Proc. 31st IEEE Conf.
[14] A. Betzler, C. Gomez, I. Demirkol, and J. Paradells, ‘‘CoAP congestion Local Comput. Netw., Nov. 2006, pp. 641–648, doi: 10.1109/LCN.2006.
control for the Internet of Things,’’ IEEE Commun. Mag., vol. 54, no. 7, 322172.
pp. 154–160, Jul. 2016, doi: 10.1109/MCOM.2016.7509394. [34] A. Dunkels, B. Gronvall, and T. Voigt, ‘‘Contiki—A lightweight and
[15] J. J. Lee, K. T. Kim, and H. Y. Youn, ‘‘Enhancement of congestion flexible operating system for tiny networked sensors,’’ in Proc. 29th
control of constrained application protocol/congestion control/advanced Annu. IEEE Int. Conf. Local Comput. Netw., Nov. 2004, pp. 455–462, doi:
for Internet of Things environment,’’ Int. J. Distrib. Sensor Netw., vol. 12, 10.1109/LCN.2004.38.
no. 11, Nov. 2016, Art. no. 155014771667627, doi: 10.1177/155014771 [35] J. W. Eaton, D. Bateman, and S. Hauberg, ‘‘GNU Octave,’’ in Network
6676274. Thoery London. Bristol, U.K.: Network Theory, 1997.

58980 VOLUME 9, 2021

P. Aimtongkham et al.: Enhanced CoAP Scheme Using Fuzzy Logic

PHET AIMTONGKHAM received the B.S., M.S., CHAKCHAI SO-IN (Senior Member, IEEE)
and Ph.D. degrees in information technology received the Ph.D. degree in computer engineer-
from the Department of Computer Science, Khon ing from Washington University, St. Louis, MO,
Kaen University, Thailand, in 2013, 2017, and USA, in 2010. He was an Intern with CNAP-NTU
2020, respectively. He is currently a Postdoctoral (SG), Cisco Systems, WiMAX Forums, and the
Researcher with the ANT Laboratory, Khon Kaen Bell Labs, USA. He is currently a Professor with
University. His research interests include com- the Department of Computer Science, Khon Kaen
puter networking, multimedia networks, the Inter- University. He has authored over 100 international
net of Things, and machine learning. publications and ten books, including some in
IEEE JSAC, IEEE Magazines, and Computer Net-
work/Network Security Labs. His research interests include the Internet of
Things (IoT), mobile computing, wireless/sensor networks, signal process-
ing, and computer networking and security. He is also a senior member of
PARAMATE HORKAEW received the B.Eng.
ACM. He has served as an Editor for IEEE ACCESS, PLOS ONE, PeerJ, and
degree in telecommunication engineering from the
ECTI-Transactions on Computer and Information Technology (ECTI-CIT)
King Mongkut’s Institute of Technology, Ladkra-
and as a Committee Member for many conferences/journals such as GLOBE-
bang, in 1999. During his undergraduate study,
COM, ICC, VTC, WCNC, ICNP, ICNC, PIMRC, IEEE TRANSACTIONS, IEEE
he was working part-time on medical informatics
LETTER, IEEE Magazines, and Computer Networks/Communications.
research with the Computed Tomography Labora-
tory, NECTEC, from 1997 to 1999. As a Research
Assistant with the institute, he was involved in
both software development, notably CalScoreR,
and FPGA design projects. He had then continued
his research, supported by the Ministry of Science, in medical image comput-
ing with the Visual Information Processing Group, Imperial College London,
from 2000 to 2004. His Ph.D. thesis focused on an efficient and automatic
method for constructing the optimal statistical deformable model for com-
plex topological shapes with application to cardiovascular imaging. Based on
the p-Harmonic analysis on manifolds and information theory, the resultant
model is not only concise but also able to capture the intrinsic morphology
of typical human hearts. As a part of his research, in collaboration with
the Royal Brompton Hospital London, he also co-wrote a computer-assisted
diagnosis software for cardiovascular magnetic resonance images (CMR-
ToolsR), currently being clinically validated by several international research
centers. He is currently a Lecturer with the School of Computer Engineering,
Suranaree University of Technology. His main research interests include
computational anatomy, digital geometry processing, computer vision and
graphics, and the evolution of harmonic maps on riemannian surfaces with
applications to nonlinear PDE.

VOLUME 9, 2021 58981