IEEE INTERNET OF THINGS JOURNAL, VOL. 6, NO. 2, APRIL 2019
3743
Machine Learning-Based Bandwidth Prediction for
Low-Latency H2M Applications
Lihua Ruan , Graduate Student Member, IEEE, Maluge Pubuduni Imali Dias,
and Elaine Wong , Senior Member, IEEE
Abstract—Human-to-machine (H2M) communications in
emerging tactile-haptic applications are characterized by
stringent low-latency transmission. To achieve low-latency transmissions over existing optical and wireless access networks, this
paper proposes a machine learning-based predictive dynamic
bandwidth allocation (DBA) algorithm, termed MLP-DBA, to
address the uplink bandwidth contention and latency bottleneck
of such networks. The proposed algorithm utilizes an artificial
neural network (ANN) at the central office (CO) to predict H2M
packet bursts arriving at each optical network unit wireless access
point (ONU-AP), thereby enabling the uplink bandwidth demand
of each ONU-AP to be estimated. As such, arriving packet bursts
at the ONU-APs can be allocated bandwidth for tranmission by
the CO without having to wait to transmit in the following transmission cycles. Extensive simulations show that the ANN-based
prediction of H2M packet bursts achieves >90% accuracy, significantly improving bandwidth demand estimation over existing
prediction algorithms. MLP-DBA also makes adaptive bandwidth
allocation decisions by classifying each ONU-AP according to
its estimated bandwidth, with results showing reduced uplink
latency and packet drop ratio as compared to conventional
predictive DBA algorithms.
Index Terms—Artificial neural network (ANN), low latency,
machine learning, predictive bandwidth allocation.
I. I NTRODUCTION
HE emerging Internet-of-Things and Tactile Internet
paradigms are instigating the evolution of machine-tomachine communications toward real-time and remotely controlled human-to-machine (H2M) communications. Effective
H2M communications, such as responsive autonomous
systems, immersive edutainment, and proactive healthcare [1],
rely on low-latency (in milliseconds order) delivery of
H2M packets from/to human operators and the corresponding remotely-controlled devices/robots [2], [3]. An example
network architecture that supports H2M communications is
illustrated in Fig. 1 [1]–[4]. As illustrated, bidirectional transmission of H2M packets traverses the different segments of
the Internet including optical access networks, the integrated
optical network unit and wireless access points (ONU-APs),
T
Manuscript received September 3, 2018; revised October 9, 2018 and
November 10, 2018; accepted December 25, 2018. Date of publication
January 1, 2019; date of current version May 8, 2019. (Corresponding author:
Lihua Ruan.)
The authors are with the Electrical and Electronic Engineering
Department, University of Melbourne, Melbourne, VIC 3010,
Australia (e-mail: ruanl@student.unimelb.edu.au; diasm@unimelb.edu.au;
ewon@unimelb.edu.au).
Digital Object Identifier 10.1109/JIOT.2018.2890563
Fig. 1.
Network architecture that supports H2M communications.
and wireless local area networks (WLANs). Specifically, stateof-the-art literature has reported that stringent end-to-end
latency (within 1∼10 ms) must be adhered to in supporting remote and real-time control/actuation/haptic feedback
delivery over such heterogeneous wireless and optical access
networks [1]–[4]. To address this latency challenge, we have
presented in [4], [7], and [8], that strategically placing control servers at the central office (CO) closer to end users can
effectively expedite the delivery of H2M packets, and consequently the overall response and actuation of devices/robots.
In this regard, bandwidth resource allocation decision made
by the CO for delivering H2M traffic over access networks is
critical to improve the latency performance.
Similar to the burstiness of Internet traffic [12], experimental
studies in [9]–[11] have revealed that H2M traffic generated in
H2M applications are also bursty with alternate ON and OFF
intervals. For example, during an ON interval, consecutive packets generated by end devices/robots are wirelessly transmitted
through the WLANs and aggregated at the ONU-APs, whereas
no packet is generated in the OFF interval. While downlink
transmissions in Fig. 1 are typically broadcast, the latency
bottleneck mainly lies in the uplink transmission due to bandwidth contention amongst multiple ONU-APs. Specifically,
uplink bandwidth allocation is mainly achieved by implementing dynamic bandwidth allocation (DBA) algorithms. In
a typical DBA algorithm, an ONU-AP demands bandwidth
by appending a REPORT message to its packet transmissions.
Accordingly, the CO grants bandwidth to the ONU-AP through
a GATE message in a round-robin manner [12]. The time interval
between consecutive transmissions of an ONU-AP is known
as a polling cycle. The uplink latency focused in this paper
hence refers to the time upon a packet arriving at an ONU-AP
until it reaches CO. Classic DBA algorithms including fixedservice DBA, exhaustive-service DBA, limited-service DBA,
c 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
2327-4662 See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Microchip Tech Inc. Downloaded on January 24,2022 at 13:12:05 UTC from IEEE Xplore. Restrictions apply.
3744
IEEE INTERNET OF THINGS JOURNAL, VOL. 6, NO. 2, APRIL 2019
and excessive-service DBA have been comprehensively studied and compared in [12]–[14]. Results show that these DBAs
exhibit high uplink latency when traffic is bursty. This is partly
attributed to the fact that arriving packets at the ONU-AP
need to be reported first by the ONU-AP and then subsequently granted bandwidth for transmission in the next polling
cycle. Moreover, packets buffered at highly loaded ONU-APs
may need to wait for multiple cycles before bandwidth is
granted for their transmission. Due to the marginal difference
in uplink latency performance between the above classic DBAs,
the limited-service DBA algorithm, an algorithm that prevents
bandwidth monopolization by heavily-loaded ONU-APs, is
commonly adopted as the baseline DBA [12].
On the other hand, predictive DBAs typically provide additional bandwidth to each ONU-AP on top of the requested
amount. This additional bandwidth, termed BWpred in this
paper, is typically estimated by predicting the amount of arriving packets (in bytes) at each ONU-AP during each polling
cycle [15]. Predictive DBAs have the potential to reduce
uplink latency as additional packets can be transmitted in
the same polling cycle without their bandwidth having to
be reported by the ONU-APs. However, when it comes to
predicting the bandwidth of bursty traffic, this remains an open
challenge. The effectiveness of a predictive DBA is dependent on its prediction accuracy. Since accurately predicting
the bandwidth of bursty traffic will reduce uplink latency [14],
it is imperative that the bandwidth prediction of bursty traffic is investigated in order to reduce the uplink latency of
networks that support H2M traffic. Further, the bandwidth
allocated by the CO should be in accordance with the uplink
traffic characteristics and load. For highly bursty H2M traffic, the ability of DBAs in allocating bandwidth during bursty
intervals and preserving bandwidth during long-idle intervals
is key to reducing the uplink latency [9].
In light of the above, in this paper, we exploit the
use of an artificial neural network (ANN) at the CO to
predict the ON and OFF status of bursty H2M traffic arriving at each ONU-AP. Prediction of the ON/OFF status
consequently enables the evaluation of the uplink bandwidth demand of each ONU-AP. In justifying prediction
performance, we present a comparative study of ANN versus
existing bursty traffic prediction methods used in predictive
DBAs. Through simulations, we first show that our trained
ANN can achieve superior accuracy (>90%) in predicting
the ON and OFF status of ONU-APs, and therefore, effectively improving the accuracy of the bandwidth demand
estimation. The proposed ANN prediction is then implemented at the CO to facilitate bandwidth allocation decisions through the proposed machine learning-based predictive
DBA (MLP-DBA). In contrast to conventional DBA/predictive
DBA algorithms that grant requested/estimated bandwidth
to all ONU-APs, MLP-DBA classifies each ONU-AP based
on: 1) the predicted ON and OFF status of ONU-APs
and 2) the estimated bandwidth demand. The CO then
adaptively allocates bandwidth to each ONU-AP. In this
paper, we show that with ANN prediction, the proposed
MLP-DBA successfully improves uplink latency performance
and reduces the packet drop ratio as compared to
the baseline limited-service DBA and existing predictive
DBA algorithms.
The rest of this paper is organized as follows. In
Section II, we introduce state-of-the-art predictive DBA algorithms and the motivations behind applying ANN in
a DBA. Our ANN-based ON/OFF status prediction is proposed
in Section III. Section IV details the proposed MLP-DBA algorithm with analytical and simulation results and discussions
presented in Section V. A summary of the main contributions
of this paper is presented in Section VI.
II. E XISTING P REDICTIVE DBA A LGORITHMS
Early predictive DBA algorithms such as constant and linear
credit DBAs typically allocate a constant BWpred to ONU-APs
in each polling cycle irrespective of traffic load. Hence, these
algorithms, frequently termed as static predictive DBAs, provide only marginal latency improvements over nonpredictive
DBAs if the considered traffic is bursty [12]. By comparison,
statistical predictive DBAs reported in [16]–[18] use arithmetic
averaging, exponential smoothing, and Bayesian estimation to
evaluate BWpred by estimating the packet interarrival time.
In terms of bursty traffic, the latency performance of these
DBAs in [16]–[18] were shown to fluctuate due to their limitation in predicting instantaneous traffic variations [4]. On
the other hand, the predictive algorithm in [19], termed mostrecently (MR) DBA, uses the MR received packets from the
previous polling cycle to estimate BWpred . Uplink latency
was shown to be reduced as compared to the limited-service
DBA. Further, the limited sharing DBA with traffic prediction
(LSTP-DBA) [20] uses a 4-order autoregressive model, i.e.,
counting on the arrival packets in the previous four polling
cycles, to estimate BWpred . Simulation results show that
LSTP-DBA outperforms classic DBA algorithms regarding
the latency performance. Yet, the prediction accuracy, a critical criterion for evaluating the prediction performance [14],
in MR- and LSTP-DBAs, remains to be investigated. Most
recently in [21], a machine learning technique, k-nearest
neighborhood (kNN), was used to evaluate BWpred for video
streams. The proposed algorithm, named data mining forecasting DBA, estimated BWpred by averaging the received packets
in k number of past polling cycles that have similar durations
to the current cycle. We verified in a preliminary study [22]
that even though kNN-based prediction was effective in
predicting constant video flows, it was less effective in bursty
scenarios, and results in large prediction errors. The prediction
performance of the algorithms mentioned above depends on
the applications supported and also the selection of traffic features used in making the prediction, e.g., interarrival time,
received packets, or polling cycle durations, etc. As such, the
prediction accuracy for busty traffic is still to be investigated.
Compared to the predictive DBAs described above that
directly predict the value of BWpred using past received packets, in this paper, we propose to exploit an ANN located
at the CO that is able to input multiple traffic features to
predict the ON/OFF status of each ONU-AP, such that BWpred
of the ONU-AP can be estimated with high accuracy for bursty
traffic. Then, using MLP-DBA, the CO adaptively classifies
Authorized licensed use limited to: Microchip Tech Inc. Downloaded on January 24,2022 at 13:12:05 UTC from IEEE Xplore. Restrictions apply.
RUAN et al.: MACHINE LEARNING-BASED BANDWIDTH PREDICTION FOR LOW-LATENCY H2M APPLICATIONS
3745
Fig. 2. Operational traffic flow diagram of a typical predictive DBA operation
algorithm.
ONU-APs and allocates bandwidth accordingly. Note that our
preliminary work in [22] presents the advantages of using an
ANN in bandwidth prediction and allocation over the existing
use of kNN for H2M applications. This paper on the other
hand, details the proposed ANN prediction model, supervised
training process, prediction performances, and comprehensive validates the effectiveness of the proposed MLP-DBA in
addressing the challenges in conventional limited-service
DBA, MR-DBA, and LSTP-DBA algorithms.
III. A RTIFICIAL N EURAL N ETWORK -BASED P REDICTION
A. Traffic Flow in Predictive DBA
Fig. 2 illustrates the operational flow diagram of a typical predictive DBA. Let us denote Tpoll (i, j) as the duration
of the ith polling cycle of ONU-AP j. As illustrated, in the
ith polling cycle, ONU-AP j transmits k(i, j) bytes, or equivalently n(i, j) number of packets, in the granted transmission
duration and reports the remaining queue length, denoted
as BWreq (i, j), in the REPORT message to the CO. Hence,
BWreq (i, j) is the requested bandwidth by ONU-AP j. The CO
then allocates bandwidth, BWdem (i+1, j), by sending a GATE
message that indicates the next transmission start time and
duration of ONU-AP j. The duration of the (i+1)th polling
cycle, Tpoll (i + 1, j), is updated at the CO. Upon receiving
the GATE message from the CO, ONU-AP j waits for the
next transmission start time which it then transmits packets
for a duration equivalent to BWdem (i + 1, j).
As shown in Fig. 2, while waiting for the transmission start
time in the (i + 1)th polling cycle, an amount of a(i + 1, j)
bytes is received by ONU-AP j for uplink transmission. To
reduce uplink latency, the CO predicts a(i + 1, j), or the
equivalent bandwidth BWpred (i + 1, j). The total bandwidth,
BWdem (i + 1, j), that is allocated by the CO to ONU-AP j,
should therefore be calculated as follows:
BWdem (i + 1, j) = BWreq (i, j) + BWpred (i + 1, j).
(1)
With BWpred (i+1, j) allocated, a(i+1, j) bytes arriving during
the (i+1)th polling cycle can be transmitted within the polling
cycle without needing to be reported to the CO, thereby reducing latency. However, the exact a(i+1, j) is not known a priori
and so accurately predicting a(i + 1, j), i.e., BWpred , is critical
to improving uplink latency performance. Let us now denote
the prediction error as follows:
(2)
e(i + 1, j) = BWpred (i + 1, j) − a(i + 1, j).
Fig. 3. Schematic of the proposed ANN architecture to predict the ONU/OFF
status (y) of the bursty traffic at ONU-AP j.
Note that as how long a packet will wait at the ONU-APs
before transmission depends on DBA operations, investigation
into bandwidth allocations for packets aggregated at ONU-APs
is our focus in the current work.
In a heterogeneous access networks, bursty network traffic
characterized by ON/OFF intervals has been rigorously studied in [26]–[28]. In H2M experiments, human operations have
also shown similar ON/OFF packet bursts [9]. In this paper,
we adopt the typical ON/OFF traffic model [26] widely used
to simulate the aggregations of H2M traffic. The sensitivity
of the proposed ANN prediction to different levels of burstiness and traffic model is evaluated anticipating the emerging
applications.
B. Artificial Neural Network
As introduced earlier, when considering bursty traffic, it is
critical for the CO to identify if an ONU-AP is currently in an
ON interval, i.e., the ONU-AP is receiving packets aggregated
from its WLAN, or otherwise in an OFF interval, i.e., an idle
interval with no arriving packet bursts. Erroneously allocating additional bandwidth to ONU-APs that are in their OFF
intervals increases polling cycle time, and thereby increases
the uplink latency of the packets at those ONU-APs.
Hence, we use an ANN to predict the ON/OFF status of
ONU-AP j and estimate BWpred (i + 1, j) from this prediction.
Fig. 3 illustrates a schematic of our proposed ANN comprising an input layer, an output layer, and some hidden layers in
between. Each layer is composed of neuron units, which are
nonlinear functions that map the associated inputs to an output. Specifically, when receiving the REPORT from ONU-AP
j in the ith polling cycle, the CO uses available traffic features, as listed in Table I to predict the ON/OFF burst status
of ONU-AP j in the (i + 1)th polling cycle. With x(i, j) =
{k(i, j), n(i, j), a(i, j), BWreq (i, j), Tpoll (i, j), Tpoll (i + 1, j)} as
the input traffic features to the ANN in the ith polling cycle,
the target output y is denoted in (1) as follows:
1, ONU-AP j is ON in (i + 1)-th polling cycle
y(i + 1, j) =
0, ONU-AP j is OFF in (i + 1)-th polling cycle.
(3)
Authorized licensed use limited to: Microchip Tech Inc. Downloaded on January 24,2022 at 13:12:05 UTC from IEEE Xplore. Restrictions apply.
3746
IEEE INTERNET OF THINGS JOURNAL, VOL. 6, NO. 2, APRIL 2019
TABLE I
T RAFFIC F EATURES C OLLECTED IN THE ith P OLLING C YCLE
An ANN learns by iteratively adjusting its weights and bias
associated with the neurons to a yield desired output. It has
a proven capability in addressing curve fitting, optimization,
pattern recognition, and classification problems [23]. Here,
based on six input traffic features, our proposed ANN devolves
into the classification of an ON (y = 1) or OFF (y = 0) status. Recall that BWpred (i + 1, j) is the estimated bandwidth
of a(i + 1, j) at ONU-AP j. Considering lastmile compatibility [20], [24] and the ON/OFF status y(i + 1, j) predicted by
the ANN, BWpred (i + 1, j) can be estimated using
BWpred (i + 1, j) ≈ y(i + 1, j)Tpoll (i + 1, j)RWLAN .
(4)
In (2), RWLAN is the data rate of the WLAN associated with
ONU-AP j. Once the ON/OFF status y(i + 1, j) is predicted,
and the BWpred (i, j) is estimated, the CO evaluates the total
bandwidth demand, BWdem (i + 1, j), of ONU-AP j in (i + 1)th
polling cycle as follows:
BWdem (i + 1, j) = BWreq (i, j) + y(i + 1, j)Tpoll (i + 1, j)RWLAN .
(5)
C. Supervised Training
Prior to using the ANN to predict the ON/OFF status of
ONU-APs, we must first train the ANN using training sets
collected during the training phase of network operation.
A training set, S, is represented as {(x(i, j), y(i + 1, j))| for i =
1, 2 . . . , N, j = 1, 2 . . . , M} with x(i, j) as the traffic features
collected by the CO in the ith polling cycle and y(i+1, j) as the
labeled output. The parameter N is the total number of polling
cycles, and M is the number of ONU-APs. Here, we generate
training sets using event-driven packet-level simulations of 16and 32-ONU-AP optical access networks. The data rate of the
optical access network and WLANs are 1 Gb/s and 100 Mb/s,
respectively. Traffic features from 250 000 polling cycles under
various network loads at each ONU-AP are collected and then
is utilized for supervised training. The Hurst parameter in
the traffic model is a primary indicator for self-similarity and
long-range dependence (LRD) of network traffic aggregated
from diverse applications [26]. A Hurst parameter of 0.8, i.e.,
H = 0.8, commonly utilized by existing traffic profiles is chosen here to simulate packet bursts in the WLANs [12]. The
sensitivity of the trained ANN is presented in the following
section, considering a variation in Hurst parameter as well as
an exponential traffic model [12] for comparison.
After supervised training with the collected training set, the
final architecture of the ANN is shown in Fig. 3, whereby
in addition to the input and output layers, there is one hidden layer of ten neurons. The optimal weight matrix for each
(a)
(b)
Fig. 4.
ON / OFF status prediction performance. (a) Prediction accuracy.
(b) ROC (traffic load = 0.6).
layer is determined using Gradient descent method [25]. Since
prediction accuracy does not improve with additional layers
and neurons, this architecture is used as a final architecture in
subsequent investigations.
Note that once the training phase is complete, the CO only
needs to store the weight matrix to map the input traffic
features to output the predicted ON/OFF status (3), estimate
BWpred (4), and evaluate (5). To verify the predict performance
of the trained ANN, we use a test set, T, with another 250 000cycle traffic features, to investigate the accuracy of its ON/OFF
status prediction. Fig. 4(a) plots the ON/OFF prediction accuracy for the 16 and 32 ONU-AP networks as a function of
aggregated network load, highlighting that the trained ANN
achieves high prediction accuracy for network loads from
0.1 to 1. The prediction accuracy is >96% under light and high
network loads due to the ONU-APs having longer idle (OFF)
and bursty (ON) intervals, respectively. By comparison, the
prediction accuracy decreases under moderate traffic load, i.e.,
0.4∼0.7. This behavior can be attributed to the frequent alternation of ON and OFF intervals. To investigate further, we plot
the receiver operating characteristic (ROC) curves at a network
load of 0.6 [Fig. 4(b)]. Here, the true positive rate is the percentage of correct ON predictions in all ON samples, and the
false positive rate is the percentage of mispredicting OFF over
all OFF samples. Note that the curves approach (0,1), indicating that the trained ANN achieves effective prediction for
both ON and OFF status. Also, the trained ANN achieves better prediction performance for 32-ONU-AP network, as also
indicated in Fig. 4(a). Considering the same network load,
each individual ONU-AP in the 32-ONU-AP network contributes to a lower network load than an individual ONU-AP
in a 16-ONU-AP network.
Authorized licensed use limited to: Microchip Tech Inc. Downloaded on January 24,2022 at 13:12:05 UTC from IEEE Xplore. Restrictions apply.
RUAN et al.: MACHINE LEARNING-BASED BANDWIDTH PREDICTION FOR LOW-LATENCY H2M APPLICATIONS
3747
(a)
(a)
(b)
(b)
Fig. 5. Evaluation of sensitivity to training set size and traffic model variation.
(a) Training set size variation. (b) Traffic model variation.
Fig. 6. Prediction performance comparison with MR-DBA and LSTP-DBA.
(a) 16 ONU-AP. (b) 32 ONU-AP.
In Fig. 5, we investigate the dependency of the ON/OFF
prediction accuracy on the training set size, Hurst parameter
variation in the Pareto traffic model, and an exponential traffic model for a 32-ONU-AP. In Fig. 5(a), results for various
network loads at 0.3, 0.6, and 0.9 are shown for illustrative
purposes. As can be observed, increasing the training set size
improves prediction accuracy. Further, the prediction accuracy varies under different network traffic loads. As shown
in Fig. 5(a), network loads of 0.3 and 0.9 require a smaller
training set to reach >90% accuracy whereas a larger training set is required for the network load of 0.6 to achieve the
same prediction accuracy. For the network loads considered,
prediction accuracy is low when the training set is small due
to an inadequate number and imbalance of ON/OFF samples.
Specifically, as each time of packet burst at each ONU-AP
differentiates significantly, a small training set only containing
features of a few ON/OFF intervals cannot be well generalized
to predict a test set. When samples are enough to include features from enough ON/OFF alternations at all 32 ONU-APs,
a clear raise in prediction accuracy can be observed. Results
show that overall, across the three network loads, a training set
size with around 104 samples is sufficient to train the ANN
whereby any further increase in training set size would not
yield further improvement in prediction accuracy.
Fig. 5(b) shows the sensitivity of the trained ANN. The
trained ANN is first applied to predict the ON/OFF status
of ONU-APs when the Hurst parameter varies from 0.6
to 0.9. Note that 0.5 < H < 1 describes the long-term
auto-correlations, i.e., LRD, of the aggregated traffic [29].
A larger H represents stronger LRD and traffic burstiness,
indicating that consecutive long ON intervals are more likely
to occur. Results in Fig. 5(b) shows that the trained ANN
achieves >95% prediction accuracy regardless of H variations
in traffic. In comparison with LRD traffic, the trained ANN
is also applied to traffic characterized by short-range dependence (SRD) with exponential ON/OFF intervals. As can be
viewed in Fig. 5(b), the prediction accuracy (green bar)
reduces, but >90% prediction accuracy can still be achieved.
Further, when retraining the ANN corresponding to the SRD
traffic, the prediction accuracy (yellow bar) is effectively
improved to 98%. Results in Fig. 5(b) validate the robustness
of the trained ANN.
D. Bandwidth Prediction
With the trained ANN predicting the ON/OFF status of all
ONU-APs using (3), and estimating BWpred using (4), we
evaluate the average prediction error using
Eavg =
M N
M N
1 1 e(i, j) =
BWpred (i, j) − a(i, j).
NM
NM
j=1 i=1
j=1 i=1
(6)
Fig. 6(a) and (b) plots Eavg for a 16- and 32-ONU-AP
network, respectively, and compares the results obtained using
our proposed ANN-based prediction method to that obtained
using the MR-DBA and LSTP-DBA. Fig. 6 shows that Eavg
increases with traffic load and that our proposed ANN-based
prediction has the best performance, especially under heavy
traffic loads. Eavg of the 32-ONU-AP network is about half
of that of the 16-ONU-AP network. Again, this is attributed
to the lower traffic load per individual ONU-APs in 32-ONUAP network when considering the same aggregated network
load.
Authorized licensed use limited to: Microchip Tech Inc. Downloaded on January 24,2022 at 13:12:05 UTC from IEEE Xplore. Restrictions apply.
3748
IEEE INTERNET OF THINGS JOURNAL, VOL. 6, NO. 2, APRIL 2019
TABLE II
P RINCIPLE OF ONU-AP C LASSIFICATION AND MLP-DBA BANDWIDTH A LLOCATION
min
Fig. 7.
Illustration of MLP-DBA procedure with three ONU-APs.
IV. MLP-DBA A LGORITHM
The CO in an optical access network will typically allocate
a bandwidth equivalent to min{BWdem , BWmax } to ONU-APs
during each polling cycle. The parameter BWmax is the
maximum bandwidth that can be allocated to an individual ONU-AP and BWdem = BWreq + BWpred as given
in (1). When ONU-APs are heavily loaded with long accumulated packet queues, BWreq exceeds BWmax . For heavily
loaded ONU-APs, reducing the time interval between consecutive transmissions and increasing the allocated bandwidth
would reduce the uplink latency of the arriving packets,
but this must be done without impacting the less heavily
loaded ONU-APs. In order to expedite transmission of arriving packets, MLP-DBA enables the CO to classify ONU-APs
based to their BWdem , and allocate bandwidth accordingly, as
described below.
Upon receiving a REPORT message from an ONU-AP, the
CO predicts the ON/OFF status of the ONU-AP, i.e., y using the
trained ANN, and estimates BWpred and BWdem as introduced
in Section III. Then, the CO classifies each ONU-AP into one
of 3 different classes following the principle listed in Table II.
1) If according to the train ANN, the status of an ONU-AP
is ON in the next polling cycle, i.e., y = 1, and meanwhile BWdem < BWmax , the ONU-AP will be classified
as Class A. For such an ONU-AP, the CO will allocate through a GATE message a bandwidth equivalent to
BWdem = BWpred + BWreq for transmission in the next
polling cycle, allowing arriving packets to be transmitted without having to be reported to the CO and without
having to wait for the next transmission time.
2) If the status of the ONU-AP is predicted to be OFF,
i.e., y = 0, with BWdem < BWmax , it will be classified
as Class B. The CO will allocate a Class-B ONU-AP
a bandwidth that is equivalent to BWreq since no packet
is expected to arrive during its OFF interval.
3) Heavily loaded ONU-AP with a BWdem > BWmax is
classified as Class C. For such an ONU-AP, the CO will
first allocate BWmax through a GATE message. Further,
to alleviate the number of queued packets, an additional
GATE message is generated for the ONU-AP by the CO
such that an additional transmission opportunity is given
to the ONU-AP before scheduling the next polling cycle.
To explain the operation of MLP-DBA for Class-A, B, and
C ONU-APs more clearly, we present the MLP-DBA operation in Fig. 7 with 3 ONU-APs for illustrative purposes. As
shown in Fig. 7, in the (i − 1)th polling cycle, ONU-AP 1, 2,
and 3 are classified as Class A, C, and B, respectively. As such,
in responding to their REPORTs, the CO first grants BWdem
to ONU-AP 1, 2, and 3 in the ith polling cycle, respectively.
Moreover, since ONU-AP 2 is classified as Class C, at the
beginning of the ith polling cycle, an additional GATE message
is generated for ONU-AP 2 by the CO such that an additional transmission opportunity is given to ONU-AP 2 before
scheduling the (i + 1)th polling cycle. As such, the transmission interval for ONU-AP 2 is reduced as presented in the
figure. As ONU-AP 3 is OFF with no packet arrivals expected,
the additional transmission opportunity interval does not affect
ONU-AP 3. Note that additional GATEs are generated and
send at the beginning of each polling cycle based on the
classification results in the previous cycle. As such, the CO
does not respond to the REPORTs received during additional
transmission opportunity intervals. The CO repeats this procedure during each polling cycle for all ONU-APs. It is worth
noting that MLP-DBA algorithm considers bursty traffic pattern and reduces latency by adaptively allocating bandwidth to
ONU-APs in ON intervals. As such, it can be applied to bursty
Authorized licensed use limited to: Microchip Tech Inc. Downloaded on January 24,2022 at 13:12:05 UTC from IEEE Xplore. Restrictions apply.
RUAN et al.: MACHINE LEARNING-BASED BANDWIDTH PREDICTION FOR LOW-LATENCY H2M APPLICATIONS
Algorithm 1 MLP-DBA:
Allocation at the CO
Pseudocode
of
Bandwidth
tscheduled — time up to which the uplink channel has been scheduled
RRT — round-trip transmission time
T g — guard time
T process — time for processing a GATE message at an ONU-AP
T REPORT — duration of a REPORT message
δ(i, j) ∈ {A, B, C} — class of ONU-AP j in ith polling cycle
tstart (i, j)/tadd_start (i, j) — the start time of granted timeslot for ONU-AP j
in the ith polling cycle
tend (i, j)/tadd_end (i, j) — the end time of granted timeslot for ONU-AP j in
the ith polling cycle
RPON — data rate in the optical network
RWLAN — data rate in the WLAN
BWadd_grant (i, j) — the granted bandwidth for ONU-AP j in the additional
opportunity if an additional GATE is generated for ONU- AP j in the ith
polling cycle
ANN(x) — trained neural network with input feature vector x
The operation for ONU-AP j at the CO at the beginning of ith polling
cycle (repeat for all i and j):
1 {
2 if δ(i, j) = C
3 // schedule the start time in the ith additional cycle
4
tadd_start (i, j) = min{LocalTime + RRT/2 + Tprocess ,
5
tscheduled – RRT/2 – T process };
6
BWadd_grant (i, j) = BWmax ;
7
tadd_end (i + 1, j) = tadd_start (i, j) + BWmax /RPON +
8
T REPORT + Tguard ;
9
generate and send a GATE indicating tstart (i + 1, j)
10
and BWadd_grant (i, j);
11
tscheduled = tadd_end (i, j); // update scheduled time
12 else
13
BWadd_grant (i, j) = 0;
14 end
15 }
The operation for ONU-AP j at the CO upon receiving the REPORT in
the ith polling cycle (repeat for all i and j):
16 {
17
// schedule the start time in the (i + 1)th polling cycle
18
tstart (i + 1, j) = min{LocalTime + RRT/2 + Tprocess , tscheduled
19
−RRT/2 − Tprocess };
20 //ANN-based prediction and classification
21 xi , j = {k(i, j), n(i, j), a(i, j), BWreq (i, j), TPOLL (i, j),
22 TPOLL (i + 1, j)};
23 y = ANN(xi , j);
24 BWpred (i + 1, j) = yRWLAN tstart (i + 1, j) − yRWLAN tstart (i,
25 j)tstart (i, j);
26 update δ(i + 1, j);
27 // following the limited-service principle, the granted BW
28 is bounded by BWmax
29 BWgrant (i + 1, j) = min{BWreq (i, j) + BWpred (i + 1, j)−
30
BWadd_grant (i, j), BWmax };
31 // update time schedule and generate a GATE message
32 tend (i + 1, j) = tstart (i + 1, j) + BWgrant (i + 1, j)/RPON +
33
TREPORT + Tguard ;
34 generate and send a GATE indicating tstart (i + 1, j) and
35 BWgrant (i + 1, j);
36 tscheduled = tend (i + 1, j); // update scheduled time
37 }
The operation for ONU-AP j at the CO upon receiving the REPORT in
the ith additional opportunity (repeat for all i and j):
38 {
39
xi,j = {k(i, j), n(i, j), a(i, j), BWreq (i, j), TPOLL (i, j),
40
TPOLL (i + 1, j)}; // note that features are collected from
41
the additional cycle
42
y = ANN(xi,j );
43
update δ(i + 1, j);
44 }
traffic with different levels of burstiness such as that studied
in Fig. 5(b). The pseudocode of the MLP-DBA algorithm is
presented in Algorithm 1.
3749
TABLE III
N ETWORK AND P ROTOCOL PARAMETERS
V. P ERFORMANCE E VALUATION
To verify the effectiveness of MLP-DBA, we implemented
packet-level simulations (MATLAB) of an optical access
network which parameters are listed in Table III. The bursty
ON and OFF intervals follows Pareto distribution with a Hurst
parameter of 0.8. We analyze the ONU-AP classifications during MLP-DBA operation and compare the uplink latency and
packet drop ratio performances of MLP-DBA with/without
ONU-AP classification, MR-DBA, LSTP-DBA, and the baseline limited-service DBA algorithms. In the case of MLPDBA without classification, BWdem is estimated using the
proposed ANN and (3–5) but bandwidth allocation follows the
conventional DBA operation whereby min{BWdem , BWmax } is
allocated to all ONU-APs. In contrast, MLP-DBA with classification utilizes the ANN to predict the ON/OFF status and
estimate BWdem of each ONU-AP and based on these results,
classify and allocate bandwidth to each ONU-AP accordingly.
A. Uplink Latency and Packet Drop Ratio
Fig. 8(a) and (b) verify the effectiveness of MLP-DBA in
predicting bandwidth demand and improving uplink latency
performance. The uplink latency is defined as the duration between the time a packet arrives at a buffer of an
ONU-AP until its uplink transmission. Latency constraints,
Dst = 1 and 10 ms, are considered in supporting H2M
interactions [1], [2]. Several observations can be highlighted.
First of all, compared to the limited-service DBA, MR-DBA,
LSTP-DBA, and the proposed MLP-DBA without classification, the uplink latency performance of the MLP-DBA with
classification is improved under light-to-moderate network
loads, i.e., 0.1∼0.5 in Fig. 8(a) and 0.1∼0.4 in Fig. 8(b).
This is attributed to the high accuracy of ON/OFF status
prediction, BWpred estimation and the subsequent allocation
of BWdem = BWpred + BWreq to ensure that arriving packets
can be transmitted without having to wait for the next transmission cycle. MLP-DBA without classification achieve better
performance compared to MR-DBA and LSTP-DBA due to its
superior ON/OFF status prediction performance of the ANN
as compared to the most-recent and the 4-order autoregressive models, respectively. However, MR-DBA, LSTP-DBA,
and MLP-DBA without classification result in high uplink
latency similar to that of the limited-service DBA algorithm when network load is beyond 0.5 in Fig. 8(a) and (b).
This is because under heavy network loads, packet bursts
at ONU-APs tend to have shorter OFF intervals, leading to
Authorized licensed use limited to: Microchip Tech Inc. Downloaded on January 24,2022 at 13:12:05 UTC from IEEE Xplore. Restrictions apply.
3750
IEEE INTERNET OF THINGS JOURNAL, VOL. 6, NO. 2, APRIL 2019
(a)
(b)
Fig. 8.
Average uplink latency with the proposed MLP-DBA.
(a) 16 ONU-APs. (b) 32 ONU-APs.
accumulated packet queues at ONU-APs. Consequently, conventional predictive DBAs that allocates min{BWdem , BWmax }
to ONU-APs are less effective when BWmax is allocated in
each polling cycle and uplink latency is mainly attributed
to multiple-cycle waiting time before a packet is allocated
bandwidth for transmission. Adaptively allocating bandwidth
to ONU-APs based on the proposed classification addresses
this latency bottleneck. As shown in Fig. 8(a) and (b), implementing MLP-DBA with ONU-AP classification effectively
reduces the uplink latency even under heavy network loads.
This is because bandwidth is allocated to heavily loaded Class
C ONU-APs, by adding additional transmission opportunities
in each polling cycle. In comparison, for network loads below
0.5 in the 16-ONU-AP network and 0.3 in the 32-ONU-AP
network, MLP-DBA with and without classification exhibit
similar latency performance since Class-A ONU-APs form the
majority type of ONU-APs. Overall, considering H2M applications, MLP-DBA is shown to support approximately 10%
more network traffic than conventional DBA algorithms under
Dst = 1 and 10 ms. As presented in Fig. 8, with MLP-DBA,
a 16-ONU-AP network can support a network traffic load of
up to 0.6 and 0.8 under Dst = 1 ms and Dst = 10 ms, respectively, and a 32-ONU-AP network can support a network load
of up to 0.4 for Dst = 1 ms and 0.6 for Dst = 10 ms.
Fig. 9 plots the packet drop ratio under limited-service
DBA, MR-DBA, LSTP-DBA, and MLP-DBA with/without
Fig. 9. Distribution of Class A, B, and C ONU-APs. (a) 16 ONU-APs.
(b) 32 ONU-APs.
classification. As shown in the figures, implementations of
MR-DBA, LSTP-DBA, and MLP-DBA without classification help alleviate packet drops under heavy network loads.
However, the improvement shown is marginal with < 5%
compared to limited-service DBA in both networks. For a 32ONU-AP network, the packet drop ratio is about twice that
in a 16-ONU-AP network. This is attributed to the fact that
under the same maximum polling cycle duration of 1 ms, the
BWmax of a 32-ONU-AP network is half of that for a 16-ONUAP network. In comparison, the proposed MLP-DBA based
on ONU-AP classification shows significant improvement in
reducing packet drops. Overall, as illustrated in the figures, the
packet drop ratio for MLP-DBA with classification is reduced
by 10% as compared to the baseline limited-service DBA at
high traffic loads. This is due to the reduced polling cycle
durations for heavily loaded Class C ONU-APs by introducing
additional transmission opportunities. Since MLP-DBA can
predict and classify ONU-APs that have long packet queues,
GATE messages are generated prior to receiving REPORT
messages from these ONU-APs. As a result, the polling
cycle duration of these ONU-APs is reduced, and additional
bandwidth is allocated, thereby improving both latency and
packet drop performances. The performance of the proposed
MLP-DBA can be further explained by viewing the varying distribution of Class-A, B, and C ONU-APs in the two
networks in Fig. 10.
Authorized licensed use limited to: Microchip Tech Inc. Downloaded on January 24,2022 at 13:12:05 UTC from IEEE Xplore. Restrictions apply.
RUAN et al.: MACHINE LEARNING-BASED BANDWIDTH PREDICTION FOR LOW-LATENCY H2M APPLICATIONS
3751
in the 32-ONU-AP network, when network load increases,
more ONU-APs are classified as Class C in the 32-ONU-AP
network. Overall, results in Fig. 10 verify the capability of
the proposed MLP-DBA in making adaptive bandwidth allocation decisions that reduces uplink latency and packet drops.
Concluding, simulations presented in Figs. 8–10 in this section highlight the importance of classification and adaptive
bandwidth allocation decisions.
VI. C ONCLUSION
(a)
(b)
Fig. 10. Distribution of Class A, B, and C ONU-APs. (a) 16 ONU-APs.
(b) 32 ONU-APs.
To improve the uplink latency performance for H2M traffic
over access networks, we proposed MLP-DBA to adaptively allocate bandwidth to classified ONU-APs based on
ANN-enabled bursty ON/OFF status prediction and bandwidth
estimation. Specifically, we first presented our exploitation
of an ANN at the CO to predict the ON/OFF status of
ONU-APs in each polling cycle. We showed that with supervised training, the proposed ANN achieved >90% accuracy in
identifying the ON/OFF status of ONU-APs, thereby yielding
superior ONU-AP bandwidth estimation as compared to MRand LSTP-DBA algorithms. Further, based on both ON/OFF
prediction and the estimated bandwidth, the proposed MLPDBA was able to: 1) allocate the estimated bandwidth to
ONU-APs that are in ON intervals (Class A); 2) reserve bandwidth from ONU-APs that are in OFF intervals (Class B);
and 3) concurrently providing additional transmission opportunities to heavily-loaded ONU-APs (Class C). Consequently,
MLP-DBA effectively reduced the uplink latency and alleviated packet drops as compared to the baseline limited-service
DBA and predictive MR-DBA and LSTP-DBA algorithms.
B. ONU-AP Classification Analysis
In each polling cycle, the CO classifies ONU-APs upon
receiving the REPORTs. Fig. 10(a) and (b) show the distribution of the classification results of ONU-APs, i.e., Class A, B,
and C, during MLP-DBA operation over 500 s simulation time
in 16-ONU-AP and 32-ONU-AP network, respectively. In the
lightly-loaded region, i.e., for normalized network loads under
0.4, Class-B ONU-APs dominate since the ONU-APs tend to
experience long OFF intervals. In this case, when an ONU-AP
occasionally encounters an ON interval, arriving packets can
be transmitted within BWmax . When the network traffic load
increases, the percentage of Class A increases and followed by
Class C, which rapidly increases and dominates the network
due to the heavy packet bursts at these ONU-APs. In this
case, conventional predictive DBAs become less effective as
BWmax is allocated in majority of the polling cycles due to the
dominated Class-C ONU-APs. Therefore, marginal improvement in reducing latency and packet drop ratio is shown in
Figs. 8 and 9 when network loads are beyond 0.5 with MRDBA, LSTP-DBA, and MLP-DBA without classification. In
contrast, MLP-DBA with the proposed ONU-AP classification adaptively allocates BWdem to Class-A ONU-APs and
enables additional transmission opportunities to heavily loaded
Class-C ONU-APs, thereby effectively reducing latency and
packet drop ratio as presented in Figs. 8 and 9. Moreover, since
bandwidth contention amongst ONU-APs is more competitive
R EFERENCES
[1] G. P. Fettweis, “The Tactile Internet: Applications and challenges,” IEEE
Veh. Technol. Mag., vol. 9, no. 1, pp. 64–70, Mar. 2014.
[2] M. Simsek, A. Aijaz, M. Dohler, J. Sachs, and G. Fettweis, “5Genabled Tactile Internet,” IEEE J. Sel. Areas Commun., vol. 34, no. 3,
pp. 460–473, Mar. 2016.
[3] M. Dohler et al., “Internet of skills: Where robotics meets AI, 5G and
the Tactile Internet,” in Proc. EuCNC, Oulu, Finland, Jun. 2017, pp. 1–5.
[4] E. Wong, M. P. I. Dias, and L. Ruan, “Predictive resource allocation
for Tactile Internet capable passive optical LANs,” J. Lightw. Technol.,
vol. 35, no. 13, pp. 2629–2641, Jul. 1, 2017.
[5] M. Chowdhury and M. Maier, “Collaborative computing for advanced
Tactile Internet human-to-robot (H2R) communications in integrated
FiWi multirobot infrastructures,” IEEE Internet Things J., vol. 4, no. 6,
pp. 2142–2158, Dec. 2017.
[6] M. Chowdhury, E. Steinbach, W. Kellerer, and M. Maier, “Context-aware
task migration for HART-centric collaboration over FiWi based Tactile
Internet infrastructures,” IEEE Trans. Parallel Distrib. Syst., vol. 29,
no. 6, pp. 1231–1246, Jun. 2018.
[7] E. Wong, M. P. I. Dias, and L. Ruan, “Tactile Internet capable passive
optical LAN for healthcare,” in Proc. OECC, Niigata, Japan, Jul. 2016,
pp. 1–3.
[8] S. Mondal, G. Das, and E. Wong, “A novel cost optimization framework
for multi-cloudlet environment over optical access networks,” in Proc.
IEEE GLOBECOM, Singapore, Dec. 2017, pp. 1–7.
[9] M. Condoluci et al., “Soft resource reservation for low-latencyed teleoperation over mobile networks,” IEEE Access, vol. 5, pp. 10445–10455,
2017.
[10] X. Xu, Q. Liu, and E. Steinbach, “Toward QoE-driven dynamic control
scheme switching for time-latencyed teleoperation systems: A dedicated
case study,” in Proc. IEEE HAVE, Oct. 2017, pp. 1–6.
Authorized licensed use limited to: Microchip Tech Inc. Downloaded on January 24,2022 at 13:12:05 UTC from IEEE Xplore. Restrictions apply.
3752
IEEE INTERNET OF THINGS JOURNAL, VOL. 6, NO. 2, APRIL 2019
[11] M. Condoluci, G. Araniti, T. Mahmoodi, and M. Dohler, “Enabling
the IoT machine age with 5G: Machine-type multicast services for
innovative real-time applications,” IEEE Access, vol. 4, pp. 5555–5569,
2016.
[12] G. Kramer, Ethernet Passive Optical Networks. New York, NY, USA:
McGraw-Hill, 2005.
[13] Y. Luo and N. Ansari, “Bandwidth allocation for multiservice access on
EPONs,” IEEE Commun. Mag., vol. 43, no. 2, pp. 16–21, Feb. 2005.
[14] M. P. Mcgarry, M. Reisslein, and M. Maier, “Ethernet passive optical
network architectures and dynamic bandwidth allocation algorithms,”
IEEE Commun. Surveys Tuts., vol. 10, no. 3, pp. 46–60, 3rd Quart.,
2008.
[15] I. Mamounakis, K. Yiannopoulos, G. Papadimitriou, and E. Varvarigos,
“On the prediction of EPON traffic using polynomial fitting in optical
network units,” in Proc. OPTICS, Vienna, Austria, Aug. 2014, pp. 1–7.
[16] R. Kubo, J. I. Kani, Y. Fujimoto, N. Yoshimoto, and K. Kumozaki,
“Adaptive power saving mechanism for 10 gigabit class PON systems,”
IEICE Trans. Commun., vol. E93–B, no. 2, pp. 280–288, 2010.
[17] M. Fiammecngo, A. Lindström, P. Monti, L. Wosinska, and B. Skubic,
“Experimental evaluation of cyclic sleep with adaptable sleep period
length for PON,” in Proc. 37th Eur. Conf. Exhibit. Opt. Commun.,
Geneva, Switzerland, Sep. 2011, pp. 1–3.
[18] M. P. I. Dias, B. S. Karunaratne, and E. Wong, “Bayesian estimation and prediction-based dynamic bandwidth allocation algorithm for
sleep/doze-mode passive optical networks,” J. Lightw. Technol., vol. 32,
no. 14, pp. 2560–2568, Jul. 15, 2014.
[19] H.-J. Byun, J. -M. Nho, and J.-T. Lim, “Dynamic bandwidth allocation
algorithm in Ethernet passive optical networks,” Elect. Lett., vol. 39,
no. 13, pp. 1001–1002, Jun. 2003.
[20] Y. Luo and N. Ansari, “Limited sharing with traffic prediction for
dynamic bandwidth allocation and QoS provisioning over EPONs,” J.
Opt. Netw., vol. 4, no. 9, pp. 561–572, Sep. 2005, .
[21] P. Sarigiannidis, D. Pliatsios, T. Zygiridis, and N. Kantartzis, “DAMA: A
data mining forecasting DBA scheme for XG-PONs,” in Proc. MOCAST,
Thessaloniki, Greece, May 2016, pp. 1–4.
[22] L. Ruan, S. Mondal, and E. Wong, “Machine learning based bandwidth prediction in Tactile heterogeneous access networks,” in Proc.
INFOCOM (WORKSHOP), Honolulu, HI, USA, Apr. 2018, pp. 1–2.
[23] N. Kato et al., “The deep learning vision for heterogeneous network traffic control: proposal, challenges, and future perspective,” IEEE Wireless
Commun., vol. 24, no. 3, pp. 146–153, Jun. 2017.
[24] E. Wong, “Current and next-generation broadband access network
technologies,” J. Lightw. Technol., vol. 30, no. 4, pp. 597–608, 2012.
[25] S. Haykin, Neural Networks: A Comprehensive Foundation, 2nd ed.
New Delhi, India: Prentice-Hall, 1999.
[26] W. H. Tranter et al., “On the self-similar nature of Ethernet traffic (extended version),” in The Best of the Best: Fifty Years of
Communications and Networking Research. Piscataway, NJ, USA: IEEE
Press, 2007, pp. 517–531.
[27] R. Fontugne et al., “Scaling in Internet traffic: A 14 year and 3 day
longitudinal study, with multiscale analyses and random projections,”
IEEE/ACM Trans. Netw., vol. 25, no. 4, pp. 2152–2165, Aug. 2017.
[28] P. Loiseau et al., “Investigating self-similarity and heavy-tailed distributions on a large-scale experimental facility,” IEEE/ACM Trans. Netw.,
vol. 18, no. 4, pp. 1261–1274, Aug. 2010.
Lihua Ruan (GS’16) received the M.S. degree from Northwestern
Polytechnical University, Xi’an, China, in 2015. She is currently pursuing
the Ph.D. degree at the University of Melbourne, Melbourne, VIC, Australia.
Her current research interests include smart body area networks, wireless
local area networks, and dynamic bandwidth resource allocation in passive
optical networks, in conjunction with the use of machine learning techniques,
for tactile Internet applications.
Maluge Pubuduni Imali Dias received the B.Sc. degree (Hons.) from the
University of Moratuwa, Moratuwa, Sri Lanka, in 2007, and the Ph.D. degree
from the University of Melbourne, Melbourne, VIC, Australia, in 2016.
She is currently engaged in teaching and research activities with the
University of Melbourne. Her current research interests include dynamic bandwidth allocation algorithms, energy-efficiency, tactile Internet, and intelligent
transportation systems.
Elaine Wong (SM’14) received the B.E. and Ph.D. degrees from the
University of Melbourne, Melbourne, VIC, Australia.
She is currently a Professor and an Australia Research Council Future
Fellow with the University of Melbourne. She has coauthored over 150 journal
and conference publications. Her current research interests include energyefficient optical and wireless networks, optical-wireless integration, broadband
applications of vertical-cavity surface-emitting lasers, wireless sensor body
area networks, and emerging optical and wireless technologies for tactile
Internet.
Dr. Wong has served on the Editorial Board of the Journal of Lightwave
Technology and the Journal of Optical Communications and Networking.
Authorized licensed use limited to: Microchip Tech Inc. Downloaded on January 24,2022 at 13:12:05 UTC from IEEE Xplore. Restrictions apply.