Optical Fiber Technology 75 (2023) 103191

Contents lists available at ScienceDirect

Optical Fiber Technology

journal homepage: www.elsevier.com/locate/yofte

Regular Articles

From 5G to beyond: Passive optical network and multi-access edge

computing integration for latency-sensitive applications
Imali Dias a , Lihua Ruan b , Chathurika Ranaweera a , Elaine Wong c ,∗
a
Faculty of Science Engineering and Built Environment, Deakin University, 3125, VIC, Australia
b
Chinese University of Hongkong, Shenzhen, 518172, PR China
c
Department of Electrical and Electronic Engineering, University of Melbourne, 3010, VIC, Australia

ARTICLE INFO ABSTRACT

Keywords: A recent paradigm shift in support of 5G-and-beyond (5GB), Human-to-Machine/Robot (H2M/R), and the
Federated learning Tactile Internet has resulted in a surge of latency-sensitive applications being delivered across communication
Human-to-Machine/Robot (H2M/R) networks. These applications in conjunction with exponential growth in connected devices, have pushed for
communications
architectural and capacity changes in both wired and wireless networks. Passive optical networks (PONs)
Low-latency applications
have evolved in terms of their capacity and function over the years and are now supporting both fixed as
Multi-access edge computing (MEC)
Passive optical network (PON)
well as x-haul solutions to deliver cost-efficient, low-latency and high-bandwidth applications to end-users.
In combination with multi-access edge computing (MEC), high computational tasks of applications can be
offloaded to servers much closer to end-users further reducing the network latency. To this end, this paper
presents a comprehensive review of the capabilities of integrated PON-MECs in delivering emerging latency-
sensitive applications. Specific use case scenarios benefiting from integrated PON-MECs are also detailed.
Shared learning through a federated PON-MEC framework to expedite resource allocation decisions is discussed
and finally, the time-savings arising from implementing the proposed framework is quantified, showing further
improvements in network latency performance.

1. Introduction away from end-users [13]. As a result, user data experienced high prop-
agation delays that prevented the underlying networks from meeting
With the proliferation of 5G networks, Internet-of-Things (IoT) [1], stringent end-to-end latency requirements of applications. To overcome
and Machine-to-Machine communication (M2M) [2], attention is now this problem, MEC enabled cloud computing and IT services to be
focused on next-generation network paradigms including 6G networks brought closer to the end-users through the placement of MEC servers
and the Tactile Internet [3,4]. Wired and wireless communication at the access network edge [9,14,15]. Rather than having the data
networks will continue to rapidly evolve both in terms of their ar- traverse through the mobile backhaul or core networks, computing,
chitectures and capabilities to support the surge in latency-sensitive storage, and caching are now rendered at the access network edge [16].
human-to-machine/robot (H2M/R) applications such as augmented Recent studies on MEC and 5G have highlighted the benefits of using
reality/virtual reality (AR/VR) [5], industry automation [6], tele- optical fiber networks to support the increasing number of connected
surgery [7], and intelligent transport systems (ITS) [8]. The expedited devices and their quality-of-service (QoS)-aware applications. Optical
growth in the number of mobile devices, projected to reach 29.3B in
fiber benefits from high capacity and low propagation delay. Taking
2023 and with each envisioned to generate large amounts of data,
advantage of optical fiber that is already deployed in a majority of ur-
has also warranted new network architectures, hardware, and software
ban areas close to residential and industrial premises, the cost-efficient
solutions [5,9,10].
and high-capacity passive optical network (PON) has become a natural
Today, the seamless delivery of latency-sensitive and reliable ser-
choice for fronthauling and backhauling of 5G and 5GB networks [17].
vices in 5G networks is realized by enabling technologies such as
The PON has evolved over the years in terms of its architecture,
multi-access edge computing (MEC) [9,11] and fog computing [12].
Prior to MEC, cloud services such as computing, storage, and caching hardware, capacity, and underlying protocols and algorithms [18].
were performed at centralized data centers, often at a long distance These developments in conjunction with optical fiber deployments that

∗ Corresponding author.
E-mail addresses: [email protected] (I. Dias), [email protected] (L. Ruan), [email protected] (C. Ranaweera),
[email protected] (E. Wong).

https://doi.org/10.1016/j.yofte.2022.103191
Received 27 April 2022; Received in revised form 3 September 2022; Accepted 30 November 2022
Available online 2 January 2023
1068-5200/© 2022 Elsevier Inc. All rights reserved.
I. Dias et al. Optical Fiber Technology 75 (2023) 103191

have now reached residential and business premises, have enabled the
delivery of bandwidth-hungry and latency-sensitive services to large
numbers of users.
The integration of PON and MEC architectures is a potential solution
to deliver latency-sensitive yet task-intensive H2M/R applications. Such
applications define the capabilities of 5GB and 6G as they require
latency in the milliseconds range, a ultra-reliable network connectivity,
and a significant amount of data processing. In order to satisfy such
stringent QoS requirements, the suitability of the converged PON-MEC
network from different aspects such as PON and MEC configurations,
resource allocation across heterogeneous network segments, and com-
plementary technologies such as machine learning that can further
reduce end-to-end latency through added intelligence at the mobile
edge, should be carefully explored. Hence, in this paper, we review
PON, MEC and PON-MEC integration in detail and investigate how
their inherent characteristics assist low-latency and task-intensive ap-
plications. To demonstrate the capabilities of integrated PON-MECs in
supporting latency-sensitive applications, we propose a bandwidth al-
location decision framework that uses reinforcement learning to reduce
bandwidth allocation decision time and sharing this learning across a
multiple PON-MECs that form a federation to further improve latency
performance.
The remainder of this paper is structured as follows. In Section 2, we
present a review of PON and MEC architectures In Section 3, we discuss Fig. 1. Multi-Access Edge Computing (MEC) architecture [33,34].
the PON-MEC integration architectures and specific use case scenarios
of emerging latency-sensitive applications that would benefit from an
integrated PON-MEC network. In Section 4, we describe our framework
• GPON and 1G-EPON
that expedites bandwidth decision making in federated PON-MEC for
H2M/R communications. Finally, the main contributions of this paper – ITU-T G.984 Gigabit PON (GPON) [21]
are summarized in Section 5. – IEEE 802.3ah Gigabit Ethernet PON (1G-EPON) [22]

• 10-Gigabit-Capable PON
2. Background: Passive optical networks and multi access edge
computing – IEEE Std. 802.3av 10 GE-PON [23]
– ITU-T G.987 10G-PON standard [24]
In this section, we provide an overview of major PON and MEC – ITU-T G.9807.1 XGS-PON [25]
architectures and their inception to recent deployments, which serves
• NG-PON2
as the foundation to discuss the main concepts covered in the rest of
the paper. – ITU-T G.989 Next-Generation PON 2 (NG-PON2) standard
[26]
2.1. PON overview
• 50-Gigabit-Capable PON
Over the last decade, we have witnessed the unrelentless penetra- – IEEE 802.3ca 25G/50G-EPON standard [27]
tion of optical fiber deeper and deeper into residential and business – ITU-T G.9804 High Speed PON (HSP) [28]
premises. According to [19], 40 percent of the world’s population will
have fixed Fiber-To-The-x (FTTx) connection by 2026. The models of 2.2. MEC overview
FTTx, namely FTTH (home), FTTC (curb) and FTTB (building), installed
either through a physical point-to-multipoint (P2MP) or point-to-point The concept of MEC evolved from mobile cloud computing (MCC)
(P2P) topology offer direct fiber connection to or close to the end- where mobile devices accessed remote centralized data centers through
users. The global FTTx market is still growing with a steep trajectory, the mobile network to offload their computation and data process-
predicted to reach 22.4B by 2026 [20]. ing [29]. The downside of MCC is the user data having to traverse
Most FTTx networks are based on the PON technology due to its through the core and access networks to reach centralized cloud re-
cost-effectiveness and low energy consumption per bit. The PON has a sources, thereby experiencing significant propagation delay and con-
P2MP topology with an unpowered optical distribution network (ODN). gesting the backhaul and core networks [30]. As the number of mobile
Specifically, the ODN comprises a feeder fiber that connects the optical devices and the QoS requirements of their applications continued to
line terminal (OLT) at the central office (CO) to a passive splitter and grow, in order to enjoy the benefits of MCC while also meeting latency
a number of distribution fibers that connects the passive splitter to and other QoS requirements, cloud and IT services had to be brought
optical network units (ONUs) at end-user locations. For N subscribers, closer to the end-users. For this purpose, Mobile Edge Computing which
a PON requires only the installation of a single transceiver at CO and N is an initiative of the ETSI Industry Specification Group (ISG), was
transceivers at the ONUs, giving rise to N +1 transceivers overall. The identified as a favorable solution to integrate edge computation into
shared feeder fiber saves CO space and allows for easy termination. mobile networks [9,31,32].
This physical topology also allows its downstream and upstream band- Fig. 1 illustrates the placement of mobile edge computing servers
width to be naturally shared through time division multiplexing (TDM) at a close proximity to end-users in a radio access network (RAN).
and time division multiple access (TDMA), respectively. The following Specifically, the servers are placed at multiple locations including the
are some of the TDM/TDMA-based PON architectures standardized in LTE macro base station, radio network controller (RNC) or the multi-
recent literature. radio access technology (RAT) cell aggregation point [31]. The servers

2
I. Dias et al. Optical Fiber Technology 75 (2023) 103191

function to either process requests from end users in its vicinity or transport all wireless data to the metro or core networks [40]. In
forward requests onto the centralized data centers for further pro- particular, emerging 6G wireless networks is envisioned to consist of
cessing. Realizing the possibility of serving both mobile and wired ultra-small wireless cells to achieve the required capacity and coverage
networks through edge computing in order to deliver latency-sensitive requirements. As higher cell densities can lead to costlier x-haul trans-
and computation-heavy applications, ETSI further redefined Mobile port connectivity [41], such cell densification calls for cost-efficient
Edge Computing as Multi-Access Edge Computing (MEC) in which the x-haul solutions in order to meet global business goals.
new MEC architecture is defined as one that serves both wired and A traditional transport network typically consists of direct P2P
wireless networks at the edge [35,36], as shown in Fig. 1. fiber/wireless links between cell sites and the network core, which is
commonly referred to as backhaul. However, to cope with the increas-
3. PON-MEC integration ing demands in a cost and energy-efficient manner, a new paradigm
of centralized processing of mobile signals was first introduced with
3.1. Network architecture 4G technology [43]. Cloud/centralized radio access network (C-RAN)
is one of the main architectures currently deployed as centralized
It has emerged in recent literature the integration of both PON
processing allows for advance wireless cooperative communication
and MEC as a favorable solution in provisioning broadband access
functionalities. In C-RAN, all functions are centralized in a baseband
to meet QoS requirements. In [13], the authors investigate how to
processing unit (BBU) which is placed at a central location. Radio
optimally plan a WDM-PON-based MEC-enabled fiber wireless access
frequency (RF) functionalities are performed at the remote radio head
network (MFWAN) for latency-sensitive applications. The proposed
(RRH), located at the antenna cell site. In this architecture, RF signals
work presents a mathematical model to minimize the end-to-end la-
are digitized and sent over fiber to the centralized BBU using Common
tency using constraints such as maximal transmission distance, maximal
PON power budget, and bandwidth requests. In [37], the authors Public Radio Interface (CPRI).
investigate different optical fronthaul architectures to support ultra- In a native C-RAN, the bandwidth required for this digitized signal
low latency and reliability required in 5G mobile fronthaul. Taking scales with the physical parameters of antenna, e.g., antenna port count
Tactile Internet as a use case, the authors of [3] present a decentralized and wireless bandwidth in use. However, with the high number of
cooperative dynamic bandwidth allocation algorithms for end-to-end antenna ports in use in 5G and beyond networks, the fronthaul between
resource coordination in a MEC-enabled fiber-wireless (FiWi) access the BBU and RRH consequently demands exceedingly high-capacity
networks using an EPON. Additionally, the authors of [38] study the links. As an example, using CPRI to support a fronthaul with 200
collaborative computation offloading in fiber-wireless networks, specif- MHz of bandwidth, 64 QAM, 8X12 MIMO, and 96 antenna ports, will
ically considering the resource allocation between a mobile device and necessitate a 896 Gb/s link [44]. The demand for such high-capacity
the MEC server, and also between the mobile device and the centralized links makes C-RAN an impractical architecture.
server. The authors of [33] also propose a dynamic resource alloca- A solution to this problem is to move some functionalities back to
tion algorithm considering both centralized cloud and MEC enabled the cell site. There are ten different standardized ways, termed func-
fiber-wireless networks. tional split options, in which functions can be split between the BBU
The architecture of an integrated PON-MEC network is schemat- and RRH [45]. In the latest 5G New Radio (NR) transport architecture,
ically shown in Fig. 1, whereby the OLT of the PON interconnects BBU functions are split into three entities. The centralized unit (CU),
a centralized cloud (via the core network and the Internet) with 𝑁 distributed unit (DU), and remote unit (RU). This new design in 5G
number of ONUs (via a passive splitter). The ONUs co-exist with mobile facilitates RAN virtualization with flexible assignment of computing
base stations (BSs) and are connected to MEC servers placed at the resources. The functional split options and their required x-haul data
edge of the RAN. Under such an integrated PON-MEC network, the rates are illustrated in Fig. 2.
MEC servers support not just the mobile users but also wired customers In comparison to other competing access technologies, PON can
through the ONUs. It is also important to note that not all mobile play a critical role in realizing high-capacity yet low-cost x-haul links
services require mobility and therefore would benefit from the fixed [46]. To date, different PON architectures have been investigated for
optical network resources connected to the RAN. this purpose [46–48]. As an example, ITU-T G.Sup66 investigated tra-
ditional and emerging PON technologies to support different functional
3.2. Use cases split options, while considering different bandwidth, latency, and other
x-haul and 5G parameter constraints [48]. It is also worth mentioning
As discussed in the previous section, both PON and MEC have
that while the PON is unable to support low layer functional split
inherent characteristics to support QoS-aware emerging use cases. In
options such as Option 8 and 7.1 due to its capacity limitation, it is still
this section, insights into how the integration of high-capacity PONs
a suitable candidate to support functional split options above 7.2 (refer
with MEC capability can address QoS requirements in some use case
to Fig. 2). Specifically, PON can easily support 5G transport with the
scenarios are highlighted.
7.2 or higher layer functional split options, and this specific use of PON
is considered in the latest RAN architectures such as Open Intelligent
3.2.1. X-hauling 5GB wireless networks
Driven by the increasing number of mobile devices and the use RAN (O-RAN) [49]. With the principles of intelligence and openness in
of smart mobile applications, mobile technology is continuously and mind, the O-RAN architecture is standardized for building a virtualized
rapidly advancing. The rapid deployment of 5G wireless technology RAN on open hardware with the help of artificial intelligence assisted
has reshaped network capabilities in an unimaginable way. To ad- radio control.
dress the ongoing issue of providing mobile connectivity to an ever There are however existing challenges that requires addressing
growing number of connected devices and with increasingly stringent if considering PON-based x-haul solutions. Depending on the actual
QoS requirements, both academia and industry have recently begun functional split option used, a key issue concerns the realization a scal-
to research the sixth generation (6G) mobile technology. According able and cost-effective deployments. Several studies have investigated
to [39], 6G mobile technology is envisioned to provide a peak data rate approaches to plan and dimension PON-based x-haul cost-effectively
of 100 Gb/s, a user-experienced data rate of 1 Gb/s and an ultra-low [42,50,51]. For example in [42], the authors propose a generalized
latency of less than 1 ms over ubiquitous 3D coverage areas. integer linear programming based optimization framework that can be
These high capacity future 5G/6G wireless networks must be sup- used to jointly plan both PON-based x-haul and 5G wireless networks
ported by complementary high-capacity and secured transport net- to provide the cost-optimal network design in any given deployment
works for backhauling/fronthauling (x-hauling) the wireless cells to scenario and irrespective of the RAN architecture used. Further, the

3
I. Dias et al. Optical Fiber Technology 75 (2023) 103191

Fig. 2. 5G functional spit options with supporting optical x-haul technologies [42].

framework takes into consideration a diverse set of network require- the actions taken by the surgeon. Due to the life-critical nature of this
ments such as network coverage and end-user capacities. Analyses on application, the end-to-end latency is expected to be less than 1 ms with
different 5G RAN and x-haul deployment scenarios show that PON- a reliability of 99.999% [36]. Applications of this nature will benefit
based x-haul options can save more than 30% to 40% of the deployment from PON-MEC integration as the PON will provide high bandwidth
cost compared to P2P fiber-based x-haul option. and reliability while the MEC ensures faster processing of data using
Another key issue revolves around resource management and alloca- ML and DL at the network edge.
tion. In both PON and 5G networks, all RUs and end-users respectively, Fig. 3 depicts integrated PON-MECs supporting different use case
share the same bandwidth. Due to this shared bandwidth, two major scenarios. More specifically, MEC servers are placed at the edge of
factors can severely affect the latency, namely the discovery window the network and are connected to the ONUs placed at the base sta-
and the bandwidth allocation algorithm. The discovery window, which tions. For health surveillance applications, data generated by the user
is periodically opened by the OLT for newly registering ONUs, can wearable sensors will be analyzed in these MEC servers in order to
cause delays in the x-haul. Different studies have looked at minimizing identify user patterns and to send warnings or notifications to carers or
the impact of discovery window on network latency. For example, emergency services in case of an emergency. The location and context-
in [52], the authors propose a mechanism to assign the discovery awareness together with high-capacity optical fiber connections will
window by using network slicing in order to reduce network latency. ensure latency-sensitive service delivery in this scenario. In applications
Effective approaches to reduce latency using intelligent bandwidth such as remote surgery, the MEC-PON integration itself may not be
allocation algorithms have also been explored in [52–56]. In [52], the able to satisfy stringent latency requirements. Additional techniques,
authors propose a cooperative bandwidth allocation by considering the including ML to predict the resource requirement of the end-users, are
actual data arrival period of fronthaul traffic to improve the latency also necessary [61]. Both MEC and centralized servers can be used for
and the bandwidth utilization. Further, in [33] the authors investigate this purpose, depending on their computation capabilities and trained
data stored at the MEC for reduced latency.
the bandwidth allocation of an integrated PON-MEC x-haul and corre-
spondingly propose a dynamic resource allocation algorithm to meet
3.2.3. Industry 4.0
network and user requirements.
Industry 4.0 involves improving the operation of traditional facto-
Looking further into the near future, integrated PON-MEC x-haul for
ries such that they are able to run more efficiently and adapt quickly
5GB networks, should include complementary technologies including
to changing requirements of the production lines [62]. A shift in this
machine learning (ML), network slicing, and NVF in order to meet
direction is made possible by novel IoT technologies, data analysis
the stringent service requirements. Further research into architectures,
techniques, and network improvements such as edge computing. As
resource allocation mechanisms and a design framework is warranted
discussed in [63], a large amount of data is being generated at factory
in order to achieve full convergence between optical and wireless access
settings. These data can be analyzed and be converted to intelligence
technologies, ML, network slicing and MEC.
which can be used to optimize different aspects of the manufactur-
ing process [64]. For example, ML-based predictive maintenance can
3.2.2. e-Health be used, where a failure of a component/device is predicted before-
Healthcare has evolved significantly over the years. Services that hand to minimize the downtime of the manufacturing process. The
previously required in-person interaction with healthcare professionals human–agent–robot teamwork (HART) is another promising concept
are now being delivered remotely. For example, intelligent health that can be used in factory floors to provide either location-dependent
surveillance systems have been proposed using wearable IoT sensors, or location-independent but computation heavy services [58]. The pro-
which collect data pertaining to user behaviors. This sensor data is then cess can be optimized to select either a resource-limited robot for tasks
processed and analyzed using ML and deep learning (DL) techniques to that requires moving from one place to the other or a resource-rich
identify user patterns and anomalies. Such applications however, are cloud agent for tasks that requires heavy computation.
latency sensitive and require significant processing resources which an The factory automation services of this nature deals with controlling
end user may not be equipped with [60]. As such, MEC has been pro- the equipment in real-time, which would need latency in the range
posed for such applications to offload their computation and processing of 25 𝜇 to 10 ms depending on the application and significant data
tasks to an edge server placed close to the end-user. Another emerging processing. Further, due to the critical nature of these applications, a
e-health application gaining interest is remote surgery, where a skilled reliability with packet loss rate no less than 10−9 is expected [65]. As
surgeon performs a surgery on a patient who is remotely located. The such, a network architecture that combines the low-latency computa-
surgeon controls a surgical robot placed at the patient side who mirrors tion of MEC and reliability and high bandwidth of PON is preferred. The

4
I. Dias et al. Optical Fiber Technology 75 (2023) 103191

Fig. 3. PON-MEC use cases [57–59].

inset of Fig. 3 illustrates a possible configuration of HART. The MEC The importance of integrated PON-MEC networks is clearly demon-
server in this case provides the required computation for ML and also strated through the use cases described above. More specifically, the
for selecting agents/robots for different tasks. The applications of this MEC assists in offloading the computation and storage close to the end-
nature specifically benefit from location and context-aware information user, thereby reducing the latency, while the PON facilitates connectiv-
available at the MEC. ity through its high-capacity optical links. It is however, important to
note that the integration of these two architectures does not guarantee
3.2.4. Augmented Reality/Virtual Reality that latency requirements will be met. For this purpose, end-to-end
resource allocation across these heterogeneous networks has to be
Augmented Reality (AR), Virtual Reality (VR) and Extended Reality
carefully managed through dynamic resource allocation algorithms, as
(XR) have become popular technologies in recent years to enhance
will be discussed below.
the digital experience of users through providing a digital component
to the real world, real life experience in the digital world, and a
4. Federated PON-MECs for H2M/R communications
combination of both. This technology is currently being deployed in
medical, educational, gaming, and streaming sectors [66]. For example,
4.1. Motivation
in healthcare, applications such as surgical training simulators and
remote medical systems are now being explored using AR, VR, and
H2M/R applications such as tele-rehabilitation and cloud AR/VR
XR [67].
play an important role in the aforementioned emerging e-health, in-
In the work reported in [57], the authors discuss AR-based remote dustrial automation and AR/VR use cases [68]. Unlike conventional
patient monitoring and robotic surgery to mitigate risk to medical human-to-human (H2H) and M2M communications, H2M/R commu-
personnel during the Covid-19 pandemic. AR-based remote monitoring nication features real-time human control of and haptic feedback
is proposed to perform daily routines carried out by the medical pro- from distant machines/robots, thus allowing immersive human inter-
fessional using robots and is expected to have comparatively high pre- action with remote environments. However, the perception of real-
cision than other conventional remote monitoring systems. AR-based time H2M/R communication requires millisecond round-trip latency
robotic surgery is also proposed as means of getting more precision between the human and machine/robot [4].
while performing invasive remote surgery, either when the surgeons Recent literature in [4,69,70] suggests a converged network solu-
are not available at the patient’s medical facility or to reduce risk tion to meet this stringent round-trip latency. Fig. 4 shows such a
of exposure. The inset of Fig. 3 illustrates these two AR-based medi- network, comprising passive optical fiber and wireless access together
cal services implemented using PON-MEC integration. As emphasized with MEC functionalities. In conjunction with high-capacity PONs, MEC
in [57], improved precision is one of the main reasons to select AR servers co-located at the COs and radio front-ends play the crucial
as a favorable option for this application. To support this requirement, role in bringing video, audio and proprioception resources closer to
a large amount of video and audio data has to be processed. Further, end-users. As importantly, through ML based traffic prediction, MEC
life-critical applications of this nature require low latency and high re- can expedite haptic feedback delivery from the machine/robot to the
liability. These three requirements are well supported by the integrated human further improving latency performance and decoupling the
PON-MEC architecture of Fig. 3. human and machine/robot distance from the achievable end-to-end

5
I. Dias et al. Optical Fiber Technology 75 (2023) 103191

Fig. 4. FedRL framework over federated PON-MECs to support low-latency H2M/R communications.

latency [4,69,71]. This also helps reduce the amount of data traversing As can be gathered, 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 is a key decision parameter in which
the network, thus relieving bandwidth to serve other latency-sensitive under-granting 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 can lead to increased waiting time while
traffic [70,72–74]. over-granting 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 can lead to bandwidth inefficiency. Ideally,
𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 will reduce the waiting time of a newly arrived packet to
4.2. Bandwidth allocation in PON-MECs 0.5 polling-cycle time. That is, newly arrived packets can be trans-
mitted immediately without having to wait for the next polling cycle.
To determine 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 , conventional dynamic bandwidth allocation
As H2M/R communications are highly interactive, the PON needs
algorithms exploit statistical prediction, i.e. using statistical features of
to support frequent data exchange amongst the MEC servers/ONU as
the data received and transmitted in previous polling cycles. Different
shown by PON-MEC 2 of Fig. 4. As uplink bandwidth is shared, the
statistical prediction methods suit different data traffic profiles. For
allocation of uplink bandwidth to each MEC server/ONU to commu-
example, Bayesian estimation [78] and linear regression [79] are used
nicate with the CO, is particularly important to maximize bandwidth
to predict traffic with favorable time-average performance whereas
efficiency and latency performance. In that regard, a REPORT-GRANT
moving average [80] is more suitable to predicting bursty traffic.
process is adopted in classic dynamic bandwidth allocation algorithms
to allocate bandwidth to ONUs [75]. Firstly, ONUs request for band-
4.3. Reinforcement learning-based bandwidth allocation in federated PON-
width 𝐵𝑊𝑅𝐸𝑄 based on their uplink buffer occupancy, in REPORT
MECs
messages to the CO in a round robin fashion. Upon receiving a REPORT
message, the CO then allocates bandwidth 𝐵𝑊𝐺𝑅𝐴𝑁𝑇 to the ONU
In recent years, supervised ML techniques have been investigated
through the GRANT messages. The time interval between an ONU’s
to improve the prediction accuracy of 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 [81–84] though the
consecutive uplink transmissions is known as a polling cycle, which
focus has not always been to meet stringent latency constraints. Specif-
in general is the scale of milliseconds [76]. The dominant latency is
ically to support H2M communications over PON-MECs of Fig. 4,
attributed to the waiting time a packet incurs while waiting to be
we proposed a federated reinforcement learning (FedRL) framework
transmitted upstream from the ONU. If an ONU is allocated insufficient
that leverages: (a) reinforcement learning (RL) to optimize 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆
𝐵𝑊𝐺𝑅𝐴𝑁𝑇 , its packets will spend an increased waiting time in the
together; with (b) federated learning (FL) between multiple PON-MECs
buffer, thus degrading the overall latency performance of the network.
to minimize the time cost in executing sub-optimal decisions [85].
In the baseline scheme, the CO grants 𝐵𝑊𝐺𝑅𝐴𝑁𝑇 equivalent to that
RL is renown for optimizing decisions via exploration–exploitation
requested by an ONU but only up to a maximum, i.e. 𝐵𝑊𝐺𝑅𝐴𝑁𝑇 = and allowing the decisions to be self-adaptive to the environment [86].
min{𝐵𝑊𝑅𝐸𝑄 , 𝐵𝑊𝑀𝐴𝑋 }. This scheme is considered a conservative In RL, an agent learns the best decision through strategic exploration of
scheme as any new packet arrivals will need to be queued and buffered various decisions and receiving the rewards. As a trade-off, an optimal
until they are reported. On average, a packet will incur a latency decision comes at the expense of time cost in executing sub-optimal
equivalent to an approximate 1.5 polling-cycle time before being trans- decisions. In the short term, the decision exploration may negatively
mitted. To reduce this latency, a common solution is to predict the affect network performance, e.g. network latency.
surplus bandwidth corresponding to new packet arrivals and allocate To overcome the limitations of RL, we leverage FL to enable ex-
this bandwidth, 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 , along with 𝐵𝑊𝑅𝐸𝑄 [77]. Hence, with perience sharing between multiple PON-MECs that form a federation.
bandwidth prediction, the CO grants bandwidth 𝐵𝑊𝐺𝑅𝐴𝑁𝑇 as follows: By definition, FL is to enable distributed devices to collaboratively
learn a shared ML model [87]. It is expected that all PON-MECs in
𝐵𝑊𝐺𝑅𝐴𝑁𝑇 = min{𝐵𝑊𝑅𝐸𝑄 + 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 , 𝐵𝑊𝑀𝐴𝑋 }. (1)
a federation have similar states, e.g., with similar architecture and
traffic environment, and thus sharing of decision experience between

6
I. Dias et al. Optical Fiber Technology 75 (2023) 103191

the PON-MECs will accelerate individual decision making, eventually Algorithm 1: FedRL framework
benefiting all PON-MECs in the federation.
1 Parameters:
In the subsections below, we first introduce local RL in individual
2 𝑡 — time stamp in unit of polling cycle
PON-MECs, and then FL to expedite local learning in the proposed
3 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆,𝑛 — decision in the 𝑛-th PON-MEC
FedRL framework. Finally, we introduce the key network features to
4 𝑄𝑛𝑒𝑡,𝑛 — local 𝑄 of the 𝑛-th PON-MEC
be considered in selecting federated PON-MECs for FL.
5 𝑄𝑓 𝑒𝑑 — shared 𝑄 by federated PON-MECs

4.3.1. 𝑄-learning in local PON-MEC 6 Operation at 𝑛-th PON-MEC ONU:

Firstly, RL is used in each local PON-MEC to make individual 7 if in transmission window then
𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 decisions. With low latency as our goal, we define the imme- 8 Send buffered packets to the CO
diate reward, termed 𝑅, of a 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 decision as the negative latency 9 if 𝑡 = polling cycle number/𝑇𝑆𝑇 𝐸𝑃 is integer then
value when the CO grants a 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 to the ONUs. Specifically, as 10 𝑑𝑎𝑣𝑔 ← average uplink latency in 𝑇𝑆𝑇 𝐸𝑃 cycles
the ONUs transmit to the CO in a round-robin manner, in every 𝑇𝑆𝑇 𝐸𝑃 11 𝑅(𝑡) ← −𝑑𝑎𝑣𝑔
polling-cycle interval, an ONU evaluates the average latency 𝑑𝑎𝑣𝑔 of its 12 end
packets and reports the reward 𝑅 = −𝑑𝑎𝑣𝑔 to the CO. As such, during the 13 𝐵𝑊𝑅𝐸𝑄 ← remaining packets in the buffer
𝑡th step, the CO acquires a decision reward pair as {𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 (𝑡), 𝑅(𝑡)}. 14 Send 𝐵𝑊𝑅𝐸𝑄,𝑛 and 𝑅(𝑡) (if 𝑅(𝑡) is updated at 𝑡)
Based on the decision reward pairs, the CO adjusts 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 for 15 end
the next 𝑇𝑆𝑇 𝐸𝑃 in order to maximize the long-term expected reward.
We adopt the 𝑄-learning RL algorithm to measure the decision value 16 Operation at 𝑛-th PON-MEC CO:
and adjust decisions accordingly [86]. Further, we use a local decision- 17 if receive 𝐵𝑊𝑅𝐸𝑄,𝑛 from ONU then
value function, denoted as 𝑄𝑛𝑒𝑡 , to indicate the expected reward of a 18 if 𝑅(𝑡) updates then
decision in long-term. When the CO receives an 𝑅(𝑡), it updates 𝑄𝑛𝑒𝑡 , 19 Update 𝑄𝑛𝑒𝑡,𝑛 (𝐵𝑊𝐵𝑂𝑁𝑈 𝑆,𝑘 (𝑡)) in (2)
as follows: 20 𝑄𝑓 𝑒𝑑 (𝐵𝑊𝐵𝑂𝑁𝑈 𝑆,𝑛 (𝑡)) ← 𝑄𝑛𝑒𝑡,𝑛 (𝐵𝑊𝐵𝑂𝑁𝑈 𝑆,𝑛 (𝑡)) in (4)
𝑄𝑛𝑒𝑡 (𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 (𝑡)) ← (1 − 𝛼)𝑄𝑛𝑒𝑡 (𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 (𝑡)) 21 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆,𝑛 (𝑡 + 1) = arg max 𝑄𝑓 𝑒𝑑 in (3)
(2) 22 else
+𝛼(𝑅(𝑡) + 𝛾(𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 (𝑡 + 1))).
23 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆,𝑛 (𝑡 + 1) = 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆,𝑛 (𝑡)
The 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 (𝑡) in (2) is chosen from a decision set  that comprises 24 end
discretized bandwidth values. The parameters 𝛼 and 𝛾 (0 ≤ 𝛼, 𝛾 ≤ 1) 25 Grant 𝐵𝑊𝐺𝑅𝐴𝑁𝑇 ,𝑛 in (1) via GRANT message
control the learning rate and the weight on future rewards, respectively. 26 end
With 𝑄𝑛𝑒𝑡 , CO determines 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 (𝑡 + 1) through a greedy policy
given by:
𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 (𝑡 + 1) = arg max 𝑄𝑛𝑒𝑡 . (3)
𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 ∈ The pseudo code of the proposed FedRL framework is described
Eq. (3) follows the Bellman optimality equation and increases the con- in Algorithm 1. Note that the operation of each PON-MEC is asyn-
vergence rate of 𝑄𝑛𝑒𝑡 and 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 . The process of (2) and (3) repeats chronous to each other. Within an individual PON-MEC, ONUs report
until 𝑄𝑛𝑒𝑡 converges. Note that there are other policies in existing reward 𝑅 to their CO following lines 9 to 12 and based on these
literature such as 𝜀-greedy or softmax policy that have been used to reports, the CO updates its corresponding 𝑄𝑛𝑒𝑡,𝑛 as per line 19. Mean-
explore decisions [86] but in general, these policies have a slower while, all federated COs maintain the federal 𝑄 function, 𝑄𝑓 𝑒𝑑 , as per
convergence time as compared to the greedy exploration. The granted line 20. This in turn allows each CO to make individual 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆
bandwidth 𝐵𝑊𝐺𝑅𝐴𝑁𝑇 to ONUs in (1) takes the updated 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 into decisions based on 𝑄𝑓 𝑒𝑑 as per line 21.
account and is sent via GRANT message. The FedRL framework has two favorable properties:

• Asynchronous learning: FedRL operates in a distributed fashion.

4.3.2. Experience sharing in federated PON-MECs The local 𝑄-learning need not to be synchronized to enable FL in
In FedRL, the use of FL to share 𝑄 experiences can help expedite (4). In addition, one PON-MEC can immediately commence its lo-
local 𝑄 learning. Federated PON-MEC shares its 𝑄𝑛𝑒𝑡,𝑛 with other fed- cal 𝑄-learning from the latest 𝑄𝑓 𝑒𝑑 and greedily make 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆
erated PON-MECs through updating a federal decision-value function, decisions using (3).
𝑄𝑓 𝑒𝑑 , as illustrated in Fig. 4. Specifically, a heuristic rule is considered • Low complexity: FedRL operates in low computation complexity.
whereby when a 𝑄𝑛𝑒𝑡 (𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 (𝑡)) is estimated, 𝑄𝑓 𝑒𝑑 (𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 (𝑡)) The computation of 𝑄 value updates in (2) and (4) is in time
immediately updates accordingly: complexity 𝑂(1). The decision update in (3) is in a linear time
𝑄𝑓 𝑒𝑑 (𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 (𝑡)) ← 𝑄𝑛𝑒𝑡 (𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 (𝑡)). (4) to the number of tabulated 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 decisions.

Then, each PON-MEC can refer to the shared 𝑄𝑓 𝑒𝑑 , instead of its Next, we elaborate on the conditions for experience sharing among
local 𝑄𝑛𝑒𝑡 , to update its 𝐵𝑊𝐵𝑁𝑂𝑈 𝑆 decision using (3). Here, adjusting PON-MECs in FedRL.
𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 decision based on the federated 𝑄𝑓 𝑒𝑑 is the major difference
between FedRL and pure local learning without FL. In summary, the Fe- 4.3.3. Key features for federated PON-MEC
dRL alternately exploits and updates a local 𝑄𝑛𝑒𝑡 and the federated 𝑄𝑓 𝑒𝑑 In FedRL, the heuristic sharing in (4) is efficient when the federated
to adjust 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 decision in individual PON-MECs. The iterative PON-MECs are homogeneous. This is because FL typically expects
learning and decision process involves the following key steps: homogeneous local learning environments for global performance and
convergence guarantee [88]. When the local learning environments are
• Local 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 decision is adjusted based on the federated 𝑄𝑓 𝑒𝑑 heterogeneous, the FL may not converge to the correct decision needed
in 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 ← arg max𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 ∈ 𝑄𝑓 𝑒𝑑 . in each local learning and the convergence rate may also be negatively
• Local 𝑄𝑛𝑒𝑡 is updated based on the reward received from execut- affected [89].
ing 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 decision. In our context, the association between 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 decision in a
• Federated 𝑄𝑓 𝑒𝑑 is updated based on local 𝑄𝑛𝑒𝑡 updates. The above PON-MEC and the corresponding latency performance is mainly related
process repeats. to the following network features based on studies in [81,90]:

7
I. Dias et al. Optical Fiber Technology 75 (2023) 103191

Fig. 5. Latency performance of tactile-haptic traffic (bursty traffic).

• number of ONUs;
• maximum OLT-ONU link distance;
• link data rate;
• traffic load; and
• packet statistics, i.e., length and arrival pattern.

In this paper, we consider PON-MEC networks having the same features

as homogeneous and are suitable to share decision experiences in FL.
Heterogeneous PON-MECs in general demand different 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 allo-
cation in reducing latency. Hence, aggregating their 𝑄 values in FL may
degrade the performance of local learning in individual networks. There
are sophisticated methods such as model distillation and clustering to
Fig. 6. Performance of FedRL with exponential UDP traffic.
deal with FL in heterogeneous environments [89,91,92], which are
beyond the scope of this paper.

results in Fig. 5 show that faster convergence is obtained with a higher

4.4. Performance evaluation
the number of federated PON-MECs, 𝑁. Fig. 5 showcases the time-
savings using FedRL to obtain an optimum 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 and stabilize
This section evaluates the performance of the proposed FedRL in
latency performance as compared to when only using local 𝑄-learning.
reducing latency and expedited learning via FL. We also illustrate the Increasing the number of networks expedited learning. A federation of
FedRL performance considering different PON-MEC features. 𝑁 = 2 PON-MECs saves above 40% time and a federation of 𝑁 = 4
PON-MECs saves around 60% time for the average network latency to
4.4.1. FedRL performance reduce to the minimum. With 𝑁 = 10 PON-MECs, 75% time-savings is
We perform simulations of the illustrative federated PON-MEC sce- achieved.
nario shown in Fig. 4, whereby the maximum number 𝑁 of federated With UDP traffic, Figs. 6(a) and 6(b) plot the average latency
10G PON-MECs is 10. Further, each PON-MEC supports 16 ONUs and performance under different traffic loads ranging from 0.1–0.7 and
the CO and ONU distance is 10 km for illustrative purposes. The the convergence time under different 𝑁, respectively. In Fig. 6(a), the
decision set  of 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 comprises 150 decisions ranging from 0 to latency performance of FedRL is shown to be superior to those of the
30 kB. The reward is reported every 10 polling cycles, i.e., one learning baseline and linear credit schemes. With FedRL, bandwidth decisions
iteration. The time to process the REPORT message together with a are optimized by exploring and experiencing decisions over time. At
reported reward 𝑅(𝑡), and to generate the GRANT message is 10 μs [93]. low to medium traffic loads, the latency of FedRL is approximately
The guard time between two adjacent ONUs’ transmissions is 1 μs. 50 μs, thus yielding a significant 60% improvement over the baseline
We consider two types of traffic over the PON-MECs to validate latency. In Fig. 6(b), the blue column highlights the convergence time-
savings of FedRL as compared to using local 𝑄-learning in learning
the effectiveness of FedRL: (a) tactile-haptic traffic that are based on
an optimal 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 decision. It can be observed that the higher
those collected from our H2M experiments; and (b) exponential user
the number of federated networks that share experiences, the higher
datagram protocol (UDP) traffic of a typical Internet environment [79].
the learning rate. A federation of 𝑁 = 5 PON-MECs achieves a time-
Based on our experimental traces, the packet arrivals of the tactile-
savings of 80%. When 𝑁 = 10, the time savings is up to 90%. On the
haptic traffic was modeled using a Generalized Pareto distribution with other hand, the marginal benefit of joining a federation diminishes as
𝛼 = 0.9 and 𝛾 = 0.01 values [69,85]. the number of PON-MECs increases. This can be observed through the
Fig. 5 shows the average uplink latency performance of tactile- orange columns representing the time-saving increments by adding a
haptic traffic. The latency performance of FedRL is compared with new PON-MEC to the federation. For 𝑁 > 5, the time-savings is less
those obtained using the baseline (zero 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 ) and linear credit than 5% as compared to a federation of 𝑁 −1 networks. Overall, Figs. 5
(prediction via linear regression) schemes [79]. These two schemes and 6 demonstrate that even at low 𝑁, federated PON-MECs sharing
result in a latency of more than 110 μs while FedRL achieves a latency of a 𝑄𝑓 𝑒𝑑 can achieve significant time-savings as compared to individual
∼70 μs, a 40% reduction as compared to the baseline scheme. Moreover, PON-MECs that use local 𝑄𝑛𝑒𝑡,𝑛 .

8
I. Dias et al. Optical Fiber Technology 75 (2023) 103191

Fig. 7. Performance with different ONU numbers (exponential traffic load 0.5 and N Fig. 8. Performance with different maximum link distances (exponential traffic load
= 4 in FedRL). 0.5 and N = 4 in FedRL).

4.4.2. Performance under different PON-MEC features allows more packets to avoid the round-trip report-grant process and
To show how the features of PON-MEC impact the learning, Figs. 7 thus reduces the latency. Results in Figs. 7 and 8 reveal that 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆
and 8 illustrate the latency performance and the 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 deci- decisions and their corresponding latency values are different depend-
sions learnt in FedRL considering learning environments with different ing on the network features of PON-MEC. For homogeneous PON-MECs
numbers of ONUs and the maximum link distances, respectively. For considered in this work, the FedRL convergence rate primarily depends
illustrative purposes, we consider an exponential UDP traffic load of on how many PON-MECs to share their decision experiences instead of
0.5 and 𝑁 = 4 PON-MECs. the detailed features of the federated PON-MECs.
Fig. 7(a) shows that when the number of ONUs increases, the
final network latency increases. This is intuitive as more ONUs com-
5. Summary
pete for bandwidth for uplink transmission will incur a longer packet
waiting time at each ONU. Nonetheless, the convergence performance
The surge in latency-sensitive and computation-intensive applica-
of FedRL is shown to be similar under different ONU numbers. In
tions due to the proliferation of 5GB, Tactile Internet and H2M/R
around 100 interactions, the average latency is lowered and stabilized
communication in conjunction with the exponential growth in con-
in the presented cases in Fig. 7(a). Further, Fig. 7(b) details the final
𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 decision learnt and allocated by the FedRL. First, the opti- nected devices, have pushed for architectural and capacity changes
mal 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 to allocate to the ONUs reduces when there are more in both wired and wireless networks. To this end, we have discussed
ONUs in a PON-MEC. For example, in an 8-ONU PON-MEC, each ONU the evolution of high-capacity PONs and MEC, and as importantly
is granted 13 kB extra bandwidth on average to lower the latency to how their inherent characteristics help in supporting latency-sensitive
45 μs. In comparison, in a 64-ONU PON-MEC, the 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 is only applications. Example use case scenarios that can benefit from inte-
around 3 kB to the ONUs. The latency value corresponding to this grated PON-MECs have also been detailed. As the integration of these
𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 decision is 180 μs, which is significantly higher than the two architectures does not guarantee that latency requirements will
8-ONU case. be met, end-to-end resource allocation across these heterogeneous net-
Fig. 8 presents the impact of the maximum link distance in a PON- works has to be carefully managed through dynamic resource allocation
MEC on the 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 decision and latency performance. Similar to algorithms. Specifically to support latency-sensitive H2M/R communi-
Figs. 7(a), 8(a) shows the dependency of network latency on link cations over PON-MECs, we proposed a framework based on federated
distance although the convergence performance of FedRL is minimally reinforcement learning that harnesses reinforcement learning to opti-
affected by it. In Fig. 8(b), we see that the 𝐵𝑊𝐵𝑂𝑁𝑈 𝑆 granted to ONUs mize bandwidth allocation decision together with shared learning be-
increases as the maximum link distance in a PON-MEC is increased. tween multiple PON-MECs to maximize time-savings in reaching such
This is because a longer distance causes a longer round-trip commu- decisions, further improving latency performance. Results show that the
nication time between the CO and ONUs. Granting more bandwidth proposed framework achieves significant time-savings as compared to

9
I. Dias et al. Optical Fiber Technology 75 (2023) 103191

the baseline and linear credit schemes. With the investigation of the [16] S. Lee, S. Lee, M.-K. Shin, Low cost MEC server placement and association in 5G
framework for PON-MEC resource allocation showing overwhelming networks, in: 2019 International Conference on Information and Communication
Technology Convergence (ICTC), 2019, pp. 879–882, http://dx.doi.org/10.1109/
benefit in supporting latency-sensitive applications, our future research
ICTC46691.2019.8939566.
focus will be on the exploration of suitable machine learning techniques [17] C. Ranaweera, P. Iannone, K. Oikonomou, K. Reichmann, R. Sinha, Cost optimiza-
to accelerate reinforcement learning decisions within each PON-MEC tion of fiber deployment for small cell backhaul, in: Optical Fiber Communication
and the identification of which specific PON-MECs to form a federation. Conference/National Fiber Optic Engineers Conference 2013, Optica Publishing
Group, 2013, p. NTh3F.2, http://dx.doi.org/10.1364/NFOEC.2013.NTh3F.2, URL
http://opg.optica.org/abstract.cfm?URI=NFOEC-2013-NTh3F.2.
CRediT authorship contribution statement
[18] E. Wong, Next-generation broadband access networks and technologies, J.
Lightwave Technol. 30 (4) (2012) 597–608, http://dx.doi.org/10.1109/JLT.2011.
Imali Dias: Writing – original draft, Investigation, Writing – review 2177960.
& editing. Lihua Ruan: Software, Experiment, Data collection, Vali- [19] Fiber and Copper Access Equipment Forecast: 2021–27, OMDIA, 2022,
dation, Writing – review & editing. Chathurika Ranaweera: Writing, [Online] https://Omdia.Tech.Informa.Com/OM022089/Fiber-and-Copper-
Access-Equipment-Forecast-202127.
Investigation, Writing – review & editing. Elaine Wong: Conceptual-
[20] Global FTTX Market to Reach $22.4B by 2026, BroadbandWorldNews,
ization, Writing – review & editing. 2021, [Online] https://Www.Broadbandworldnews.Com/Document.Asp?Doc_id=
771259.
Declaration of competing interest [21] Gigabit-capable Passive Optical Networks (G-PON), in: ITU-T Recommendation
G.984 Series, 2008.
The authors declare that they have no known competing finan- [22] IEEE Standard for Local and Metropolitan Area Networks - Specific requirements
- Part 3: Carrier Sense Multiple Access with Collision Detection (CSMA/CD)
cial interests or personal relationships that could have appeared to
Access Method and Physical Layer Specifications- Amendment: Media Access
influence the work reported in this paper. Control Parameters, Physical Layers, and Management Parameters for Subscriber
Access Networks, IEEE Standard 802.3ah - 2004, 2004.
Data availability [23] IEEE Standard for Information technology– Local and Metropolitan Area
Networks– Specific requirements– Part 3: CSMA/CD Access Method and Physical
Layer Specifications Amendment 1: Physical Layer Specifications and Manage-
The data that has been used is confidential.
ment Parameters for 10 Gb/s Passive Optical Networks., IEEE 802.3av-2009,
2009.
References [24] 10-Gigabit-capable Passive Optical Network (XG-PON) Systems: Definitions,
Abbreviations and Acronyms, in: ITU-T Recommendation G.987 Series, 2012.
[1] A. Zanella, N. Bui, A. Castellani, L. Vangelista, M. Zorzi, Internet of things for [25] 10-Gigabit-capable Symmetric Passive Optical Networks (XGS-PON), in: ITU-T
smart cities, IEEE Internet Things J. 1 (1) (2014) 22–32, http://dx.doi.org/10. Recommendation G.9807.1 Series, 2016.
1109/JIOT.2014.2306328. [26] 40-Gigabit-capable Passive Optical Networks (NG-PON2): Definitions, Abbrevia-
[2] M. Agiwal, A. Roy, N. Saxena, Next generation 5G wireless networks: A tions and Acronyms, in: ITU-T Recommendation G.989 Series, 2015.
comprehensive survey, IEEE Commun. Surv. Tutor. 18 (3) (2016) 1617–1655, [27] IEEE Standard for Ethernet Amendment 9: Physical Layer Specifications and
http://dx.doi.org/10.1109/COMST.2016.2532458. Management Parameters for 25 Gb/s and 50 Gb/s Passive Optical Networks.,
[3] M. Maier, A. Ebrahimzadeh, Towards immersive tactile internet experiences: IEEE 802.3ca-2020, 2020.
Low-latency FiWi enhanced mobile networks with edge intelligence [Invited], J. [28] 50-Gigabit-capable Passive Optical Networks (50G-PON): Physical Media
Opt. Commun. Netw. 11 (4) (2019) B10–B25, http://dx.doi.org/10.1364/JOCN. Dependent (PMD) Layer specification, ITU-T Recommendation G.9804.3, 2021.
11.000B10. [29] J. Liu, G. Shou, Y. Liu, Y. Hu, Z. Guo, Performance evaluation of integrated
[4] E. Wong, M. Pubudini Imali Dias, L. Ruan, Predictive resource allocation for multi-access edge computing and fiber-wireless access networks, IEEE Access 6
tactile internet capable passive optical LANs, J. Lightwave Technol. 35 (13) (2018) 30269–30279, http://dx.doi.org/10.1109/ACCESS.2018.2833619.
(2017) 2629–2641, http://dx.doi.org/10.1109/JLT.2017.2654365. [30] Q.-V. Pham, F. Fang, V.N. Ha, M.J. Piran, M. Le, L.B. Le, W.-J. Hwang, Z. Ding,
[5] N. Al-Falahy, O.Y. Alani, Technologies for 5G networks: Challenges and oppor- A survey of multi-access edge computing in 5G and beyond: Fundamentals, tech-
tunities, IT Prof. 19 (1) (2017) 12–20, http://dx.doi.org/10.1109/MITP.2017. nology integration, and state-of-the-art, IEEE Access 8 (2020) 116974–117017,
9. http://dx.doi.org/10.1109/ACCESS.2020.3001277.
[6] L.D. Xu, W. He, S. Li, Internet of things in industries: A survey, IEEE Trans. Ind.
[31] Y.C. Hu, M. Patel, D. Sabella, N. Sprecher, V. Young, Mobile edge computing: A
Inform. 10 (4) (2014) 2233–2243, http://dx.doi.org/10.1109/TII.2014.2300753.
key technology towards 5G, WhitePaper, ETSI, 2015.
[7] S.M.R. Islam, D. Kwak, M.H. Kabir, M. Hossain, K.-S. Kwak, The internet of
[32] X. Fu, F.R. Yu, J. Wang, Q. Qi, J. Liao, Resource allocation for blockchain-
things for health care: A comprehensive survey, IEEE Access 3 (2015) 678–708,
enabled distributed network function virtualization (NFV) with mobile edge
http://dx.doi.org/10.1109/ACCESS.2015.2437951.
cloud (MEC), in: IEEE INFOCOM 2019 - IEEE Conference on Computer
[8] M.P.I. Dias, E. Grigoreva, C.M. Machuca, L. Wosinska, E. Wong, Delay-
Communications Workshops (INFOCOM WKSHPS), 2019, pp. 1–6.
constrained framework for road safety and energy-efficient intelligent transporta-
[33] B.P. Rimal, D. Pham Van, M. Maier, Mobile-edge computing versus centralized
tion systems, in: 2017 European Conference on Optical Communication (ECOC),
cloud computing over a converged FiWi access network, IEEE Trans. Netw. Serv.
2017, pp. 1–3, http://dx.doi.org/10.1109/ECOC.2017.8346010.
Manag. 14 (3) (2017) 498–513, http://dx.doi.org/10.1109/TNSM.2017.2706085.
[9] T. Taleb, K. Samdanis, B. Mada, H. Flinck, S. Dutta, D. Sabella, On multi-access
[34] T. Wang, S. Jie, B. Jiang, S.K. Bose, G. Shen, Distributed backup resource
edge computing: A survey of the emerging 5G network edge cloud architecture
allocation in fiber-wireless (FiWi) access networks supporting mobile edge com-
and orchestration, IEEE Commun. Surv. Tutor. 19 (3) (2017) 1657–1681, http:
puting, in: 2019 21st International Conference on Transparent Optical Networks
//dx.doi.org/10.1109/COMST.2017.2705720.
(ICTON), 2019, pp. 1–4, http://dx.doi.org/10.1109/ICTON.2019.8840327.
[10] Cisco Annual Internet Report (2018–2023), White Paper, 2020, [Online]
https://www.cisco.com/c/en/us/solutions/collateral/executive-perspectives/ [35] I. Morris, ETSI Drops ‘‘Mobile’’ From MEC, Light Reading, New York, NY, USA,
annual-internet-report/whitepaper-c11-741490.html. 2016.
[11] Multi Access Edge Computing (MEC); Phase2: Use Cases and Requirements, [36] J. Vora, S. Kaneriya, S. Tanwar, S. Tyagi, N. Kumar, M. Obaidat, TILAA: Tactile
Standard ETSI GS MEC 002 V2.1.1, 2018. internet-based ambient assistant living in fog environment, Future Gener. Com-
[12] Z. Zou, Y. Jin, P. Nevalainen, Y. Huan, J. Heikkonen, T. Westerlund, Edge and put. Syst. 98 (2019) 635–649, http://dx.doi.org/10.1016/j.future.2019.01.035,
fog computing enabled AI for IoT-An overview, in: 2019 IEEE International URL https://www.sciencedirect.com/science/article/pii/S0167739X18326839.
Conference on Artificial Intelligence Circuits and Systems (AICAS), 2019, pp. [37] C. Ranaweera, E. Wong, A. Nirmalathas, C. Jayasundara, C. Lim, 5G C-RAN
51–56, http://dx.doi.org/10.1109/AICAS.2019.8771621. architecture: A comparison of multiple optical fronthaul networks, in: 2017
[13] X. Wang, Y. Ji, J. Zhang, L. Bai, M. Zhang, Low-latency oriented network International Conference on Optical Network Design and Modeling (ONDM),
planning for MEC-enabled WDM-PON based fiber-wireless access networks, 2017, pp. 1–6, http://dx.doi.org/10.23919/ONDM.2017.7958544.
IEEE Access 7 (2019) 183383–183395, http://dx.doi.org/10.1109/ACCESS.2019. [38] H. Guo, J. Liu, Collaborative computation offloading for multiaccess edge
2926795. computing over fiber–wireless networks, IEEE Trans. Veh. Technol. 67 (5) (2018)
[14] J. Zhang, W. Xie, F. Yang, Q. Bi, Mobile edge computing and field trial results 4514–4526, http://dx.doi.org/10.1109/TVT.2018.2790421.
for 5G low latency scenario, China Commun. 13 (Supplement2) (2016) 174–182, [39] M. Giordani, M. Polese, M. Mezzavilla, S. Rangan, M. Zorzi, Toward 6G networks:
http://dx.doi.org/10.1109/CC.2016.7833471. Use cases and technologies, IEEE Commun. Mag. 58 (3) (2020) 55–61.
[15] Y. Yu, Mobile edge computing towards 5G: Vision, recent progress, and open [40] C. Ranaweera, J. Kua, I. Dias, E. Wong, C. Lim, A. Nirmalathas, 4G to 6G:
challenges, China Commun. 13 (Supplement2) (2016) 89–99, http://dx.doi.org/ Disruptions and drivers for optical access [Invited], J. Opt. Commun. Netw. 14
10.1109/CC.2016.7833463. (2) (2022) A143–A153, http://dx.doi.org/10.1364/JOCN.440798.

10
I. Dias et al. Optical Fiber Technology 75 (2023) 103191

[41] C. Ranaweera, A. Nirmalathas, E. Wong, C. Lim, P. Monti, M. Furdek, L. [65] P. Schulz, M. Matthe, H. Klessig, M. Simsek, G. Fettweis, J. Ansari, S.A. Ashraf,
Wosinska, B. Skubic, C.M. Machuca, Rethinking of optical transport network B. Almeroth, J. Voigt, I. Riedel, A. Puschmann, A. Mitschele-Thiel, M. Muller, T.
design for 5G/6G mobile communication, IEEE Future Netw. (12) (2021). Elste, M. Windisch, Latency critical IoT applications in 5G: Perspective on the
[42] C. Ranaweera, P. Monti, B. Skubic, E. Wong, M. Furdek, L. Wosinska, C.M. design of radio interface and network architecture, IEEE Commun. Mag. 55 (2)
Machuca, A. Nirmalathas, C. Lim, Optical transport network design for 5G (2017) 70–78, http://dx.doi.org/10.1109/MCOM.2017.1600435CM.
fixed wireless access, J. Lightwave Technol. 37 (16) (2019) 3893–3901, http: [66] X. Jiang, F. Yu, T. Song, V. Leung, A survey on multi-access edge computing
//dx.doi.org/10.1109/JLT.2019.2921378. applied to video streaming: Some research issues and challenges, IEEE Commun.
[43] China Mobile Research Institute, C-RAN: The road towards green RAN, 2013, Surv. Tutor. 23 (2) (2021) 871–903.
URL whitepaper. [67] S. de Ribaupierre, R. Armstrong, D. Noltie, M. Kramers, R. Eagleson, VR and AR
[44] B. Skubic, M. Fiorani, S. Tombaz, A. Furuskär, J. Mårtensson, P. Monti, Optical simulator for neurosurgical training, in: 2015 IEEE Virtual Reality (VR), 2015,
transport solutions for 5G fixed wireless access [Invited], J. Opt. Commun. Netw. pp. 147–148, http://dx.doi.org/10.1109/VR.2015.7223338.
9 (9) (2017) D10–D18. [68] M. Maier, A. Ebrahimzadeh, A. Beniiche, S. Rostami, The art of 6G (TAO
[45] T. 3GPP, Radio access architecture and interfaces (release 14), 2016. 6G): How to wire society 5.0 [Invited], J. Opt. Commun. Netw. 14 (2) (2022)
[46] J.S. Wey, Y. Luo, T. Pfeiffer, 5G wireless transport in a PON context: An A101–A112, http://dx.doi.org/10.1364/JOCN.438522.
overview, IEEE Commun. Stand. Mag. 4 (1) (2020) 50–56, http://dx.doi.org/ [69] L. Ruan, M.P.I. Dias, E. Wong, Achieving low-latency human-to-machine (H2M)
10.1109/MCOMSTD.001.1900043. applications: An understanding of H2M traffic for AI-facilitated bandwidth
[47] J.S. Wey, J. Zhang, Passive optical networks for 5G transport: Technology and allocation, IEEE Internet Things J. 8 (1) (2021) 626–635, http://dx.doi.org/10.
standards, J. Lightwave Technol. 37 (12) (2019) 2830–2837, http://dx.doi.org/ 1109/JIOT.2020.3007947.
10.1109/JLT.2018.2856828. [70] A. Ebrahimzadeh, M. Maier, Delay-constrained teleoperation task scheduling and
assignment for human+machine hybrid activities over FiWi enhanced networks,
[48] ITU-t g.sup66 : 5G wireless fronthaul requirements in a passive optical network,
IEEE Trans. Netw. Serv. Manag. 16 (4) (2019) 1840–1854, http://dx.doi.org/10.
2020.
1109/TNSM.2019.2937020.
[49] O-RAN alliance: O-RAN, 2020, [Online] https://www.o-ran.org.
[71] S. Mondal, L. Ruan, M. Maier, D. Larrabeiti, G. Das, E. Wong, Enabling remote
[50] C. Ranaweera, P. Monti, B. Skubic, M. Furdek, L. Wosinska, A. Nirmalathas,
human-to-machine applications with AI-enhanced servers over access networks,
C. Lim, E. Wong, Optical X-haul options for 5G fixed wireless access: Which
IEEE Open J. Commun. Soc. 1 (2020) 889–899, http://dx.doi.org/10.1109/
one to choose? in: IEEE INFOCOM 2018 - IEEE Conference on Computer
OJCOMS.2020.3009023.
Communications Workshops (INFOCOM WKSHPS), 2018, pp. 1–2, http://dx.doi.
[72] X. Xu, H. Singh, Q. Liu, M. Panzirsch, T. Hulin, E. Steinbach, A novel energy
org/10.1109/INFCOMW.2018.8406906.
compensation scheme for quality enhancement in time-delayed teleoperation
[51] C. Ranaweera, E. Wong, C. Lim, C. Jayasundara, A. Nirmalathas, Optimal design
with multi-dof haptic data reduction and communication, IEEE Trans. Haptics
and backhauling of small-cell network: Implication of energy cost, in: 2016 21st
14 (4) (2021) 936–944, http://dx.doi.org/10.1109/TOH.2021.3103379.
OptoElectronics and Communications Conference (OECC) Held Jointly with 2016
[73] M. Condoluci, T. Mahmoodi, E. Steinbach, M. Dohler, Soft resource reservation
International Conference on Photonics in Switching (PS), 2016, pp. 1–3.
for low-delayed teleoperation over mobile networks, IEEE Access 5 (2017)
[52] H. Uzawa, K. Honda, H. Nakamura, Y. Hirano, K. ichi Nakura, S. Kozaki,
10445–10455, http://dx.doi.org/10.1109/ACCESS.2017.2707319.
J. Terada, Dynamic bandwidth allocation scheme for network-slicing-based
[74] X. Xu, L.-Y.D. Liu, E.G. Steinbach, Toward qoe-driven dynamic control scheme
TDM-PON toward the beyond-5G era, J. Opt. Commun. Netw. 12 (2) (2020)
switching for time-delayed tele-operation systems: A dedicated case study, in:
A135–A143.
2017 IEEE International Symposium on Haptic, Audio and Visual Environments
[53] J. Zhang, X. Liu, M. Liang, H. Yu, Y. Ji, Low latency DWBA scheme for mini-slot and Games (HAVE), 2017, pp. 1–6.
based 5G new radio in a fixed and mobile converged TWDM-PON, J. Lightwave
[75] T. Thangappan, B. Therese, A. Suvarnamma, G.S. Swapna, Review on dynamic
Technol. 40 (1) (2022) 3–13, http://dx.doi.org/10.1109/JLT.2021.3117972.
bandwidth allocation of GPON and EPON, J. Electron. Sci. Technol. 18 (4) (2020)
[54] Q. Yuan, L. Wang, L. Wang, X. Chen, Traffic aware wavelength and resource 100044, http://dx.doi.org/10.1016/j.jnlest.2020.100044.
allocation scheme for TWDM-PON based 5G fronthaul, in: Asia Communications [76] G. Kramer, Ethernet Passive Optical Networks, McGraw-Hill, New York, NY, USA,
and Photonics Conference, Optica Publishing Group, 2017, p. M3B.3. 2005.
[55] Y. Liu, C. Ranaweera, C. Lim, L. Guo, A. Nirmalathas, E. Wong, Convergence [77] K.A. Memon, K.H. Mohammadani, A.A. Laghari, R. Yadav, B. Das, W.U. Khan
of 5G RAN and optical access: A coordinated resource allocation framework, Tareen, N. ul Ain Memon, X. Xin, Dynamic bandwidth allocation algorithm
in: 2018 Asia Communications and Photonics Conference (ACP), 2018, pp. 1–3, with demand forecasting mechanism for bandwidth allocations in 10-gigabit-
http://dx.doi.org/10.1109/ACP.2018.8595806. capable passive optical network, Optik 183 (2019) 1032–1042, http://dx.doi.
[56] D. Hisano, Y. Nakayama, Two-stage optimization of uplink forwarding order org/10.1016/j.ijleo.2019.03.003.
with cooperative DBA to accommodate a TDM-PON-based fronthaul link, J. Opt. [78] M.P.I. Dias, B.S. Karunaratne, E. Wong, Bayesian estimation and prediction-based
Commun. Netw. 12 (5) (2020) 109–119. dynamic bandwidth allocation algorithm for sleep/doze-mode passive optical
[57] P. Ranaweera, M. Liyanage, A.D. Jurcut, Novel MEC based approaches for smart networks, J. Lightwave Technol. 32 (14) (2014) 2560–2568, http://dx.doi.org/
hospitals to combat COVID-19 pandemic, IEEE Consum. Electron. Mag. 10 (2) 10.1109/JLT.2014.2327629.
(2021) 80–91, http://dx.doi.org/10.1109/MCE.2020.3031261. [79] N. Hanaya, Y. Nakayama, M. Yoshino, K.-I. Suzuki, R. Kubo, Remotely controlled
[58] M. Chowdhury, M. Maier, Toward dynamic HART-centric task offloading over XG-PON DBA with linear prediction for flexible access system architecture, in:
FiWi infrastructures in the tactile internet era, IEEE Commun. Mag. 57 (11) 2018 Optical Fiber Communications Conference and Exposition (OFC), 2018, pp.
(2019) 123–128, http://dx.doi.org/10.1109/MCOM.001.1800669. 1–3.
[59] F. Boabang, A. Ebrahimzadeh, R.H. Glitho, H. Elbiaze, M. Maier, F. Belqasmi, [80] Y. Luo, N. Ansari, Limited sharing with traffic prediction for dynamic bandwidth
A machine learning framework for handling delayed/lost packets in tactile allocation and QoS provisioning over ethernet passive optical networks, J. Opt.
internet remote robotic surgery, IEEE Trans. Netw. Serv. Manag. 18 (4) (2021) Netw. 4 (9) (2005) 561–572, http://dx.doi.org/10.1364/JON.4.000561.
4829–4845, http://dx.doi.org/10.1109/TNSM.2021.3106577. [81] L. Ruan, M.P.I. Dias, E. Wong, Enhancing latency performance through intelligent
[60] L. Ruan, M.P.I. Dias, E. Wong, Deep neural network supervised bandwidth bandwidth allocation decisions: A survey and comparative study of machine
allocation decisions for low-latency heterogeneous E-health networks, J. Light- learning techniques, J. Opt. Commun. Netw. 12 (4) (2020) B20–B32, http:
wave Technol. 37 (16) (2019) 4147–4154, http://dx.doi.org/10.1109/JLT.2019. //dx.doi.org/10.1364/JOCN.379715.
2928821. [82] J.A. Hatem, A.R. Dhaini, S. Elbassuoni, Deep learning-based dynamic bandwidth
[61] F. Boabang, R. Glitho, H. Elbiaze, F. Belqami, O. Alfandi, A framework for allocation for future optical access networks, IEEE Access 7 (2019) 97307–97318,
predicting haptic feedback in needle insertion in 5G remote robotic surgery, http://dx.doi.org/10.1109/ACCESS.2019.2929480.
in: 2020 IEEE 17th Annual Consumer Communications Networking Conference [83] J. Yu, Y. Fu, F. Xiao, H. Jia, P. Qin, Z. Zhao, J. You, F. Liu, S. Han, J. Wang,
(CCNC), 2020, pp. 1–6, http://dx.doi.org/10.1109/CCNC46108.2020.9045432. RNN based EPON dynamic bandwidth allocation algorithm for complex network,
[62] T.M. Fernández-Caramés, P. Fraga-Lamas, A review on the application of J. Opt. Commun. (2022) http://dx.doi.org/10.1515/joc-2021-0250.
blockchain to the next generation of cybersecure industry 4.0 smart factories, [84] Y. Feng, L. Ruan, A. Nirmalathas, E. Wong, Intelligent radio resource allocation
IEEE Access 7 (2019) 45201–45218, http://dx.doi.org/10.1109/ACCESS.2019. for human-robot collaboration, IEEE Open J. Commun. Soc. 3 (2022) 144–158,
2908780. http://dx.doi.org/10.1109/OJCOMS.2022.3141201.
[63] B. Bajic, N. Suzic, N. Simeunovic, S. Moraca, A. Rikalovic, Real-time data [85] L. Ruan, S. Mondal, I. Dias, E. Wong, Low-latency federated reinforcement
analytics edge computing application for industry 4.0: The Mahalanobis-Taguchi learning-based resource allocation in converged access networks, in: 2020 Optical
approach, Int. J. Ind. Eng. Manage. 11 (3) (2020) 146–156. Fiber Communications Conference and Exhibition (OFC), 2020, pp. 1–3.
[64] M.-P. Pacaux-Lemoine, D. Trentesaux, G. Zambrano Rey, P. Millot, Designing [86] R.S. Sutton, A.G. Barto, Reinforcement Learning: An Introduction, The MIT Press,
intelligent manufacturing systems through human-machine cooperation princi- 2018.
ples: A human-centered approach, Comput. Ind. Eng. 111 (2017) 581–595, http: [87] H.B. McMahan, E. Moore, D. Ramage, S. Hampson, B.A. Arcas, Communication-
//dx.doi.org/10.1016/j.cie.2017.05.014, URL https://www.sciencedirect.com/ efficient learning of deep networks from decentralized data, in: AISTATS,
science/article/pii/S0360835217302188. 2017.

11
I. Dias et al. Optical Fiber Technology 75 (2023) 103191

[88] L. Buşoniu, R. Babuška, B.D. Schutter, Multi-agent reinforcement learning: An Lihua Ruan (S’16-M’20) received the Ph.D. degree from the University of Melbourne,
overview, Innov. Multi-Agent Syst. Appl.-1 (2010) 183–221. Australia, in 2020. She is currently a Postdoctoral Fellow of the Chinese University of
[89] M. Zaheer, M. Sanjabi, A. Talwalkar, T. Li, A. Sahu, V. Smith, Federated Hong Kong, Shenzhen. She has coauthored more than 30 papers. Her research interests
optimization in heterogeneous networks, in: Machine Learning and Systems, include low-latency communications and networking, and machine learning-assisted
2020, pp. 429–450. low-latency applications. She was a recipient of the Melbourne University’s John Collier
[90] I.D. L. Ruan, E. Wong, Machine intelligence in supervising bandwidth allocation Scholarship and the John Melvin Scholarship for the Best Ph.D. Thesis in Engineering
for low-latency communications, in: IEEE 20th International Conference on High and IT.
Performance Switching and Routing (HPSR), 2019, pp. 1–6.
[91] J.H. Z. Zhu, J. Zhou, Data-free knowledge distillation for heterogeneous federated
learning, in: The International Conference on Machine Learning (PMLR), 2021, Chathurika Ranaweera received B.Sc. and Ph.D. degrees from the University of
pp. 12878–12889. Peradeniya, Sri Lanka, and the University of Melbourne, Australia, respectively. She is
[92] D.Y. A. Ghosh, K. Ramchandran, An efficient framework for clustered federated currently a Senior Lecturer and Course Director of Master of Information Technology
learning, in: The 34th International Conference on Neural Information Processing at Deakin University, Melbourne, Australia. She has published more than 50 papers
Systems (NIPS’20), 2020, pp. 19586–19597. in prestigious journals and international conferences so far. Her research interests
[93] R.B. S. Bidkar, T. Pfeiffer, Low-latency TDM-pon for 5G Xhaul, in: The 22nd include network optimization, optical wireless convergence, optical transport network
International Conference on Transparent Optical Networks (ICTON), 2020, pp. architectures, secure IoT connectivity, quality of service management, energy efficient
1–4. communications and networks, smart grid communication, and machine learning
applications in networking.

Imali Dias received the B.Sc. degree (Hons.) from the University of Moratuwa, Elaine Wong (M’02–SM’14) received B.E. and Ph.D. degrees from the University of
Moratuwa, Sri Lanka, in 2007, and the Ph.D. from the University of Melbourne in 2016. Melbourne, Australia. She is currently Associate Dean and Professor at the Faculty of
She has published her work in multiple peer-reviewed journals and conferences. Her Engineering and Information Technology, University of Melbourne. Her research inter-
current research interests include Internet of Things, Tactile Internet, optical wireless ests include energy-efficient optical and wireless networks, optical-wireless integration,
convergence, dynamic bandwidth allocation algorithms, passive optical networks, broadband applications of vertical-cavity surface-emitting lasers, wireless sensor body
Bayesian estimation and low-latency communication. area networks, and emerging optical and wireless technologies for human-to-machine
applications. She has coauthored more than 150 journal and conference publications.
She has served on the editorial board of the Journal of Lightwave Technology and the
Journal of Optical Communications and Networking.

12