Optical and Quantum Electronics (2024) 56:1075

https://doi.org/10.1007/s11082-024-07016-z

Capacity, resilience and virtual embedding in elastic optical

networks planning with adopted machine learning

Talison Augusto Correia de Melo1 · Alex Ferreira dos Santos2 ·

Raul Camelo de Andrade Almeida Jr.3 · Karcius Day Rosário Assis1

Received: 13 January 2024 / Accepted: 24 April 2024 / Published online: 13 May 2024
© The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2024

Abstract
This work presents supervised machine learning techniques for the problem of virtualiza-
tion design with protection over elastic optical networks (EONs) for predicting the total
number of used spectrum slots to support all traffic demands. It considers virtual optical
networks (VONs) subject to protection and proposes learning techniques to solve the link
capacity problem of EONs with virtualization faster than traditional integer linear program-
ming (ILP) formulations, but keeping the finds near to the optimal ones. The performance
of the models were evaluated using statistical metrics, along with the time for training and
performing inferences, both using quantitative and qualitative analysis. They showed that
the proposed method is effective for predicting the number of required slots (bandwidth) on
physical substrate subject to several VONs.

Keywords Elastic optical networks · Optimization · Protection · Virtualization · Machine

learning

1 Introduction

The optical network is the wired infrastructure present in nowadays telecommunication

infrastructure to support data transmission with the advantage of large transmission capac-
ity, small transmission loss, good anti-interference, strong confidentiality, high speed and
low energy consumption (Zhang et al. 2020, 2022). Optical network planning is important
to ensure that the network is capable of meeting users’ traffic and bandwidth demands. It
prevents corporations from being able to face failures, server outages and other daily prob-
lems without major losses.

* Karcius Day Rosário Assis

[email protected]
1
Department of Electrical Engineering, Federal University of Bahia, Salvador, Bahia, Brazil
2
Center of Science and Technology in Energy and Sustainability, Federal University of Reconcavo
of Bahia. Postgraduate Program in Computer Science, State University of Feira de Santana,
Feira de Santana, Bahia, Brazil
3
Photonics Group, Electronics and Systems Department, Technology Center and Geosciences,
Federal University of Pernambuco, Recife, Pernambuco, Brazil

13
Vol.:(0123456789)
1075 Page 2 of 21 T. A. C. de Melo et al.

Rather than using fixed wavelength channels as in traditional WDM (wavelength-divi-

sion multiplexing) networks, Elastic Optical Networks (EONs) use bands of variable band-
width waves that share the entire spectrum on one fiber (Assis et al. 2023). Furthermore,
it is possible to create virtual optical networks (VONs) on top of an optical network infra-
structure using network virtualization techniques.
The planning of EONs is important because it provides flexibility and scalability in
spectrum allocation that is superior to traditional networks based on wavelength division
multiplexing. This makes them suitable to support the increasing demand of Internet traffic
(Moura et al. 2020). Planning must take into account factors such as network topology, the
number of available slots, the bandwidth and the modulation format used.
In EONs, the problem of finding unoccupied spectrum resources so that a lightpath may
be established is called the Routing and Spectrum Assignment (RSA) problem. RSA con-
cerns assigning a contiguous fraction of frequency spectrum to a connection request - we
refer to it as the spectrum contiguity constraint–subject to the fact that the same spectrum
range has to be repeated in all links along the route (continuity constraint) and no fre-
quency overlapping in network links (Klinkowski 2013). The RSA optimization problem is
NP hard (Klinkowski and Walkowiak 2011).
EONs are not immune to events that can compromise communication between users,
such as fiber cuts, which can cause substantial data loss or even interruption of data trans-
mission. To diminish or even eliminate the likelihood of network outages, it is imperative
to incorporate survivability mechanisms that can restore network connectivity in the event
of failures. These mechanisms must be taken into account during the network design phase.
Without such measures, a single network failure, like a fiber cut, could lead to numerous
broken connections. This is because a specific physical resource may be shared among
multiple lightpaths, which, in turn, can disrupt the VON (Bui et al. 2013).
In terms of planning capacity and resiliency, it’s crucial to inquire about how the
EON substrate will support the composition of multiple isolated virtual optical networks
(VONs), described as slices, coexisting over the same physical optical network infrastruc-
ture simultaneously (Hammad 2013; Davalos and Baran 2018). The primary issue is to
design optical link virtualization by flexibly inserting virtual optical links (i.e. lightpaths)
onto a common set of spectrum slots in a physical optical infrastructure in such a way that
bandwidth requirements are satisfied under an efficient use of spectrum.
Developing models that allow each VON to select an appropriate resiliency or surviv-
ability mechanism has been a fine idea since bandwidth requirements vary for a wide range
of services (along with other requirements). The use of squeezing protection (Sone 2011)
for virtualization is a promising approach to ensure the survival of multiple VONs compet-
ing for resources.
Two proposed protection mechanisms that could be applied to an EON using virtual-
ization can be seen in Fig. 1: (1) dedicated protection (DP), Fig. 1a, (2) and/or dedicated
protection with squeezing (DP+S), Fig. 1b, Shahriar (2020); Assis (2016); Assis et al.
(2021). In DP+S, the backup path capacity can be limited to a fraction of the requested
working traffic by allowing traffic to be squeezed into the used protection mechanism. Fig-
ure 1b presents a case where the fraction of the requested traffic is kept at 50%, QoS =
0.5 (throughout this paper, the term QoS will refer to quality-of-service-aware protection),
under a network failure occurrence and these values are pre-accorded in the SLA (Service
Level Agreement). Therefore, a QoS = 0.5 means that 50% of traffic working is guaranteed
in backup path. This protection strategy could also be implemented for all lightpaths of
VON 2 or even for other VONs as well. Furthermore, Fig. 1c depicts the load generated
by VONs 1 and 2 on three physical links of the network. Slice is a term that encompasses

13
Capacity, resilience and virtual embedding in elastic optical… Page 3 of 21 1075

Fig. 1  Different protection mechanism examples for two VONs (red and blue one): a DP, b DP+S, c load
on physical links, Dias (2021)

many concepts (such as Spectrum-sliced Elastic Optical Path), but it is specifically used to
refer to VON in this paper.
The planning of EONs involves the consideration of various parameters and configura-
tions to optimize network performance. Conducting tests with different configurations is a
common practice to find the best design, taking into account factors such as spectral effi-
ciency, latency, scalability, and reliability. Observing the need for multiple tests, a model
could streamline the planning process by providing quick and accurate estimations for the
required number of slots based on varying network topologies. This efficiency can lead to
time savings and better decision-making in the design and optimization of EONs. Further-
more, this would avoid the need for each topology or design configuration of an EON to
require adaptation and execution of heuristics.
The issue of long computational running time is a significant consideration when
designing VONs models, especially when using the integer linear programming (ILP) for-
mulation from Dias (2021) for VONs design with protection. A multicommodity network
flow problem (Pióro and Medhi 2004) is a type of problem that involves multiple demands
(or commodities) that need to be routed in the network simultaneously and they compete
for available resources (for example, link capacities). Communication and computer net-
works frequently encounter this scenario, and the formulation cannot be practical for virtu-
alization planning in terms of running time, even small instances, as we have observed in
Dias (2021).
In order to attenuate this problem and contribute to the current state of the art, in this
paper we propose a supervised machine learning (ML) technique using a selection of pre-
computed sets of pair of link-disjoint paths p and protection levels for the traffic demands
of each VON, which form the so-called dataset. Essentially, a regression model which is
trained with several previously labeled p configurations (solved by a heuristic algorithm)
is capable of reporting near-optimal solution (in terms of maximum number of slots) much
faster than ILPs and Heuristics.
It is important to note that a large and growing body of literature has investigated the
management and planning of optical networks using machine learning techniques, Renjith
et al. (2023), Wang (2023). Despite this, to the best of our knowledge, this is the first study

13
1075 Page 4 of 21 T. A. C. de Melo et al.

that shows how to map several VONs with different protection levels that can be tackled
with ML models.
Under this scenario, the contributions of this paper are:

(a) We model through a ML technique a cost-effective multiple VONs survivability plan-

ning, over a single network substrate, while minimizing resources, in terms of the total
number of slots;
(b) The proposed model is suitable for large networks for which the original formulation
would otherwise be unfeasible and expensive to solve in terms of computational power
and required time. Therefore, once trained, the ML model can be queried to solve the
problem within a few microseconds;
(c) A general model is designed to estimate the number of required slots on physical
substrate subject to several VONs for a wide variety of physical topologies, in other
words, a single model can be used to the prediction using multiple variables as input;
(d) We showcase numerical results on the realistic 35 network topologies to demonstrate
the efficacy of the designed ML models in which service requests should be deployed
on the substrate EON.

The remainder of this work is organized as follows: Sect. 2 reviews how RSA can be mod-
eled as a supervised regression problem to be trained in an offline fashion, and further
shows its applicability in a scenario with virtualization and protection. Section 3 specifies
how the dataset was generated. Sections 4 and 5 provide a summary of the experiments
carried out and the results obtained, both from a ML perspective and networking view-
point. Finally, Sect. 6 concludes this paper.

2 Methodology

Much of the current literature on EON pays particular attention to the routing and spectrum
assignment (RSA) problem (Chatterjee and Oki 2020) and their constraints. In fact, these
constraints increase outstandingly the complexity of ILP formulations of RSA. This key
research issue adds more complexity if we have several VONs (Jinno 2009; Velasco 2012).

2.1 The RSA problem by a channel‑based plus virtualization approach

In some works in literature, the channel-based approach has been used by ILP (Velasco
2012; Hai 2020). Therefore, we can find a proper allocation of demands with different pro-
tection requirements, referred to as QoS-aware protection, with routes (working and protec-
tion) and slots for each demand over each VON using the channel-based approach. Note
that by precomputing the set of slots and predefined paths that can be assigned to each
demand, the complexity added by the contiguity and route constraints are removed. In such
strategy, demand lightpaths (either working or protection paths) chosen for a demand on
virtual topology VON can be embedded on the physical substrate by using a predefined
route r and a spectral channel that does not overlap (i.e. has no slot in common) with a
spectral channel assigned to any other working or protection path that shares any link with
r. Furthermore, the existence of source-destination node pairs with diverse slot demands
and paths for working and protection paths, the different protection requirements included
through virtualization introduce another dimension, since each virtual topology is proposed

13
Capacity, resilience and virtual embedding in elastic optical… Page 5 of 21 1075

to represent the assignment of different QoS aware protection services in the common
physical substrate (Assis et al. 2022). Please see our previous work (Assis et al. 2022) for
details about the channel-based plus virtualization approach.

2.2 The choice of the paths

The model in ILP node-link formulations encompasses all feasible regions and fields, pro-
viding a global optimal solution for the problem at the cost of possibly high processing
time, even for networks with few nodes, making its use not practical. For example, the
ILP simulation time for the 6-node network shown in Fig. 2 demands about 16 h (Assis
et al. 2022). Therefore, using pre-defined paths can decrease that complexity for the major-
ity of network planning problems. In our ML model, the form of representing a path for
demands associated to a pair of nodes is formed by both the working and protection paths
of the demand, as exemplified in Fig. 3, (r, b) is a pair of working and protection paths, and
it is assumed that there are three candidate p pairs {(r1 , b1 ), (r2 , b2 ), (r3 , b3 )}. Let r ∩ b be
the set of physical links simultaneously present in routes r and b. In this work, in order to
ensure protection, for any pair of routes (r, b) for an arbitrary demand ij, it is necessary that
r ∩ b = �.
Let us define F as the number of used spectrum slots to support all traffic
demands. For instance, Fig. 4 shows a 6-node network (Fig. 2) with i = 2 and j = 5.
(r1 , b1 ), (r2 , b2 ) and (r3 , b3 ) are three candidate p pairs of disjoint routing paths for demand
ij. Assuming that there are 3 VONS and that they all choose pair (r1 , b1 ), we then have dif-
ferent possibilities of squeezing survivability (QoS = 1, 2/3 or 1/3) for each VON. And
finally F is equal to 9.

2.3 The ML model

For this work, we decided to apply the supervised learning technique to train the Multilayer
Perceptron (MLP) neural network using a set of known inputs and corresponding outputs.
Figure 5 summarizes the adopted strategy, where i represents source, j represents desti-
nation, and t represents virtual topology. The dataset is split into train, validation, and test
sets, with no overlapping between sets. We tested different MLP architectures, changing
the number of layers and modifying the model’s hyperparameters. The final model repre-
sents the best architecture found. The dataset was built with a reduced and random number
of QoS combinations, this way, new combinations can be mapped creating new features,
which can be applied to the final model in order to generate new predicted values.
The MLP was trained with backpropagation training algorithm to predict the total num-
ber of spectrum slots F required to support all traffic demands depending on the topology

Fig. 2  6-node topology

13
1075 Page 6 of 21 T. A. C. de Melo et al.

Fig. 3  Example of three can-

didates path pairs (Assis et al.
2022)

characteristics, set of working and protection pair of paths (p) and the adopted QoS con-
figuration. For example, in Table 1, we can see in the hypothetical scenario of 6 instances,
as inputs, we have structural characteristics of topologies (for the example of the 6-node
topology illustrated in Fig. 2), the number of link-disjoint pairs of paths (p = 1, 2 or 3) in
each VON, different possibilities of survivability, QoS = 1 (3 slots in work path and 3 slots
in backup path), QoS = 2/3 (3 slots in work path and 2 slots in backup path) and QoS= 1/3
(3 slots in work path and 1 slot in backup path) for each VON and as output, the total num-
ber of spectrum slots F, a variable we will predict using a supervised approach. The value
of F was calculated based on an algorithm inspired by the original heuristic for RSA from
Wang et al. (2011), please see Sect. 3.1 for more details.

3 Dataset generation

The dataset were generated through of pre-computed sets of pair of link-disjoint paths p
and protection levels for the traffic demands of each VON, structural topology character-
istic data and the total number of spectrum slots F required to support all traffic demands

13
Capacity, resilience and virtual embedding in elastic optical… Page 7 of 21 1075

Fig. 4  Example of three VONs over a same physical substrate

Fig. 5  Summary of approach based on computational intelligence model for solving the RSA

depending on the topology characteristics, set of working and protection pair of paths (p)
and the adopted QoS configuration.
The number of candidate pairs of link-disjoint paths, p, for each demand is what deter-
mines the performance of routing network design methods (ILP with pre-defined paths or
heuristics) since, for example, the use of a single link-disjoint pair of paths p could lower

13
 1075

13
Page 8 of 21

Table 1  Example of some inputs and outputs from the dataset with 3 VONs in the 6-node topology
Instance N (number L (links) Nodal degree QoS (VON1) (p for VON1) QoS (VON2) (p for VON2) QoS (VON3) (p for VON3) Output (F)
of nodes) (mean)

1 6 2x7 2.34 1 3 1/3 2 1 3 138

2 6 2x7 2.34 1 3 1 2 2/3 2 57
3 6 2x7 2.34 1/3 1 1/3 1 2/3 2 82
4 6 2x7 2.34 1 3 1/3 3 2/3 1 104
5 6 2x7 2.34 2/3 1 2/3 2 2/3 3 71
6 6 2x7 2.34 1 2 1 2 1 2 96
T. A. C. de Melo et al.
Capacity, resilience and virtual embedding in elastic optical… Page 9 of 21 1075

the CPU time required, but it would also reduce the performance in the number of slots
required, while the probability of finding an optimal (or near-optimal) solution is increased
by three candidate link-disjoint pairs of paths, but at the expense of higher CPU time.
The Yen’s algorithm is capable of determining the k shortest link-disjoint candi-
date pairs of working and protection paths for each demand.The shortest path in unitary
weighted graphs is equivalent to computing the k routes with the lowest number of hops.
In order to get k link-disjoint paths for the dataset, this paper precomputes pair of link-
disjoint paths for each demand with k = 3 using Yen’s Algorithm. Therefore, p can be 1, 2,
or 3.
We apply an algorithm inspired by original heuristics for the RSA from Wang et al.
(2011) to define the total number of spectrum slots F. It is an algorithm based on routing
with shortest path and described below.

3.1 Shortest path with maximum spectrum reuse (SPSR)

Let us assume that a given set of requests on VON ∈ T is given with their amount of
requested traffic. The RSA problem is proposed to use the shortest path with maximum
spectrum reuse (SP-M) algorithm, which combines the shortest path routing with the maxi-
mum reuse spectrum allocation (MRSA) algorithm shown in Algorithm 1 from Wang et al.
(2011). In this approach, the spectrum disjoint pair p paths from every demand over all
VONs are first sorted according to the number of required slots, since the slot contigu-
ity constraint makes it harder to find available consecutive slots for demands for a larger
number of slots. After that, spectrum assignment is performed to the paths by following
the order in which they were sorted. Note that only fiber-disjoint spectrum paths may reuse
the same slots. We hence record the set of spectrum paths that are accommodated in the
current iteration and employ a First-Fit spectrum assignment strategy to find the first set of
available consecutive slots.
Besides, depending on the number of VONs, the number of possible QoS combinations
can increase significantly. In this research, we worked with three and ten VONs. For three
VONs, the amount of QoS combinations would be approximately twenty thousand, while
for ten VONs it would be 2 ⋅ 1014. We propose to work only with a random sample of this
data. We used 2322 and 52488 samples of different QoS for three and ten VONs, respec-
tively, for each topology.
The dataset was divided as follows: 70% of original data samples were used for training
and the remaining 30% for testing. The validation data set represents 30% of the training
data.

4 Setup and MLP performance assessment

We present two kinds of ML models:

1. The specialized model, which serves for a single topology;

2. The general model, which it serves for any from 34 topologies.

In the specialized model, we disregard structural topology characteristic data as input

since they wouldn’t change; thus, the inputs would consist only of the number of

13
1075 Page 10 of 21 T. A. C. de Melo et al.

Fig. 6  MLP architecture for specialized model

Fig. 7  MLP architecture for general model

link-disjoint pairs of paths (p = 1, 2 or 3) in each VON and different possibilities of

squeezing survivability (QoS = 1, 2/3 or 1/3) for each VON.
In the general model, the inputs would consist of structural topology characteristic
data, the number of link-disjoint pairs of paths (p = 1, 2 or 3) in each VON and different
possibilities of squeezing survivability (QoS = 1, 2/3 or 1/3) for each VON.
Empirical variations of ML architectures were examined regarding some hyperpa-
rameters such as the number of neurons in the inner hidden layers and the number of
hidden layers. After tests, the best ANN configuration found, considering the aforemen-
tioned hyperparameters, for the specialized model (6-node topology) is the following
(see Fig. 6): two hidden layers with 18 neurons in the first hidden layer and 12 neurons
in the second hidden layer.
In order to obtain satisfactory results in the construction of the general model, it was
necessary to increase the complexity of the model through the inclusion of a new hid-
den layer, in addition to increasing the number of neurons in the layers. The architecture
is represented in Fig. 7: three hidden layers with 64 neurons in the first hidden layer,
32 neurons in the second hidden layer and 18 neurons in the third hidden layer. These
architectural diagrams (Figs. 6 and 7) were generated using Netron (Roeder 2017).
For the models, we restricted the use of the Adam (Kingma and Ba 2014) optimizer
with 𝛽1 = 0.9, 𝛽2 = 0.999 and 𝜖 = 10−7 . The output and hidden layers used Rectified Lin-
ear Unit (ReLU) activation function. The learning rate was set to 0.001.
Early stopping is a method that interrupts training once the model performance stops
improving on a hold out validation dataset. Keras supports the early stopping of training
via a callback called EarlyStopping. The Early stopping method was applied in all simu-
lations with patience = 30. The number of maximum iterations (epochs) was 1000. In
all experiments, we normalised the features by scaling them ranging from 0 to 1 through
the minimum-maximum scaler procedure.
The statistical indicators employed to analyze the results are the Root Mean Squared
Error (RMSE), Mean Absolute Error (MAE), Pearson’s correlation coefficient ( 𝜌(r)),
fraction of data that lies within a factor of two (Fac2), Normalised Mean Square Error
(NMSE) and Coefficient of Determination ( R2),

13
Capacity, resilience and virtual embedding in elastic optical… Page 11 of 21 1075

√
√ n
√1 ∑
RMSE =√ (x − xi� )2 (1)
n i=1 i

n
1 ∑|
MAE = x − xi� || (2)
n i=1 | i

∑n
(xi − x) ⋅ (xi� − x� )
i=1
𝜌(r) = √ (3)
var(x) ⋅ var(x� )

∑n xi
i=1 [0.5 ≤ xi�
≤ 2.0]
Fac2 = (4)
N

1 ∑n
(xi − xi� )2
NMSE = n i=1
(5)
var(x)

∑n
(xi − xi� )2
(6)
2 i=1
R =1 −
(xi − x)2

where n is the number of samples, x and x′ are the observed and prediction sets, respec-
tively, xi and xi′ are the i-th observed sample and predicted value, respectively, and var() is
the variance of an input set of variables.
For these statistical metrics, values close to 0.0 are adequate for the RMSE, MAE,
NMSE. Values close to 1.0 are adequate for the Pearson, Factor of Two and coefficient of
determination. Machine learning algorithms were implemented in Python and performed
on an Intel(R) Core(TM) i7-1255U CPU 1.70GHz processor with 16GB RAM and we used
Intel(R) Core(TM) i7-3632QM CPU 2.20GHz processor with 8GB RAM for the applica-
tion of the SPSR-S algorithm.

5 Results and discussion

For our experiments, we used a 6-node and 7-link topology (as shown in Fig. 2) to vali-
date and quantify the performance of our approach based on computational intelligence
model (specialized model), and then we constructed a single database with information
from thirty-four topologies, see Table 2, to implement a general model. In this research, we
simulated scenarios with a fixed number of VONs.

5.1 Small network–specialized model

Firstly, we used a 6-node network with two scenarios, we assumed three ( VON1, VON2
and VON3) and ten ( VON1, VON2 , VON3,..., VON9 and VON10 ) virtual topologies with
75 Gbit/s traffic requests between all node-pair. The protection traffic requirement was

13
1075 Page 12 of 21 T. A. C. de Melo et al.

Table 2  The structural Network id Name N (number L (links) Nodal

characteristics of topologies of nodes) degree
(mean)

(a) Eight nodes Net 8 2 × 12 3.00

(b) Via network 9 2 × 12 2.67
(c) Bren 10 2 × 11 2.2
(d) Learn 10 2 × 12 2.4
(e) Metro optical 10 2 × 15 3.0
(f) RNP 10 2 × 12 2.4
(g) Abilene 11 2 × 14 2.55
(h) Compuserve 11 2 × 14 2.55
(i) Brasileira 12 2 × 19 3.17
(j) Cesnet 12 2 × 19 3.17
(k) VBNS 12 2 × 17 2.84
(l) Deutsche 14 2 × 23 3.29
(m) Italy 14 2 × 29 4.15
(n) Japan 14 2 × 22 3,15
(o) NSFNet 14 2 × 21 3.00
(p) Aconet 15 2 × 22 2.94
(q) Germany 17 2 × 26 3.06
(r) Pacific Bell 17 2 × 23 2.71
(s) Span 17 2 × 28 3.30
(t) Canarie 19 2 × 26 2.74
(u) EON 19 2 × 38 4.0
(v) Memorex 19 2 × 24 2.53
(w) Arpanet 20 2 × 32 3.2
(x) Sweden 20 2 × 24 2.4
(y) Pionier 21 2 × 25 2.39
(z) Cox 24 2 × 40 3.34
(aa) Usanet 24 2 × 43 3.59
(bb) Sanet 25 2 × 28 2.24
(cc) Newnet 26 2 × 31 2.39
(dd) Portugal 26 2 × 36 2.77
(ee) Renater 27 2 × 35 2.60
(ff) Omnicom 38 2 × 54 2.85
(gg) Internet 56 2 × 61 2.18
(hh) Coronet 75 2 × 99 2.64

determined by the protection level (QoS) adopted by each VON (Table 3 displays the
slot occupancy for each bit rate).
Table 4 presents the evaluation metrics of prediction of the number of required slots
for testing set with three and ten VONs. In the last column of Table 4, the range of the
number of required slots for a 6-node topology is specified. For specialized model, the
experiment with ten VONs presented the best results for all metrics, except for Fac2,
which has the same value.

13
Capacity, resilience and virtual embedding in elastic optical… Page 13 of 21 1075

Table 3  QoS-aware protection Bit rate Prot. Level Baud rate Slots occupancy
with PDM-QPSK transponder, (QoS) (12.5 GHz/slot)
Hai (2020)
75 Gbit/s 1 24 GBd 3
50 Gbit/s 2/3 16 GBd 2
25 Gbit/s 1/3 8 GBd 1

Fig. 8  Evolution of the loss of MLP model during training and validation in the 6-node topology with (left)
three and (right) ten VONs

Table 4  Results of statistical analyses of the 6-node topology for different QoS
VONs RMSE MAE 𝜌(r) Fac2 NMSE R2 Slot number range

3 5.25 4.09 0.97 1 0.05 0.93 49–151

10 4.06 2.91 0.99 1 0.01 0.97 144–451

Table 5  Comparison between Observed data Predicted data

the observed and predicted value
(the number of required slots)
138 139
for some QoS in the experiments
with 3 VONs of the 6-node 57 59
topology 82 78
104 106
71 77
96 87
51 50
65 57
78 76
126 127

Figure 8 shows the evolution of the loss during the training and validation processes of
the MLP model with (left) three and (right) ten VONs. When analysing the loss curves, in
both cases, it is clear that the models generalize well, presenting no overfitting or underfit-
ting, and the convergence behavior was similar for both training and validation loss.
Tables 5 and 6 present, respectively, comparison between the observed and predicted
value (the number of required slots) for some instances of the test set when 3 and 10 VONs

13
1075 Page 14 of 21 T. A. C. de Melo et al.

Table 6  Comparison between Observed data Predicted data

the observed and predicted value
(the number of required slots)
318 319
for some QoS in the experiments
with 10 VONs of the 6-node 304 301
topology 248 250
267 269
293 295
213 214
236 233
206 203
280 277
336 333

are assumed in the 6-node topology. Predicted data have been rounded to the nearest inte-
ger. In the examples presented in the Tables 5 and 6, the largest absolute errors are 9 for the
simulation with 3 VONs and 3 with 10 VONs.

5.2 Large networks–general model

It is important to know the behavior of the prediction target variable. We then propose an
exploratory analysis of the prediction variable (the number of required slots), for that we
use a data visualization through Box Plot. Figures 9 and 10 show the values of the first and
third quantiles, the median in the distribution, the width between the third and first quan-
tiles (interquartile range–IQR), spot outliers and the minimum and maximum values in the
dataset excluding outliers. More information about the range of the number of required
slots for each topology with 3 and 10 VONs can be viewed of the last column of Tables 4,

Fig. 9  A box plot visualization of the number of required slots for different QoS applied to six topologies
with 3 VONs

13
Capacity, resilience and virtual embedding in elastic optical… Page 15 of 21 1075

Fig. 10  A box plot visualization of the number of required slots for different QoS applied to six topologies
with 10 VONs

7 and 8. Due to the large number of topologies, Figs. 9 and 10 are restricted to a few topol-
ogies only.
In our experiments with large networks we considered the same methodology applied in
the 6-node network and same setup as Table 3. We assumed a high enough network capac-
ity to support all demands.
In order to measure the effectiveness of the proposed model, we conducted experiments
on various topologies. Figure 11 presents examples of topologies used in simulations and
Table 2 presents structural characteristics of all topologies used.
All results exposed so far are related to the first simulation approach for which a special-
ized model was built for the 6-node topology. For the second approach, we constructed a
single database with information from thirty-four topologies, see Table 2, to train a general
model.

Fig. 11  Some physical topologies used in the simulations to build the general model

13
1075 Page 16 of 21 T. A. C. de Melo et al.

Table 7  Results of statistical analyses–three VONs for different QoS (general model)
Network id Name RMSE MAE Pearson R Fac2 NMSE R2 Slot number
range

(a) Eight Nodes 5.03379 3.98372 0.986686 1.0 0.043785 0.9156 45–180
Net
(b) Via Network 7.2823 5.23531 0.993553 1.0 0.0266831 0.9575 68–297
(c) Bren 10.3304 8.2814 0.988814 1.0 0.0254679 0.9734 106–468
(d) Learn 8.97508 6.8599 0.990188 1.0 0.02248 0.9775 105–459
(e) Metro Optical 4.72199 3.75437 0.992991 1.0 0.0196232 0.9444 66–252
(f) RNP 7.17494 5.78994 0.993636 1.0 0.0155089 0.9741 93–414
(g) Abilene 8.70322 6.72231 0.992454 1.0 0.0189857 0.9567 108–522
(h) Compuserve 3.88754 2.96453 0.997917 1.0 0.00450439 0.9782 111–432
(i) Brasileira 4.30631 3.39491 0.997387 1.0 0.00876682 0.9741 90–324
(j) Cesnet 5.58634 4.40226 0.995463 1.0 0.0101839 0.9839 96–414
(k) VBNS 3.98352 3.07879 0.998518 1.0 0.00429597 0.9894 135–468
(l) Deutsche 8.00686 6.45735 0.993936 1.0 0.0148178 0.9834 116–477
(m) Italiana 6.98398 5.67136 0.991073 1.0 0.0250092 0.9750 88–324
(n) Japan 5.88626 4.72426 0.996869 1.0 0.00882808 0.9709 133–441
(o) NFSNet 5.04367 3.90303 0.996679 1.0 0.00736694 0.9616 107–441
(p) Aconet 6.01122 4.66141 0.997126 1.0 0.00749683 0.9690 134–540
(q) Germany 4.88907 3.99521 0.999213 1.0 0.00280099 0.9938 186–666
(r) Pacific Bell 8.13718 6.53111 0.997488 1.0 0.00516793 0.9888 213–882
(s) Span 12.4456 8.97418 0.993832 1.0 0.0179894 0.9823 194–702
(t) Canarie 9.27136 7.65162 0.998864 1.0 0.00357995 0.9956 306–1179
(u) EON 8.14688 6.64115 0.99576 1.0 0.0152401 0.9864 134–513
(v) Memorex 8.62595 6.53667 0.998609 1.0 0.00297412 0.9961 333–1134
(w) Arpanet 7.33577 5.80576 0.99867 1.0 0.00432849 0.9753 223–819
(x) Sweden 8.90064 6.83206 0.999533 1.0 0.00125233 0.9971 342–1863
(y) Pionier 4.90914 3.70953 0.999785 1.0 0.000453905 0.9972 411–1620
(z) Cox 7.8055 6.04381 0.999496 1.0 0.00128832 0.9968 420–1638
(aa) Usanet 6.54014 5.11239 0.998495 1.0 0.00319201 0.9954 249–792
(bb) Sanet 18.2084 14.3258 0.998955 1.0 0.00264342 0.9968 668–2916
(cc) Newnet 8.99101 7.23281 0.999798 1.0 0.000668293 0.9974 759–2457
(dd) Portugal 11.2774 8.48886 0.999302 1.0 0.00140179 0.9968 627–2205
(ee) Renater 12.5973 10.2386 0.999449 1.0 0.00143749 0.9945 702–2520
(ff) Omnicom 29.692 25.097 0.999355 1.0 0.00436117 0.9996 1015–3249
(gg) Internet 26.9952 21.0388 0.99985 1.0 0.000314972 0.9998 3387–10,719
(hh) Coronet 19.8852 15.6292 0.999969 1.0 0.000104082 1.0 4320–13,068
Mean – 9.3 7.4 0.99 1.0 0.0098 0.98 –

In the general model, there was a need to include a new resource (physical topology
name) to differentiate the topologies, in addition to the structural characteristic data. Due
to the categorical nature of this new feature and high cardinality, we use binary encoding
method to encode categorical features to integers and then convert them to binary code.
Tables 7 and 8 present the evaluation metrics of prediction of the number of required
slots for testing set of thirty four topologies with three and ten VONs for general model,

13
Capacity, resilience and virtual embedding in elastic optical… Page 17 of 21 1075

Table 8  Results of statistical analyses–Ten VONs for different QoS (general model)
Network id Name RMSE MAE Pearson R Fac2 NMSE R2 Slot number
range

(a) Eight nodes 20.903 18.1479 0.947667 1.0 0.289679 0.7626 178–516
net
(b) Via network 19.447 16.4236 0.982921 1.0 0.0956085 0.9444 264–844
(c) Bren 21.1543 15.3296 0.984174 1.0 0.0403394 0.9618 432–1338
(d) Learn 20.6941 15.5996 0.981041 1.0 0.040935 0.9484 426–1314
(e) Metro opti- 16.0562 12.9287 0.969804 1.0 0.0917297 0.8518 274–709
cal
(f) RNP 26.2797 21.1003 0.96282 1.0 0.0739401 0.9325
406–1196
(g) Abilene 20.032 14.958 0.982061 1.0 0.0362742 0.9356
420–1445
(h) Compuserve 11.8189 9.36014 0.992315 1.0 0.0158642 0.9575
428–1261
(i) Brasileira 10.0976 7.82707 0.995057 1.0 0.0199087 0.9633
366–969
(j) Cesnet 14.9172 11.5872 0.987335 1.0 0.0289189 0.9594
416–1201
(k) VBNS 12.2811 9.19824 0.994011 1.0 0.0153156 0.9707
504–1388
(l) Deutsche 10.6009 8.04217 0.995018 1.0 0.0114682 0.9733
496–1363
(m) Italiana 24.6395 18.2478 0.966392 1.0 0.0700558 0.8733
402–1220
(n) Japan 14.4806 11.3032 0.987781 1.0 0.0246133 0.9684
484–1258
(o) NFSNet 22.6372 18.7735 0.984011 1.0 0.0714141 0.9582
398–1132
(p) Aconet 21.274 16.4028 0.984987 1.0 0.031284 0.9726
546–1596
(q) Germany 14.3037 9.85323 0.996288 1.0 0.0096781 0.9815
736–2029
(r) Pacific bell 29.8613 24.8956 0.994234 1.0 0.0207539 0.9883
872–2639
(s) Span 21.1786 15.6386 0.986388 1.0 0.0295126 0.9346
666–1705
(t) Canarie 20.2264 15.6386 0.99672 1.0 0.008027 0.9849
1180–3114
(u) EON 18.7176 14.7288 0.991087 1.0 0.0317568 0.9236
550–1425
(v) Memorex 27.9518 22.7668 0.996892 1.0 0.0131445 0.9820
1304–3262
(w) Arpanet 26.0924 22.3304 0.995095 1.0 0.0227438 0.9849
860–2399
(x) Sweden 34.3735 24.2804 0.996473 1.0 0.007617 0.9950
1648–4991
(y) Pionier 29.3554 22.3457 0.997165 1.0 0.00662356 0.9958
1744–4766
(z) Cox 32.8297 26.5328 0.997414 1.0 0.00925181 0.9941
1644–4720
(aa) Usanet 23.4371 18.4102 0.99149 1.0 0.0173287 0.9603
866–2428
(bb) Sanet 52.8509 40.2887 0.996778 1.0 0.00767565 0.9798
2518–8194
(cc) Newnet 18.6676 14.2514 0.999568 1.0 0.00117279 0.9972
2848–7416
(dd) Portugal 32.9649 26.2141 0.998132 1.0 0.0050854 0.9919
2388–6310
(ee) Renater 167.941 43.2878 0.955411 0.999936 0.105545 0.9901
2592–7283
(ff) Omnicom 58.0888 50.8515 0.998954 1.0 0.00678853 0.9964
3714–9723
(gg) Internet 89.1948 66.9358 0.999295 1.0 0.00142861 0.9971
12,512–
32,288
(hh) Coronet 90.5329 71.7812 0.999607 1.0 0.00098569 0.9992 15,632–
38,903
Mean – 31.64 22.18 0.98 0.99 0.037 0.95 –

13
1075 Page 18 of 21 T. A. C. de Melo et al.

Fig. 12  Evolution of the loss of MLP model during training and validation in the general model with (left)
three and (right) ten VONs

respectively, and Fig. 12 shows the evolution of the loss during the training and validation
processes of the general models with (left) three and (right) ten VONs.
Considering the construction of general models, Fig. 13 shows the prediction of the
number of required slots for a sample of 200 instances of the test set in the Deutsche topol-
ogy with three VONs, while Fig. 14 illustrates the prediction of the number of required
slots for a sample of 200 instances of the test set in the Coronet topology with ten VONs.
Each point on the x-axis (different QoS) in Figs. 13 and 14 corresponds to an input instance
(feature vector) composed by: structural topology characteristic data, the number of link-
disjoint pairs of paths ( p = 1, 2 or 3) in each VON, different possibilities of squeezing

Fig. 13  Prediction of the number of required slots for different QoS in the Deutsche topology with three
VONs (general model)

Fig. 14  Prediction of the number of required slots for different QoS in the Coronet topology with ten VONs
(general model)

13
Capacity, resilience and virtual embedding in elastic optical… Page 19 of 21 1075

Table 9  Training and inference Time VONs General model

times of the experiments with
three and ten VONs for general
Training 3 446.3521s
model
Inference 3 0.00011345s
Training 10 5041.2033s
Inference 10 0.00021462s

Table 10  Results of statistical Model RMSE MAE Fac2 NMSE

𝜌(r) R2
analyses of the Deutsche
topology for different QoS with
Specialized 4.337 3.349 0.998 1.0 0.005 0.9965
three VONs
General 8.006 6.457 0.993 1.0 0.014 0.9834

survivability (QoS = 1, 2/3 or 1/3) for each VON, and physical topology name, which, due
to its categorical nature, we used binary encoding method to encode categorical features to
integers and then convert them to binary code.
When analysing the loss curves of general models (Fig. 12), it is clear that the models
trained with 3 VONs needed more epochs to reach the EarlyStopping and finish training,
while for the specialized model (Fig. 8) it was the opposite. There are no under/overfitting
issues in the models.
For general models, on average, the experiments with three VONs presented better
results in relation to error and correlation metrics compared to the experiments with ten
VONs, while for the specialized model it was the opposite.
Due to the amount of training data, we observed that the training time for the simula-
tion with 3 VONs was smaller than the training time for the simulation with 10 VONs, see
Table 9. None of the models reached the maximum number of epochs (1000) due to the
activation of EarlyStopping, however, in the case of the general model, the model with 10
VONs completes the training in a much smaller number of epochs. Regarding the time of
inference, the time for a single inference is a little smaller for model with 3 VONs com-
pared to model with 10 VONs, see Table 9.
Table 10 presents a comparison of the statistical metrics between the implementation
of a specialized model (based on the architecture shown in Fig. 6) for the Deutsche topol-
ogy and the result obtained with the general model for the same topology, both with three
VONs. We observed that there was a worsening in the evaluation metrics comparing the
results obtained by the general model in relation to the specialized model. This is due to
the fact that the general model had to generalize to characteristics of different topologies,
making it less accurate for each topology.
The general models were able to generalize to characteristics of thirty-four different
topologies and the results obtained demonstrated that the proposed method is effective for
predicting the number of required slots (bandwidth) on physical substrate subject to several
VONs.
As regards link capacity, the proposed model provides an estimate on the number of
slots in the network (i.e. spectrum resource) required to attend the provided number of
VONs under assumed traffic and resource usage conditions. The model may be used to
estimate the necessary amount of resources either for a network that must be installed/
upgraded within few months or a long term, given some future expected traffic and

13
1075 Page 20 of 21 T. A. C. de Melo et al.

resource usage conditions. In long term, it is expected a large amount of traffic in backbone
elastic optical networks, which is expected to be resolved by the use of space division mul-
tiplexing (SDM) under multi fiber/core, multi-band or both, in which case the total number
of expected slots is considerably large. That is the reason we have not limited in our analy-
sis the number of slots for the case of a single fibre/single band.

6 Conclusions

In this paper, we proposed a supervised machine learning (ML) technique using a selec-
tion of pre-computed sets of pair of link-disjoint paths p and protection levels for the traf-
fic demands of each VON, thus solving the problem of substrate capacity in EONs with
survivability. The performance of the models was assessed by the statistical metrics, along
with the evaluation of the time required for training and for making inferences.The pro-
posed general models are able to deal with multiple topologies and proved to generalize
well during the training and validation phases, presenting no under/overfitting. The MLP
architecture is effective for predicting the total number of spectrum slots to support all traf-
fic demands depending on the structural characteristics of the topologies, pair of link-dis-
joint paths p and the adopted QoS configuration. This study extends our knowledge of RSA
problem with virtualization and based on the proposed methodology, investigations and
on the findings, we believe that this research can provide contributions to the planning of
capacity of EONs considering survivability.
As a suggestion for future work, it proposes to improve the solution to work in scenarios
in which the number of VONs may change dynamically based on network conditions, traf-
fic load, or other parameters.
Acknowledgements This study was financed in part by the Coordenação de Aperfeiçoamento de Pessoal
de Nível Superior – Brasil (CAPES) – Finance Code 001. The authors thank to CNPq for scholarships and
grants, and to UFBA for its educational support.

Funding The authors have not disclosed any funding.

References
Assis, K., et al.: Network virtualization over elastic optical networks with different protection schemes.
IEEE/OSA J. Opt. Commun. Netw. 8(4), 272–281 (2016)
Assis, K.D.R., Almeida, R.C.D.A., Dos Santos, A.F., Dias, L.P., Giozza, W.F., De Sousa, R.T., Jaumard, B.,
Nejabati, R., Simeonidou, D.: Channel-based rsa approaches for qos protection of slices over elastic
optical networks. IEEE Access 10, 20714–20726 (2022)
Assis, K.D.R., Almeida, R.C., Dias, L.P., Waldman, H.: Squeezed protection in elastic optical networks sub-
ject to multiple link failures. IEEE Trans. Netw. Serv. Manage. 18(3), 2612–2626 (2021)
Assis, K., Almeida, R., Jr., Reed, M., Santos, A., Dinarte, H., Chaves, D., Li, H., Yan, S., Nejabati, R.,
Simeonidou, D.: Protection by diversity in elastic optical networks subject to single link failure. Opt.
Fiber Technol. 75, 103208 (2023)
Bui, M., Jaumard, B., Cavdar, C., Mukherjee, B.: Design of a survivable vpn topology over a service pro-
vider network. In: 2013 9th International Conference on the Design of Reliable Communication Net-
works (DRCN), pp. 71–78. IEEE (2013)
Chatterjee, B., Oki, E.: Elastic Optical Networks: Fundamentals. Control, and Management. CRC Press,
Design (2020)
Davalos, E.J., Baran, B.: A survey on algorithmic aspects of virtual optical network embedding for cloud
networks. IEEE 6, 20893–20906 (2018)

13
Capacity, resilience and virtual embedding in elastic optical… Page 21 of 21 1075

Dias, L., et al.: Channel-based rsa approach for virtualization and qos-aware protection in optical networks.
In: IEEE International Conference on Communications (ICC 2021), pp. 1–6. IEEE (2021)
Hai, D.T.: On the spectrum-efficiency of qos-aware protection in elastic optical networks. Optik 202,
163563 (2020)
Hammad, A., et al.: Novel approaches for composition of online virtual optical networks utilizing o-ofdm
technology. In: 39th European Conference and Exhibition on Optical Communication (ECOC 2013),
pp. 1–3. IET (2013)
Jinno, M., et al.: Spectrum-efficient and scalable elastic optical path network: Architecture, benefits, and
enabling technologies. IEEE Commun. Mag. 47(11), 66–73 (2009)
Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:​1412.​6980 (2014)
Klinkowski, M.: A genetic algorithm for solving rsa problem in elastic optical networks with dedicated path
protection. In: International joint conference CISIS’12-ICEUTE ‘12-SOCO’ 12 Special Sessions, pp.
167–176. Springer (2013)
Klinkowski, M., Walkowiak, K.: Routing and spectrum assignment in spectrum sliced elastic optical path
network. IEEE Commun. Lett. 15(8), 884–886 (2011)
Moura, T., Vieira, R., Almeida, R., Giozza, W., Rodrigues, J., Assis, K.D.R.: Planejamento de redes Ópticas
elásticas por meio da solução do max-rmsa. In: Anais do XXXVIII Simpósio Brasileiro de Redes de
Computadores e Sistemas Distribuídos, pp. 868–881. SBC (2020)
Pióro, M., Medhi, D.: Routing, flow, and capacity design in communication and computer networks. Else-
vier (2004)
Renjith, P., Sujatha, G., Vinoth, M., Vignesh, G., Prabhu, M.R., Mouleswararao, B.: Deep reinforcement
learning for comprehensive route optimization in elastic optical networks using generative strategies.
Opt. Quant. Electron. 55(13), 1197 (2023)
Roeder, L.: Netron - a viewer for neural network, deep learning and machine learning models (2017). https://​
github.​com/​lutzr​oeder/​netron
Shahriar, N.: Survivable virtual network embedding in transport networks (2020). http://​hdl.​handle.​net/​
10012/​16085
Sone, Y., et al.: Bandwidth squeezed restoration in spectrum-sliced elastic optical path networks (slice).
IEEE/OSA J. Opt. Commun. Netw. 3(3), 223–233 (2011)
Velasco, L., et al.: Modeling the routing and spectrum allocation problem for flexgrid optical networks. Pho-
ton Netw. Commun. 24(3), 177–186 (2012)
Wang, L.: Elastic optical network routing analysis using intelligent machine learning architecture with secu-
rity analysis. Opt. Quant. Electron. 55(14), 1238 (2023)
Wang, Y., Cao, X., Pan, Y.: A study of the routing and spectrum allocation in spectrum-sliced elastic optical
path networks. In: IEEE International Conference on Computer Communications (INFOCOM), pp.
1503–1511 (2011)
Zhang, Y., Xin, J., Li, X., Huang, S.: Overview on routing and resource allocation based machine learning
in optical networks. Opt. Fiber Technol. 60, 102355 (2020)
Zhang, L., Li, X., Tang, Y., Xin, J., Huang, S.: A survey on qot prediction using machine learning in optical
networks. Opt. Fiber Technol. 68, 102804 (2022)

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.

Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under
a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted
manuscript version of this article is solely governed by the terms of such publishing agreement and applicable
law.

13