Content uploaded by Stanislaw Kozdrowski
Author content
All content in this area was uploaded by Stanislaw Kozdrowski on Nov 18, 2020
Content may be subject to copyright.
PROCEEDINGS OF SPIE
SPIEDigitalLibrary.org/conference-proceedings-of-spie
Application of machine learning
methods in provisioning of DWDM
channels
Piotr Paziewski, Sławomir Sujecki, Stanisław Kozdrowski
Piotr Paziewski, Sławomir Sujecki, Stanisław Kozdrowski, "Application of
machine learning methods in provisioning of DWDM channels," Proc. SPIE
11204, 14th Conference on Integrated Optics: Sensors, Sensing Structures,
and Methods, 1120407 (13 September 2019); doi: 10.1117/12.2536656
Event: Fourteenth Integrated Optics-Sensors, Sensing Structures and
Methods Conference, 2019, Szcyrk-Gliwice, Poland
Downloaded From: https://www.spiedigitallibrary.org/conference-proceedings-of-spie on 16 Sep 2019 Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
Application of Machine Learning Methods in provisioning of
DWDM channels
Piotr Paziewskia, S lawomir Sujeckib, and Stanis law Kozdrowskia
aInstitute of Computer Science, Faculty of Electronics and Information Technology,
Warsaw University of Technology, Nowowiejska 15/19, 00-665 Warsaw, Poland
bDepartment of Telecommunications and Teleinformatics, Faculty of Electronics, Wroclaw
University of Science and Technology, Wyb. Wyspianskiego 27, 50-370 Wroclaw, Poland
ABSTRACT
Complexity and size of modern optic-fiber networks start to challenge the traditional methods of managing them
and yet majority of telecommunication companies still report rapid growth of their optical networks. One of
essential problems in managing optic-fiber networks is calculating the Quality of Transmission (QoT) of given
path in network. The unit responsible for this task is Optical Performance Unit (OPU) which communicates
with Network Management System (NMS). OPU’s task is to determine whether it is possible to transmit signal
through a given path. Modern OPUs are still operating based on traditional algorithms e.g. these systems take
into consideration known physics rules and information about the network parameters, calculating transmission
losses for each path. Main parameter that determines the OPUs result is Optical Signal to Noise Ratio (OSNR).
However, measuring its value from NMS level is often not practical. An alternative solution to this problem might
prove the application of Machine Learning (ML) algorithms for the estimation of OSNR. In this contribution
an application of Artificial Neural Network (ANN) to an evaluation of OSNR in an optical Dense Wavelength
Division Multiplexing (DWDM) network is investigated.
Keywords: WDM optical networks, Quality of transmission, Machine learning, Artificial neural networks.
1. INTRODUCTION
In the last decade, due to the demand for high network throughput, the number of optical channels has increased
rapidly. Therefore, it is important to increase the quality of transmission1along with the increase of the bit
rate in the transmission networks.2,3Optical Signal-to-Noise Ratio (OSNR) is used as one of the main metrics
to evaluate the quality of transmission for DWDM (Dense Wavelength Division Multiplexing) channels.4The
knowledge of OSNR is very valuable since it helps estimating the BER (Bit Error Ratio) level.
DWDM channels information capacity is strongly limited by ASE (Amplified Spontaneous Emission). ASE
is generated by optical amplifiers. When it comes to DWDM channels the main source of ASE are signal
boosters, in-line and pre- amplifiers, which are placed in various places in optical network. ASE from all optical
amplifiers accumulates along an optical path and may make the OSNR value too low, for reliable transmission
of information.
Experimental measurement of OSNR is usually performed using an optical spectrum analyzer. However,
such approach is often not practical in an operating DWDM network. Instead one can try to apply Machine
Learning Algorithms (MLA) e.g. Artificial Neural Networks (ANN), Support Vector Machines (SVM) or Linear
Regression (LR)56 7 to estimate OSNR.
In this contribution a specific focus is given to ANN. MLAs based on ANNs proved effective in recognizing
patterns and data classification. An ANN is build from neurons, which are connected to each other. Usually,
neurons are organized in layers while a particular neuron is connected with all neurons in previous layer. In an
ANN one can distinguish input layer (which holds the input data), the output layer (which holds the result of
Further author information: (Send correspondence to Stanis law Kozdrowski)
E-mail: s.kozdrowski@elka.pw.edu.pl, Telephone: +48 22 234 5048
14th Conference on Integrated Optics: Sensors, Sensing Structures, and Methods,
edited by Przemyslaw Struk, Tadeusz Pustelny, Proc. of SPIE Vol. 11204, 1120407
© 2019 SPIE · CCC code: 0277-786X/19/$21 · doi: 10.1117/12.2536656
Proc. of SPIE Vol. 11204 1120407-1
Downloaded From: https://www.spiedigitallibrary.org/conference-proceedings-of-spie on 16 Sep 2019
Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
the calculation) and a number of hidden layers. There are many types of neural networks for example: feed-
forward neural networks, in which information travels in only one direction, from input to output, recurrent
neural networks, in which neuron connections may form loops and cycles, convolutional neural networks, which
are a class of deep neural networks which have similar structure to feed-forward networks but are designed for
image analysis. In a neural network single neuron holds a float value between 0 and 1. The learning process of
the neural network consists in adjusting the weights of each neuron, so that the last layer of the neural network
returns the results.8ML, in particular ANN, is being proposed for lightpath quality of transmission (QoS)
estimation in the literature.910
This work describes and studies the application of ANN to the estimation of OSNR in DWDM networks.
2. PROBLEM FORMULATION
The simulations are performed for two optical networks, Polish network and German one. In order to evaluate
the robustness of the ML method the data base was generated using realistic DWDM networks. Figure 1shows
the network topology for networks studied, which include cities and optic connections between cities.
(a)(b)
Figure 1: Analyzed network topologies: (a) Polish, (b) German national transmission optical backbone network.
Polish network has 12 cities, 18 connections which gives 2457 paths. German one has 17 cities, 23 connections
and 4527 paths. Every connection has defined length which equals to real distance between given cities. In our
model, every optic connection has signal boosters placed approximately every 70km (exact value depends on
connection length). Boosters compensate for signal loss between the network but also act as source of noise in
the transmission. Loss between boosters depends on the distance between them. OSNR value of a particular
path is calculated based on total noise from all boosters in the path.
The developed software used python library called mxnet to implement feed-forward neural network and
learning process. Input data for ANN is an array with integer values that represent the path in the optical
DWDM network. Each integer represents the city (its value equals the city identifier). The order of identifiers
corresponds to the occurrences of cities on the path. The length of every input array was fixed and equal to
the longest path in a network. The input array was filled with path nodes identifiers and if particular path was
shorter than the longest path, the end of array was filed with value -1.The output of ANN was an integer value
in the range 0 - 3. Every output value corresponds to the specific range of OSNR values. The boundary values of
OSNR ranges, which divide paths into 4 categories, were designated in such a way, that number of paths in each
category was similar. For German network the ranges were defined as: less then 20.78, between 20.78 and 25.09,
between 25.09 and 31.81 and greater then 31.81 while for Polish network: less then 19.01, between 19.01 and
22.97, between 22.97 and 29.73 and greater then 29.73. The exemplary sets of OSNR values and corresponding
to them estimated categories are shown in 2. At the beginning a list of input arrays and list of corresponding
Proc. of SPIE Vol. 11204 1120407-2
Downloaded From: https://www.spiedigitallibrary.org/conference-proceedings-of-spie on 16 Sep 2019
Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
correct outputs was prepared. Obviously, the order of paths in the list was randomized. Then this data was
divided into training and test data.
Figure 2: OSNR values for example paths in the both, Polish and German, networks and estimated by ANN
category for each of them. Red color indicates a wrong estimation.
The considered ANN consists of two hidden layers (128 and 64 neurons), input layer and output layer. All
layers are fully connected with layers before them. Figure 3presents a diagram of an artificial neural network
used in the paper.
Figure 3: Diagram of the artificial neural network ANN used in the article; n (in the input layer) corresponds to
the longest path in the optical network.
In the learning process training data was divided into batches of 15 input paths. Training process is divided
into iterations, in the performed simulations the number of iterations equaled 300. In every iteration the whole
training data was processed, one batch at the time. Network parameters were updated after processing single
batch. After a defined number of iterations is completed one can start testing the network on data it has not
seen before.
Proc. of SPIE Vol. 11204 1120407-3
Downloaded From: https://www.spiedigitallibrary.org/conference-proceedings-of-spie on 16 Sep 2019
Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
3. RESULTS AND DISCUSSION
There are two parameters that are used to estimate quality of the ANN results. First one is the train accuracy
which shows how good the network is at training data classification. This parameter only shows the network
performance in classifying data that it is training on. However, its high value shows that neural network is
complex enough to keep information about the optical network. The other parameter is validation accuracy.
This parameter estimates how the network works with data that it has not seen before, and thus with a real data
network. Its value equals to ratio between correctly estimated validation examples (examples that were not used
while training) and all validation examples. Validation accuracy is usually calculated after the learning process.
Because of that, it is hard to tell how good the machine learning model actually is until its validation accuracy
is calculated on new data.11
Figure 4: Dependence of validation accuracy on number of example paths in data base of both, Polish and
German, networks.
In order to model training in real time we trained the neural network with increasing amounts of training
data, starting from very low number of examples. Figure 4shows relationship between validation accuracy and
number of example paths in data base for both networks. These results show that neural network achieves high
accuracy in classifying new input data when given enough training examples. In order to properly estimate
OSNR of test paths, network has to accumulate information about distances between cities. Figure 4suggest
that even low number of examples, approximately 25% of all possible paths, enables gathering such information.
The factor that might prevent network from achieving validation accuracy close to 1.0 is a random additional
part in signal loss of an optical fiber.
Proc. of SPIE Vol. 11204 1120407-4
Downloaded From: https://www.spiedigitallibrary.org/conference-proceedings-of-spie on 16 Sep 2019
Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
4. CONCLUSION
One of the key problems for network operators of a self-driving optical network is the automatic provisioning of
lightpath in DWDM networks. In the paper, ANN was applied for estimating OSNR values in a simulation of
DWDM network. The numerical results demonstrate the potential of artificial neural networks in estimating the
quality of transmission.
REFERENCES
[1] Behnam Shariati, Antonia Mastropaolo, N.-P. D. J. M. R.-M. D. K. I. T., “Physical-layer aware perfor-
mance evaluation of sdm networks based on smf bundles, mcfs, and fmfs,” International Journal of Optical
Communications and Networking (August 2018).
[2] Kozdrowski, S., ˙
Zotkiewicz, M., and Sujecki, S., “Resource optimization in fully flexible optical node archi-
tectures,” in [20th International Conference on Transparent Optical Networks (ICTON) ], (July 2018).
[3] Kozdrowski, S., ˙
Zotkiewicz, M., and Sujecki, S., “Optimization of optical networks based on cdc-roadm
technology,” Applied Sciences 9, 4648–4658 (Jan 2019).
[4] Metropolis, N., Rosenbluth, A. W., Rosenbluth, M. N., Teller, A. H., and Teller, E., “Evaluating machine
learning models for qot estimation,” In Proceedings of the 20th International Conference on Transparent
Optical Networks (ICTON). (15 July 2018.).
[5] F. Musumeci, C. Rottondi, A. N. I. M. D. Z.-M. R. M. T., “An overview on application of machine learning
techniques in optical networks,” IEEE Communications Surveys Tutorials (08 November 2018).
[6] Faisal Nadeem Khan, C. L. and Lau, A. P. T., “Machine learning methods for optical communication
systems,” (2427 July 2017).
[7] L. Peshkin, V. S., “Reinforcement learning for adaptive routing,” (12-17 May 2002).
[8] Wawrzy´nski, P., [Podstawy Sztucznej inteligencji], Oficyna Wydawnicza Politechniki Warszawskiej (2015.).
[9] Jimnez, T., Aguado, J. C., de Miguel, I., Durn, R. J., Angelou, M., Merayo, N., Fernndez, P., Lorenzo,
R. M., Tomkos, I., and Abril, E. J., “A cognitive quality of transmission estimator for core optical networks,”
Journal of Lightwave Technology 31(6), 942–951 (2013).
[10] Barletta, L., Giusti, A., Rottondi, C., and Tornatore, M., “Qot estimation for unestablished lighpaths using
machine learning,” Optical Fiber Communication Conference (Los Angeles, California United States, 1923
March, 2017).
[11] Cichosz, P., [Data Mining Algorithms: Explained Using R], Wiley (2014).
Proc. of SPIE Vol. 11204 1120407-5
Downloaded From: https://www.spiedigitallibrary.org/conference-proceedings-of-spie on 16 Sep 2019
Terms of Use: https://www.spiedigitallibrary.org/terms-of-use