accelerating the design of optical networks using surrogate models
TRANSCRIPT
Accelerating the Design of Optical Networks using Surrogate Models
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
Carmelo J. Bastos-Filho (Assoc. Prof. UPE) ,
Danilo R. B. Araújo (Ph.D. Student, UFPE)
Erick A. Barboza (Ph.D. Student, UFPE)
Joaquim F. Martins-fi lho (Assoc. Prof. UFPE)
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
Major question for this presentation
Is it possible to bring knowledge from other areas to improve the solutions for optical
networks?
In order to answer the general question, we will try to answer more specific questions!How to evaluate the current available methods to assess optical networks?
• An overview about the trade-off between accuracy and performance of the available tools
What is machine learning?• A brief overview on Artificial Neural Networks• What kind of applications we can develop?
What is Network Science?• A brief introduction to network metrics and generative models
How can we develop surrogate models to assess optical networks?• The major challenges related to the use of alternative procedures to assess optical networks• Network sciences + Artificial Neural Nets + Physical layer information Can we develop a suitable
surrogate model to assess optical networks???
What is the impact of using these surrogate models to design optical networks?
• A comparative study between “traditional” approaches and surrogate-based approaches
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
Designing Optical Networks In order to design optical networks, one must define:
1. The physical topology;
2. The equipments to be deployed (amplifiers, ROADMs, number of TX cards);
3. The deployed modulation format, grooming scheme, etc., ….
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
Designing Optical Networks In order to design optical networks, one must define:
1. The physical topology;
2. The equipments to be deployed (amplifiers, ROADMs, number of TX cards);
3. The deployed modulation format, grooming scheme, etc., ….
This means that you have a lot of variables!!!!
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
Designing Optical Networks (Let’s try to simplify)
One has to define:
• Which nodes should be connected [a1,2; a1,3; ... ; an-1,n]? • ai,j=1 if i and j are connected, and ai,j=0 otherwise;
• Which type of amplifier should be deployed in each link [amp1,2; amp1,3; ... ; ampn-1,n]? • ampi,j can assume different labels depending on the availability and suitability;
• How many wavelengths must be available in each link [w1,2; w1,3; ... ; wn-1,n]? • wi,j is the number of wavelengths between node i and j;
• Which equipments should be installed in each node [ROADM1; ROADM2; ... ; ROADMn]? Even if we try to simplify even more by using the same type of amplifier and ROADM in the entire network, and we deploy the same number of wavelengths for all links ◦ We still have (n2-n)/2 + 2 variables◦ [a1,2; a1,3; ... ; an-1,n; ROADM; w]
Examples using this description
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
For n = 14 |X| = 93
The number of variables grows quickly when larger networks are used
Examples using this description
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
For n = 14 |X| = 93For n = 34 |X| = 563
The number of variables grows quickly when larger networks are used
How to find the best “configuration”?
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
OPTIMIZATION METHODS!!!!
EXAMPLES:Integer Linear Programming;Evolutionary Algorithms;Swarm Intelligence;Multi-objective optimization.
How to find the best “configuration”?
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
The state-of-art optimization algorithms can not garantee the optimum solution for such a high dimensionality!!!!
OPTIMIZATION METHODS!!!!
EXAMPLES:Integer Linear Programming;Evolutionary Algorithms;Swarm Intelligence algorithms;Multi-objective optimization.
How to find the best “configuration”?
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
Besides, in all cases it is mandatory to have metrics to guide the optimization process! Objective functions:CAPEX;OPEX;Energy consumption;Network performance metrics.
How to find the best “configuration”?
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
Network performance metrics Examples: Blocking Probability (in dynamic traffic networks)Utilization rate (for static networks)Etc…
How to evaluate the “Objective function”?
Fide
lity
Resource efficiency
Experimental Measures
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
Fidelity of the analysis in comparison with real dataResource efficiency in terms of financial or computational costs
How to evaluate the “Objective function”?
Fide
lity
Resource efficiency
Experimental Measures
Simulations Based on Numerical Models
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
Fidelity of the analysis in comparison with real dataResource efficiency in terms of financial or computational costs
How to evaluate the “Objective function”?
Fide
lity
Resource efficiency
Experimental Measures
𝑦= 𝑓 (𝑥 )Simulations Based on Numerical Models
Simulations Based on Analytical Models
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
Fidelity of the analysis in comparison with real dataResource efficiency in terms of financial or computational costs
How to evaluate the “Objective function”?
Fide
lity
Resource efficiency
Experimental Measures
𝑦= 𝑓 (𝑥 )Simulations Based on Numerical Models
Simulations Based on Analytical Models
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
Fidelity of the analysis in comparison with real dataResource efficiency in terms of financial or computational costs
Is there any other possibility??
How to evaluate the “Objective function”?
Fide
lity
Resource efficiency
Experimental Measures
Surrogate Models Based on Machine Learning
𝑦= 𝑓 (𝑥 )Simulations Based on Numerical Models
Simulations Based on Analytical Models
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
Fidelity of the analysis in comparison with real dataResource efficiency in terms of financial or computational costs
How to evaluate the current available methods for optical networks analysis?
Fide
lity
Resource efficiency
Experimental Measures
Surrogate Models Based on Machine Learning
𝑦= 𝑓 (𝑥 )Simulations Based on Numerical Models
Simulations Based on Analytical Models
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
How to improve the fidelity of surrogate models based
on machine learning?
?
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
What should we know to develop good machine learning surrogates? Network Science
◦Metrics◦Generative models
Machine learning techniques◦Artificial Neural Networks
What is Network Science?
[1] T. G. LEWIS. NETWORK SCIENCE - THEORY AND APPLICATIONS. JOHN WILEY & SONS, 2009.
Network science is a multidisciplinary research field that can be applied to any problem modeled by graphs, in which the inputs or the topology of the graph can vary along the time
Recent developments in Network Science include:• Proposal of metrics that can explain the structure and the behaviour of
real world networks• Proposal of generative models that can represent the topology structures
of real world networks
What is Network Science?
[1] T. G. LEWIS. NETWORK SCIENCE - THEORY AND APPLICATIONS. JOHN WILEY & SONS, 2009.
Some well known metrics:• Average path length (APL): the average of the minimum path for all pairs of
nodes (source, destination) – mean value of the shortest path routes• Algebraic connectivity (AC): second smaller eigenvalue of the Laplacian
matrix – it is related to the robustness of the network• Density (d): ratio between the number of established links and the maximum
number of possible links• Diameter (D): the longest shortest path• Entropy (I): measures the uncertainty regarding the degree of a given node• Clustering coefficient (CC): it is calculated based on the number of
triangulations between nodes
What is Network Science?
[1] T. G. LEWIS. NETWORK SCIENCE - THEORY AND APPLICATIONS. JOHN WILEY & SONS, 2009.
Some well known generative models:• K-Regular: consists in linking each node i with the following k nodes
• entropy equal to zero and the diameter/APL depend on k• Ring networks are a special case
What is Network Science?
[1] T. G. LEWIS. NETWORK SCIENCE - THEORY AND APPLICATIONS. JOHN WILEY & SONS, 2009.
Some well known generative models:• K-Regular: consists in linking each node i with the following k nodes
• entropy equal to zero and the diameter/APL depend on k• Erdos-Renyi (ER): a link between i and j is randomly established, according to the
probability p• High entropy, lower APL/CC• Not applied to real world networks
What is Network Science?
[1] T. G. LEWIS. NETWORK SCIENCE - THEORY AND APPLICATIONS. JOHN WILEY & SONS, 2009.
Some well known generative models:• K-Regular: consists in linking each node i with the following k nodes
• entropy equal to zero and the diameter/APL depend on k• Erdos-Renyi (ER): a link between i and j is randomly established, according
to the probability p• High entropy, lower APL/CC
• Watts-Strogatz (WS): starts with k-regular networks and performs rewiring processes with the probability rp• Lower entropy, low APL and high CC• It can be suitable for transport networks!
What is Network Science?
[1] T. G. LEWIS. NETWORK SCIENCE - THEORY AND APPLICATIONS. JOHN WILEY & SONS, 2009.
Some well known generative models:
• K-Regular: consists in linking each node i with the following k nodes• entropy equal to zero and the diameter/APL depend on k
• Erdos-Renyi (ER): a link between i and j is randomly established, according to the probability p• High entropy, lower APL/CC
• Watts-Strogatz (WS): starts with k-regular networks and after that rewires new connections with the probability rp• Lower entropy, low APL and high CC
• Barabási-Albert (BA): starts with 3 nodes fully connected and each new nodes is added by using the preferential attachment concept (hubs attracts new connections)• High entropy and lower APL/diameter• presence of hubs – can be used for access networks
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
Artificial Neural Networks It is not a “magical” black box tool, instead it is a distributed tool for function approximation ◦ It was demonstrated some decades ago
Each neuron applies a non-linear function over the weighted sum of the inputs
If <the number of inputs forms a complete set regarding the required output> and <there are enough neurons in the hidden layer> and <the number of patterns presented to adjust the weights of the neurons is enough> then ◦ <an ANN can be used to approximate one desired measure, i.e. the output>
◦ *there are some well known algorithms to train the ANN. We used the backpropagation one (widely and most used)
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
ANN Applications for Optical Networks
•We can use Multi-Layer Perceptron to map the NF and GF as a function of the input and output powers applied to the amplifier. •MLPs may avoid the necessity of a small step to obtain a high resolution characterization.
• One can measure operation points with a gain interval of 3 dB, which results presenting errors as low as of 0.1 dB.
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
ANN Applications for Optical Cognitive Networks
We have developed an approach to adjust the operating point of a cascade of amplifiers (6 in a row for the results in the figure) based on the delta rule deployed to train the ANN
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
Let’s get back to Surrogates
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
How can we define a surrogate model to assess optical networks?
We handle the following problem:• Given: a RWA algorithm, the fiber topology and the
specification of the optical devices• Goal: To estimate the blocking probability (BP);• Subject to: the lack of an available wavelength or
unacceptable QoT due to physical impairments.
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
Some possible surrogates to estimate BP:• Simulations based on Monte Carlo experiments based on
discrete events• It can be precise, but it needs a lot of time!
How can we define a surrogate model to assess optical networks?
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
Some possible surrogates to estimate BP:• Simulations based on Monte Carlo experiments based on
discrete events• It can be precise, but it needs a lot of time!
• Closed analytical expressions to estimate BP• Fast, but can not represent all practical situations!
How can we define a surrogate model to assess optical networks?
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
Some possible surrogates to estimate BP:• Simulations based on Monte Carlo experiments that evaluate QoT
by using analytical expressions• It can be precise, but needs a lot of time!
• Closed analytical expressions to estimate BP• Quick, but not precise!
• Artificial Neural Networks (ANNs) obtained by using a database of previously evaluated optical networks
How can we define a surrogate model to assess optical networks?
How can we define a surrogate model to assess optical networks?
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
PROPOSAL:• Artificial Neural Networks as an approximation tool• Networks science metrics to “catch” the network behaviour• General physical layer information to include general information regarding QoT
Output LayerHidden LayerInput Layer
X1
X2
Xp
Z1
Z2
ZM
BP......
How can we define a surrogate model to assess optical networks?
Output LayerHidden LayerInput LayerX1
X2
Xp
Z1
Z2
ZM
BP......
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
First challenge: How to define the input set X?
How can we use surrogates to assess networks?Our proposed methodology is based on combining PCA and best selection
D. R. B. ARAUJO, C. J. A. BASTOS-FILHO, J. F. MARTINS-FILHO, METHODOLOGY TO OBTAIN A FAST AND ACCURATE ESTIMATOR FOR BLOCKING PROBABILITY OF OPTICAL NETWORKS, JOURNAL OF OPTICAL COMMUNICATIONS AND NETWORKING. 7 (5) (2015) 380-391.
Begin
End
1. Select a superset of input variables
2. Create a random dataset of WRONs
3. Evaluate the dataset of WRONs by simulations
5. Define p = 2
8. Use the p variables and the ANN to estimate BP
6. Test all sets of p variables as inputs of the ANN
ΔMSE > 0.05
7. Define p = p + 1
No
Yes
4. Use PCA to remove redundant variables
How can we use surrogates to assess networks?The role of each part of our complete solution [2]
[2] D. R. B. ARAUJO, C. J. A. BASTOS-FILHO, J. F. MARTINS-FILHO, METHODOLOGY TO OBTAIN A FAST AND ACCURATE ESTIMATOR FOR BLOCKING PROBABILITY OF OPTICAL NETWORKS, JOURNAL OF OPTICAL COMMUNICATIONS AND NETWORKING. 7 (5) (2015) 380{391.
Evaluator
Learning Engine
Complex Networks Engine
Start evaluation
Calculate the ANN outcome
Compute inputs
Start training
Discrete event simulator (network
simulator)
Calculate the weights of
ANN
Create a dataset of
WRONs
WRON BP
inputs
dataset
Trained ANN
WRON
WRON
Topological properties
BP
Topological properties
WRON
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
How can we use surrogates to assess networks?Superset of variables used to build an accurate estimator for BP of WRONs
Index Variable DefinitionX1 W Number of wavelengths
X2 δ OXC isolation factor
X3 CC Clustering coefficient
X4 d Density
X5 Entropy of the DFT of the Laplacian eigenvalues
X6 AC Algebraic connectivity
X7 NC Natural connectivity
X8 Average degree
X9 APL Average path length (hops)
X10 D Diameter (hops)
X11 I(G) Entropy
X12 Dkm Diameter (km)
X13 APLkm Average path length (km)
X14 ρ Spectral radius
X15 CR Concentration of routes
X16 L Traffic load
X17 σPL Standard deviation of the minimum path lengths
X18 d(km) Fiber link density
X19 ∆OSNR Average OSNR margin
X20 σ∆OSNR Standard deviation of the ∆OSNR
An illustrative case study on the performance of our proposal
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
13
12
14
10
6
4
1
2 9
11
8
7
5
3
26 k
m27
km
42 k
m
76 km
35 km
22 km
100 km
24 km
25 km72 km
55 km
25 km
21 km22 km
35 km
48 km
45 km
48 k
m
28 km
30 km
65 k
m
How can we use surrogates to assess networks?
[2] D. R. B. ARAUJO, C. J. A. BASTOS-FILHO, J. F. MARTINS-FILHO, METHODOLOGY TO OBTAIN A FAST AND ACCURATE ESTIMATOR FOR BLOCKING PROBABILITY OF OPTICAL NETWORKS, JOURNAL OF OPTICAL COMMUNICATIONS AND NETWORKING. 7 (5) (2015) 380{391.
An illustrative case study about the performance of our proposal [2]
p Camada de entrada
2 W, δ 4.02E-3 4.6E-5 -
3 W, δ, d 5.62E-4 2.4E-4 0.86
4 W, δ, d, CR 3.23E-4 1.6E-5 0.43
5 W, δ, d, CR, CC 2.83E-4 1.5E-5 0.12
6 W, δ, d, CR, CC, 2.66E-4 6.8E-6 0.06
7 W, δ, d, CR, CC, , APL (km)
2.57E-4 9.7E-6 0.03
p = 3 p = 4 p = 5 p = 6 p = 7 Results of [8]
SIMTON A
SIMTON B
1.0E-04
2.0E-04
3.0E-04
4.0E-04
5.0E-04
6.0E-04
Method to estimation of BP
MS
E
What is the impact of surrogates and network science to design optical networks?
[3] D. R. B. ARA´UJO, C. J. A. BASTOS-FILHO, AND J. F. MARTINS-FILHO. NA EVOLUTIONARY APPROACH WITH SURROGATE MODELS AND NETWORK SCIENCE CONCEPTS TO DESIGN OPTICAL NETWORKS. ENGINEERING APPLICATIONS OF ARTIFICIAL INTELLIGENCE, 43(08):67–80, 2015.
We have successfully used surrogates and network science concepts to design optical networks
The two main fields under investigation are:• Proposal of generative procedures to create good fiber topologies• Proposal of schemes to combine surrogates and discrete event simulatior to accelerate the
convergence of EA-based approaches
We studied the impact of our proposal to design the 14-node network [3]• Our goal is to find network configurations that presents good trade-off in terms of CAPEX
and blocking probability• We compared our proposal with traditional EA-based approaches
• Previous approaches used random generators and used only network simulations to assess the quality of network configurations
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
Watt-Strogatz model driven by traffic for seed generation
What is the impact of surrogates and network science to design optical networks?
0.0001 0.001 0.01 0.1 10
2000
4000
6000
8000
10000
12000
CHAVES [7]ARAUJO [8]WS-T
Blocking Probability
Cost
[7] D. A. R. CHAVES, C. J. A. BASTOS-FILHO, J. F. MARTINS-FILHO, MULTIOBJECTIVE PHYSICAL TOPOLOGY DESIGN OF ALL-OPTICAL NETWORKS CONSIDERING QOS AND CAPEX, OPTICAL FIBER COMMUNICATION. OFC 2010, 1{3.[8] D. R. B. ARAUJO, C. J. A. BASTOS-FILHO, E. A. BARBOZA, D. A. R. CHAVES, J. F. MARTINS-FILHO, AN ECIENT MULTI-OBJECTIVE EVOLUTIONARY OPTIMIZER TO DESIGN ALL-OPTICAL NETWORKS CONSIDERING PHYSICAL IMPAIRMENTS AND CAPEX, IN:
INTELLIGENT SYSTEMS DESIGN AND APPLICATIONS (ISDA), 2011 11TH INTERNATIONAL CONFERENCE ON, 2011, PP. 76{81.
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
Cascade of Surrogate Models (CSM)
0.00001 0.0001 0.001 0.01 0.1 10
2000
4000
6000
8000
10000
12000
PrefEA-NSEA-CSM
Blocking Probability
Cost
(m.u
.)
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
Not using surrogates
A reference Pareto front and two fronts obtained after 15 minutes of execution for EA-NS and for our
new proposal. Pareto fronts are for the non-uniform traffic scenario. The cost is given in generic
monetary units (m.u.).
0.00001 0.0001 0.001 0.01 0.1 10
2000
4000
6000
8000
10000
12000
PrefEA-NSEA-CSM
Blocking Probability
Cost
(m.u
.)
C
B
D
A
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
A reference Pareto front and two fronts obtained after 15 minutes of execution for EA-NS and for our
new proposal. Pareto fronts are for the non-uniform traffic scenario. The cost is given in generic
monetary units (m.u.).
0.00001 0.0001 0.001 0.01 0.1 10
2000
4000
6000
8000
10000
12000
PrefEA-NSEA-CSM
Blocking Probability
Cost
(m.u
.)
C
B
D
A
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
Not using surrogates
A reference Pareto front and two fronts obtained after 15 minutes of execution for EA-NS and for our
new proposal. Pareto fronts are for the non-uniform traffic scenario. The cost is given in generic
monetary units (m.u.).
EXECUTION TIME
Conclusions
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
Current solutions for analysis and design of optical networks present a trade-off in terms of fidelity and resource efficiency
Network science is a promising research field that can contribute to the development of new network analysis tools
Global performance metrics for optical networks such as blocking probability can be assessed by surrogate models based on machine learning techniques
Topological metrics from network science that summarize the fiber topology are natural candidates to offer reduction of dimensionality for networks assessment
ANNs can be used to forecast the blocking probability of optical networks when the right set of inputs is chosen
Surrogates and generative models can be used together to assist the design of optical networks
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
Thanks for your attention!
Accelerating the Design of Optical Networks using Surrogate Models
IV INTERNATIONAL WORKSHOP ON TRENDS IN OPTICAL TECHNOLOGIES – PROF. CARMELO J. A. BASTOS-FILHO, PHD
Carmelo J. Bastos-FILHO (Assoc. Prof. UPE) ,
Danilo R. B. Araújo (Ph.D. Student, UFPE)
Erick A. Barboza (Ph.D. Student, UFPE)
Joaquim F. Martins-fi lho (Assoc. Prof. UFPE)