PaperlayerView exact revision on arXiv

Source-linked AI summary

Machine Learning for Intelligent Optical Networks: A Comprehensive Survey

Rentao Gu, Zeyuan Yang, Yuefeng Ji

arXiv:2003.05290v1cs.NIcs.LG

TL;DR

Growing traffic and operational complexity motivate greater intelligence and automation in optical networks. This paper surveys ML applications and paradigms across optical-network control, resource management, monitoring, and survivability. It identifies the field as still immature and discusses challenges and possible solutions.

  • Problem

    Growing traffic and operational complexity motivate efforts to improve intelligence and automation in optical networks.

    Source: §1 · p. 1

  • Method

    The paper surveys ML use cases in optical networks and discusses paradigms and motivations for applying ML.

    Source: §7 · p. 29

  • Results

    The survey reviews use cases in optical-network control and resource management, and in network monitoring and survivability.

    Source: §7 · p. 29

  • Takeaways & Limitations

    The authors present possible solutions to challenges in applying ML to optical networks to inspire future research directions.

    Source: §1 · p. 4

    The field of applying ML in optical networks remains far from maturity and under investigation.

    Source: §1 · p. 4

Abstract

from arXiv · show

With the rapid development of Internet and communication systems, both in services and technologies, communication networks have been suffering increasing complexity. It is imperative to improve intelligence in communication network, and several aspects have been incorporating with Artificial Intelligence (AI) and Machine Learning (ML). Optical network, which plays an important role both in core and access network in communication networks, also faces great challenges of system complexity and the requirement of manual operations. To overcome the current limitations and address the issues of future optical networks, it is essential to deploy more intelligence capability to enable autonomous and exible network operations. ML techniques are proved to have superiority on solving complex problems; and thus recently, ML techniques have been used for many optical network applications. In this paper, a detailed survey of existing applications of ML for intelligent optical networks is presented. The applications of ML are classified in terms of their use cases, which are categorized into optical network control and resource management, and optical networks monitoring and survivability. The use cases are analyzed and compared according to the used ML techniques. Besides, a tutorial for ML applications is provided from the aspects of the introduction of common ML algorithms, paradigms of ML, and motivations of applying ML. Lastly, challenges and possible solutions of ML application in optical networks are also discussed, which intends to inspire future innovations in leveraging ML to build intelligent optical networks.

1. Introduction

Optical networks face growing operational complexity, while manual control and management struggle to meet future demands. This survey reviews ML applications across control, resource management, monitoring, and survivability, alongside ML paradigms and open challenges.

1. Introduction

  • Optical-network operation combines multidimensional resource allocation with service and network complexity, while manual approaches can be slow and locally rather than globally optimizing.Resources include wavelength, spectrum, and time slots; the introduction also identifies differentiated QoS and real-time service provision as challenges.Source: §1 · p. 2
  • The survey distinguishes its ML focus from prior reviews that covered optical communications broadly or emphasized algorithms, workflows, evaluation, and network-management support.It notes that one prior review only briefly introduced ML applications, which this paper focuses on.Source: §1 · p. 3
  • The survey reviews ML applications in optical-network control and resource management, and in network monitoring and survivability.Its scope includes traffic prediction, routing and wavelength or spectrum assignment, and performance monitoring, QoT estimation, and failure management.Source: §1 · p. 3
  • The paper summarizes regression, classification, and decision-making paradigms, discusses an ML-supporting system architecture, and analyzes motivations for applying ML in optical networks.It considers motivations in terms of ML characteristics and enabling techniques in the optical-network domain.Source: §1 · p. 3
  • The paper identifies data-access issues, ML-model drawbacks, security, and the simulation-to-real-network gap as challenges in a field it describes as immature.It also presents possible solutions intended to inspire further research.Source: §1 · p. 4

2. Paradigms and Motivations of Applying Machine Learning in Optical Networks

The section outlines how ML agents interact with optical networks through data collection and learning paradigms, then motivates their use through historical-data utilization, lower online computation, reduced feature engineering, and enabling technologies.

2.1. System Architecture and Application Paradigms

  • Functional elements collect and preprocess network data for ML agents, which use regression, classification, or environment-interacting decision-making workflows.In decision-making, agents update strategies according to rewards returned after network actions.Source: §2.1 · pp. 5–6

2.2.1. Historical Data Utilization

  • ML can learn dependencies from historical datasets, improving robustness and accuracy under noise compared with traditional methods that use only current network state.The section also notes that trained models need not rerun the learning algorithm when network state changes only slightly.Source: §2.2.1 · p. 6

2.2.2. Reduce the Online Computation Requirements

  • Offline training can use datacenter resources while online inference runs in the network, and trained neural networks are faster than Split Step Fourier computation for OSNR monitoring.This division addresses dynamic traffic and low-latency services, although model training may be computationally expensive.Source: §2.2.2 · pp. 6–7

2.2.3. Reduce Feature Engineering and Expert Knowledge Requirement

  • Deep learning can extract features directly from raw data, simplifying feature engineering and reducing requirements for expert knowledge and manual effort.For eye-diagram OSNR monitoring, the section says analytic methods are limited by the lack of an explicit pixel-to-OSNR relationship, while DL accuracy can meet requirements.Source: §2.2.3 · p. 7

2.2.4. Related Enabling Techniques

  • Programmable SDON architectures, cognitive-plane designs, and streaming telemetry support deploying ML functions and gathering network data.The section also describes increasingly mature optical-device processing and embedded OTDR as improving network awareness and access to state information.Source: §2.2.4 · p. 7

3. An Overview of Machine Learning in Optical Networks

The section places ML among techniques already used across several computational and communication domains. It introduces three learning categories used in optical networks: supervised, unsupervised, and reinforcement learning.

3. An Overview of Machine Learning in Optical Networks

  • ML is already used in image and speech recognition, natural language processing, game control, and recommendation systems.The section also notes its use in several communication systems.Source: §3 · pp. 7–8
  • The overview frames optical-network ML within a broader pattern of adoption across communication systems and other application areas.Source: §3 · pp. 7–8
  • The section introduces supervised learning, unsupervised learning, and reinforcement learning as the three categories of ML algorithms used in optical networks.Source: §3 · p. 8

3.1. Supervised Learning

Supervised learning trains models on input vectors paired with target outputs to infer a mapping. The section introduces SVMs and neural-network variants, describing their core mechanisms and uses.

3.1. Supervised Learning

  • Supervised learning uses input vectors paired with target outputs to infer a function mapping inputs to outputs.The section introduces several supervised algorithms used in optical networks.Source: §3.1 · p. 8

3.1.1. Support Vector Machine (SVM)

  • SVMs seek a separating hyperplane with maximal margin, maximizing the distance to the closest data vectors while separating two classes.The optimization is equivalent to minimizing the norm of the weight vector.Source: §3.1.1 · p. 8
  • When data are not linearly separable, SVM kernels map features to a higher-dimensional space; δ-margin hyperplanes can allow some misclassifications and guarantee better generalization.The resulting hyperplane may be nonlinear in the original space, and the Gaussian kernel is widely used.Source: §3.1.1 · p. 9

3.1.2. Neural Networks (NNs)

  • Neural-network use also requires attention to sensitivity to training hyperparameters, overfitting, and large computing-resource requirements.The section notes these disadvantages alongside the benefits of neural networks.Source: §3.1.2 · p. 9
  • An ANN passes inputs through neurons and nonlinear activation functions; training evaluates a loss such as MSE and uses backpropagation with gradient descent to update weights and biases.Iterations stop when MSE falls below a threshold.Source: §3.1.2 · pp. 9–10
  • CNNs use convolution, pooling, and fully connected layers: kernels detect local features, max pooling summarizes each rectangular area, and final layers produce classification or regression results.CNNs leverage sparse interactions, parameter sharing, and equivariant representations to reduce complexity and improve performance [35].Source: §3.1.2 · p. 10
  • RNNs process sequence data whose outputs can depend on previous inputs; LSTMs add gated memory and internal self-loops to store and update contextual information.LSTMs can overcome conventional RNN gradient vanishing and blow-up issues.Source: §3.1.2 · p. 11

3.2. Unsupervised Learning

Unsupervised learning trains on unlabeled data and includes clustering, dimensionality reduction, and density or parameter estimation. The section focuses on K-means and PCA, both used in optical networks.

3.2. Unsupervised Learning

  • Unsupervised learning trains on unlabeled data and includes clustering, dimensionality reduction, and density or parameter estimation.The section discusses K-means and PCA as methods used in optical networks; EM has been used in optical communications but, to the authors’ knowledge, not yet in optical networks.Source: §3.2 · pp. 11–12

3.2.1. K-means Clustering

  • K-means minimizes within-class distance by assigning samples to the nearest of K prototypes and repeatedly updating prototypes until assignments stop changing.The assignment uses Euclidean distance.Source: §3.2.1 · p. 12
  • K-means requires the cluster count to be predefined, may converge only to a local optimum, and is sensitive to noise and abnormal data.The authors advise repeated runs with different starting points to seek an optimal solution.Source: §3.2.1 · p. 13

3.2.2. Principal Component Analysis (PCA)

  • PCA projects N-dimensional data into an M-dimensional space, M<N, with least information loss; it maximizes projected variance using eigenvectors associated with the largest eigenvalues.PCA can serve as preprocessing to reduce the complexity of subsequent tasks.Source: §3.2.2 · p. 13

3.3. Reinforcement Learning

Reinforcement learning (RL) learns decisions by interacting with an environment and seeking a policy that maximizes expected discounted return. Deep reinforcement learning (DRL) uses deep networks to represent states and actions when conventional RL struggles with large state and action spaces.

3.3. Reinforcement Learning

  • RL agents observe a state, choose an action, receive a performance-based reward, and seek a policy maximizing expected discounted return.The objective is to maximize expected return over iterations, with the discount rate γ in [0, 1).Source: §3.3 · p. 14
  • Dynamic programming, Monte Carlo methods, and temporal-difference learning are policy-optimization approaches; Q-learning does not require known transition probabilities.The transition probability is often hard to know in practice.Source: §3.3 · pp. 14–15
  • ε-greedy action selection balances exploitation and exploration by choosing the current best action with probability 1−ε and a uniformly sampled action with probability ε.The current best action may not be optimal, while random action selection draws equally from the available action set.Source: §3.3 · p. 15

3.3.1. Deep Reinforcement Learning (DRL)

  • DRL addresses large state and action spaces by replacing a conventional Q-table with deep networks, such as CNNs or RNNs, to represent states and estimate decision values.The network is trained using observed states and corresponding action rewards.Source: §3.3.1 · p. 15

4. Machine Learning for Intelligent Optical Networks Control and Resource Management

Optical networks map service requests to physical-resource configurations subject to QoS, and traffic prediction can support advance network reconfiguration and resource allocation. The surveyed approaches use predicted traffic and resource needs in several settings, from virtual and datacenter networks to hybrid electro-optical datacenters.

4. Machine Learning for Intelligent Optical Networks Control and Resource Management

  • Traffic prediction supports advance network reconfiguration and resource allocation, while optical networks map service requests to physical resources under QoS requirements.The section reviews ML approaches to traffic prediction and resource allocation.Source: §4.1 · p. 15, §4 · p. 15

4.1. Optical Network Traffic and Resource Requirement Prediction

  • In VONs, historical traffic data informs next-slot bandwidth-requirement prediction, and transfer learning modifies the DNN with a small increase in time complexity.When allocated resources significantly mismatch future traffic, the infrastructure provider initiates VON reconfiguration; the method reports better blocking probability and spectrum utilization than fixed-capacity allocation.Source: §4.1 · p. 16
  • A data-poisoning scheme demonstrates that adversarial samples can contaminate DNN training and degrade resource reallocation performance.The described vulnerability concerns DNN training through Machine-Learning-as-a-Service.Source: §4.1 · p. 16
  • A DNN traffic predictor for datacenter networks uses predicted arrival times and resource consumption to adjust traffic queues and lower path blocking probability.The method uses predicted traffic information for resource allocation and aims to improve resource utilization.Source: §4.1 · pp. 16–17
  • For hybrid electro-optical datacenters, NARNN traffic prediction can prioritize future heavy streams for optical switching and offload EPS traffic.The architecture uses OCS for heavy traffic and EPS for short-lived, bursty flows.Source: §4.1 · p. 17
  • LSTM-based traffic remaining-time prediction supports traffic aggregation in optical datacenters, where heavy-tailed holding times make exponential assumptions invalid.LSTM learns the nonlinear relationship between historical traffic and holding-time information.Source: §4.1 · p. 17

4.2. Routing in Optical Networks

Optical-network routing faces utilization and blocking concerns with shortest-path routing, while heuristic methods can be computationally complex at scale. The section surveys supervised and reinforcement-learning approaches, including routing that reacts to contention in optical burst switching (OBS).

4.2. Routing in Optical Networks

  • Conventional shortest-path routing may yield low resource utilization and high blocking, while heuristic route planning becomes computationally complex on large topologies.ML approaches model routing either as supervised classification or regression from historical routes, or as reinforcement-learning decision-making.Source: §4.2 · p. 17

4.2.1. Supervised Learning-based Routing

  • GPRM uses Bayesian-network routing tables to select less-utilized OBS links, and experiments report lower burst loss ratio than SPF.The reported routing objective is reducing burst loss without affecting end-to-end delay.Source: §4.2.1 · pp. 17–18
  • MLRC classifies traffic matrices into precomputed optimal routing solutions and completes routing decisions from traffic-matrix acquisition to SDN flow-rule installation in 80ms.The approach avoids solving an ILP during the routing decision process.Source: §4.2.1 · p. 18
  • For multi-domain routing with private intra-domain status, LSTM learns route-generation rules from sparse historical trajectories and public inputs to return feasible inter-domain routes.The described approach is intended not to divulge private domain information when generating routes.Source: §4.2.1 · p. 18

4.2.2. Reinforcement Learning-based Routing

  • MARL agents coordinate through shared Q-tables and a central server that detects heavily shared links and can select second-best routes; experiments report lower burst loss probability and higher link utilization than a single-agent method.The comparison is reported for the network as a whole.Source: §4.2.2 · pp. 18–19

4.2.3. Reinforcement Learning-based Deflection Routing in OBS

  • RLDRS selects a deflection link with the largest Q-value when wavelength contention occurs, while limiting repeated deflections with a maximum authorized count.The deflection table indexes Q-values by destination and neighbor; bursts exceeding the limit may be dropped.Source: §4.2.3 · p. 19
  • RLDRS, RLAR, and IRLRCR report lower burst loss probability than SPF and SPDR but larger mean burst end-to-end delay than SPF.PQDR outperforms RLDRS in burst loss probability and deflection ratio, but has larger average end-to-end delay, attributed to its five lookup tables and more complicated Q-value updates.Source: §4.2.3 · p. 20

4.3. RWA and RSA in Optical Networks

Route and wavelength assignment (RWA) and routing, modulation, and spectrum assignment (RMSA) allocate multiple optical resources, and RWA and RSA are NP-complete. The surveyed supervised and reinforcement-learning methods learn assignments from network states or interact with the network to select resources.

4.3. RWA and RSA in Optical Networks

  • RWA assigns a route and wavelength per IP service flow, while RSA assigns spectrum slots to a lightpath; both problems are NP-complete.RWA and RSA are typically formulated as constrained optimization problems.Source: §4.3 · p. 20

4.3.1. Supervised Learning-based RWA

  • Supervised RWA learns mappings from network states to ILP-computed optimal configurations, avoiding online ILP computation and enabling time-efficient real-time configuration.Training states include topology, capacity, available wavelengths, and traffic demands.Source: §4.3.1 · p. 20

4.3.2. Reinforcement Learning-based RWA

  • An ingress-node RL method learns path and wavelength selection from successful burst-transmission rewards, and its RWA method outperforms SPF, DWNV, and Self-Learning schemes.The method assumes no wavelength converters at core nodes and does not use deflection routing.Source: §4.3.2 · p. 21
  • A physical-impairment-aware RL RWA method checks wavelength continuity and lightpath QoT, considering four impairments and BER as the QoT metric.It reports lower network blocking probability than shortest-path and uniform path-selection schemes.Source: §4.3.2 · p. 21

4.3.3. Reinforcement Learning-based RSA

  • For EON RMSA, DRL uses a deep Q-network to learn assignment policies from network states and lightpath requests, with CNN layers extracting state features.The considered state includes topology, connectivity, and spectrum utilization.Source: §4.3.3 · p. 21

5. Machine learning for Intelligent Optical Networks Monitoring and Survivability

This section surveys machine-learning applications for optical-network monitoring and survivability, including physical-layer and channel-state-based QoT estimation, failure prediction, identification, and localization. It also covers model-based diagnosis and operator-support approaches, framing network-state awareness as useful for control and restoration.

5.1. Optical Performance Monitoring

  • OPM targets physical-layer impairments such as OSNR, CD, PMD, optical power, and fiber nonlinearity; accurate, real-time monitoring is important to network performance [83, 84].The survey covers OPM in directly detected and digital coherent systems.Source: §5.1 · p. 22

5.1.1. Neural Networks-based OPM techniques

  • Jargon et al. used an ANN with eye-diagram features to estimate OSNR, CD, and PMD simultaneously, demonstrating effectiveness at 10 Gb/s NRZ-OOK and 40 Gb/s RZ-DPSK while assuming impairment-free fibers.Manually extracted input features may require expert knowledge.Source: §5.1.1 · pp. 22–23
  • As an accuracy metric, the reported correlation coefficients were 0.96 with direct detection and 0.995 with balanced detection.Source: §5.1.1 · p. 22
  • CNN-based OSNR estimation with compressed eye diagrams performed better than other ML algorithms, including Decision Trees, KNN, and BP-ANN.The diagrams were converted to grayscale and downsampled before being supplied to the CNN.Source: §5.1.1 · p. 23

5.1.2. PCAs-based OPM

  • PCA reduces the computational burden of image-based OPM by projecting asynchronous delay-tap diagrams into principal components and matching them to reference vectors [92].The impairment values from the reference sample with minimum Euclidean distance are assigned to the estimate.Source: §5.1.2 · p. 23

5.2. Quality of Transmission Estimation

  • ML-based QoT estimation learns relationships between network status and transmission quality from historical data, aiming to reduce uncertainty, excess system margin, and CAPEX.The survey distinguishes lightpath- and channel-level QoT estimation from monitoring physical signal parameters.Source: §5.2 · p. 23

5.2.1. Lightpath QoT Estimation

  • Barletta et al. model whether a lightpath’s BER meets the QoT requirement as a classification task using random forests; neural networks estimate blocking probability in bufferless OBS/OPS networks.The selected random forest had 100 estimators, providing the best trade-off between performance and computational time.Source: §5.2.1 · p. 24
  • A case-based reasoning method estimates lightpath Q-factor by comparing attributes with stored examples, assigning the most similar sample’s value, and classifying against a threshold.Source: §5.2.1 · p. 24
  • The survey notes techniques for optimizing the knowledge base to decrease its complexity.Source: §5.2.1 · p. 24

5.2.2. QoT Related to Channel Usage

  • A method using WDM-channel ON/OFF states reaches relatively low RMSE, while KBR predicts power-excursion STD with low MSE.KBR encodes an ON channel as 1 and an OFF channel as 0.Source: §5.2.2 · pp. 24–25

5.3. Failure Management

  • Failures in optical networks may degrade performance or cause huge data loss; manual operation may not handle massive alarms and invisible failures flexibly or promptly, motivating proactive detection of potential failures.Source: §5.3 · p. 25

5.3.1. Failure Prediction

  • Wang et al. predict board failures in software-defined metropolitan networks by forecasting performance indicators with DES and classifying the forecasts with kernel-based SVM.Historical indicator values are used to predict future values, which the SVM uses to decide whether failure will occur.Source: §5.3.1 · p. 25

5.3.2. Failure Identification

  • ANN-based proactive failure detection outperforms threshold-based reactive detection in accuracy and response time.The response-time measure is the interval from detection to the typical failure state.Source: §5.3.2 · p. 26
  • A Bayesian Network uses pre-FEC BER and received-power statistics to estimate whether a link failure exists and classify its type, after which the system reconfigures the lightpath.Source: §5.3.2 · p. 26

5.3.3. Failure Localization

  • Because one failure can trigger many alarms, DNEN analyzes alarm patterns to localize actual failures in WDM networks.The method extracts hidden failure features and uses crossover and mutation to develop networks until accuracy requirements are met.Source: §5.3.3 · p. 26
  • DNEN achieves the highest failure-localization accuracy versus SVM- and DNN-based methods and is described as time-efficient for practical use.It generates neural networks through crossover and mutation rather than gradient descent, using real optical-network training data.Source: §5.3.3 · p. 26

6. Challenges and possible Solutions

Although ML can improve optical-network efficiency and accuracy, challenges remain for practical deployment. The section identifies barriers across data access, model transparency and generalization, computation, security and reliability, and validation in real networks.

6.1. Open Dataset Access

  • Open datasets enable fairer comparisons and reduce repeated work, but few are available and synthetic-data results may lack credibility and comparability.The survey notes that real-network data are also difficult to collect and open.Source: §6.1 · p. 27
  • Real-network data can be imbalanced because reserved resource margins leave few negative samples, while privacy concerns impede publication.The proposed responses include standardized collection and anonymization, plus federated learning that trains without direct data access.Source: §6.1 · p. 27

6.2. Model Interpretability and Traceability

  • Operators need interpretable monitoring and traceability from outputs to input features, since black-box models can hinder troubleshooting and obscure failure causes.Different failure reasons can require different operational actions.Source: §6.2 · p. 27

6.3. Model Generalization Ability

  • Models are tightly coupled to network environments, so a changed environment can require costly retraining; proposed remedies include local-state features and online learning.The survey also suggests decoupling input features from network state.Source: §6.3 · p. 28

6.4. Algorithm Computational Complexity

  • Because real optical-network tasks are time-sensitive, the survey recommends evaluating algorithms under time thresholds and comparing their suboptimal solutions.Many prior experiments show effectiveness without time restrictions, while few analyze time complexity.Source: §6.4 · p. 28

6.5. System Security and Reliability

  • ML models offer no performance guarantee, creating security and reliability concerns that may hinder real-network use; proposed safeguards include manual intervention and ML-aided operation.The survey also proposes periodic effectiveness evaluation and performance-degradation alarms.Source: §6.5 · pp. 28–29

6.6. Reality Gap Between Simulation and Real Network

  • Most methods have been demonstrated in simulation or small-scale networks, and the gap between emulation and real-network performance is difficult to assess.Security and privacy concerns make real-network testing difficult, while the survey calls for more field tests.Source: §6.6 · p. 29

7. Conclusion

This survey reviews how ML has been used and may be used in intelligent optical networks. It organizes applications across network control and resource management, and monitoring and survivability, while outlining challenges and possible solutions.

7. Conclusion

  • It explains regression, classification, and decision-making paradigms, and analyzes motivations for using ML from algorithm characteristics and enabling techniques.The survey also reviews ML algorithms used in optical networks.Source: §7 · p. 29
  • The survey reviews ML applications in optical-network control and resource management, and in monitoring and survivability.It also discusses background and challenges in communication and optical networks.Source: §7 · p. 29
  • The authors identify challenges for future ML applications in intelligent optical networks and propose possible solutions.They describe the survey as an exploration of how ML has been and should be used in optical networks.Source: §7 · p. 29

RWA and RSA

The supplied entries list reinforcement-learning approaches for routing and routing-modulation-spectrum assignment, each with a stated network setting.

RWA and RSA

  • An RL approach is listed for RWA in the NSFNET setting.Source: RWA and RSA · p. 39
  • The entries distinguish the RWA and RMSA examples by both learning approach and network setting.Source: RWA and RSA · p. 39
  • A DRL approach is listed for RMSA using a link-and-spectrum-usage table on a six-node EON topology.Source: RWA and RSA · p. 39

Related papers

Loading 2003.05290v1…