Periodic Reporting for period 1 - MENTOR (Machine LEarning in Optical NeTwORks)
Période du rapport: 2021-01-01 au 2022-12-31
To date, optical fibre technologies have been a major driver of our societal progress as they enable global telecommunication and all high-capacity transmission systems.
Optical networks have changed rapidly over the last decade, moving from fixed grid with direct-detection using C-band-only systems to current elastic optical networks with coherent technology and software defined networking (SDN).
Now, C-band systems started being upgraded to C+L. This rapid evolution is a consequence of an incoming exhaustion of the C-band due to the increase of bandwidth demand, mainly caused by Internet video streaming and cloud computing. The exploitation of bands beyond C is crucial and it is seen as one of the most promising techniques.
The EID MENTOR aims to address the most relevant challenges of modern optical communication by combining ML&AI techniques and MB transmission. Its objective is to provide a world class advanced training for six early-stage researchers (ESRs) to become pioneering leaders in this sector.
Thanks to the composition of MENTOR’s consortium, we can access real datasets such as network management and monitoring data, component devices, fibre propagation impairments, and data traffic, essential for the training phase of algorithms.
Alongside Coordinator the Aston Institute of Photonic Technologies, Aston University, the consortium consists of the Polytechnic University of Catalonia (UPC), the Technical University of Denmark (DTU), Sant'Anna School of Advanced Studies (SSSA), Telecom Italia Mobile (TIM), Infinera Germany, Infinera Portugal and Orange.
Optical networks have changed rapidly over the last decade, moving from fixed grid with direct-detection using C-band-only systems to current elastic optical networks with coherent technology and software defined networking (SDN).
Now, C-band systems started being upgraded to C+L. This rapid evolution is a consequence of an incoming exhaustion of the C-band due to the increase of bandwidth demand, mainly caused by Internet video streaming and cloud computing. The exploitation of bands beyond C is crucial and it is seen as one of the most promising techniques.
The EID MENTOR aims to address the most relevant challenges of modern optical communication by combining ML&AI techniques and MB transmission. Its objective is to provide a world class advanced training for six early-stage researchers (ESRs) to become pioneering leaders in this sector.
Thanks to the composition of MENTOR’s consortium, we can access real datasets such as network management and monitoring data, component devices, fibre propagation impairments, and data traffic, essential for the training phase of algorithms.
Alongside Coordinator the Aston Institute of Photonic Technologies, Aston University, the consortium consists of the Polytechnic University of Catalonia (UPC), the Technical University of Denmark (DTU), Sant'Anna School of Advanced Studies (SSSA), Telecom Italia Mobile (TIM), Infinera Germany, Infinera Portugal and Orange.
WP1, Complexity reduction and data augmentation for ML&AI-based optical systems
ESR1 (Sasipim Srivallapanondh) has focused on complexity reduction, the real hardware implementation of the optical equalizers and parallelizability problems of the recurrent-based equalizers, enabling less inference latency with the feed-forward architecture. To reduce the computational complexity of the NN, the nonlinear activation functions can be a crucial factor, especially the LSTM-based networks, which have sigmoid and tanh as their activation functions. The study of the different approaches for the nonlinear activation function was accomplished.
WP2, ML&AI for component characterisation and optimisation
ESR2 (Li Zhang) has worked on time domain power fluctuations modelling, predictions and optimizations over time-variant atmospheric turbulence channels. To describe the complex and changeable turbulent free-space optical channel, and further mitigate the turbulence effects, a gamma gamma time-variant channel has been modelled and predicted by a Gaussian process regression predictor.
WP3, AI-based control and management for ultra-wideband optical networks
ESR3 (Lareb Zar Khan) has focused on improving the quality of training data for ML based management of optical networks, specifically in the area of failure management. Results show that by using data augmentation techniques, the following is possible:
(i) Reduction in training time of neural networks (NNs) used for failure management.
(ii) Improvement in classification accuracy of NNs.
(iii) Reduction in computational complexity during the inference phase.
ESR4 (Abdennour Ben Terki) has been developing the Routing and Spectrum Assignment (RSA) strategies using reinforcement learning (RL) for MB optical networks to improve RSA and network performance. Such strategies are also driven by the Generalized Signal to Noise Ratio (GSNR) metric in a multi-band scenario that accounts for Stimulated Raman Scattering (SRS), a common nonlinear effect in wideband transmission.
WP4, Distributed intelligence for network surveillance
ESR5 (Mariano Devigili) has investigated the exploitation of a digital twin (DG) of the optical layer for failure management and quality of transmission estimation. The models of such a DT are being exploited to design algorithms for failure detection and severity estimation.
ESR6 (Prasunika Khare) presented a computer model for a non-dispersive multichannel optical communication, providing a solution for a non-dispersive SRS in WDM optical fiber system. This work can be extended to multiband (O, E, S, C, L) optical communication.
ESR1 (Sasipim Srivallapanondh) has focused on complexity reduction, the real hardware implementation of the optical equalizers and parallelizability problems of the recurrent-based equalizers, enabling less inference latency with the feed-forward architecture. To reduce the computational complexity of the NN, the nonlinear activation functions can be a crucial factor, especially the LSTM-based networks, which have sigmoid and tanh as their activation functions. The study of the different approaches for the nonlinear activation function was accomplished.
WP2, ML&AI for component characterisation and optimisation
ESR2 (Li Zhang) has worked on time domain power fluctuations modelling, predictions and optimizations over time-variant atmospheric turbulence channels. To describe the complex and changeable turbulent free-space optical channel, and further mitigate the turbulence effects, a gamma gamma time-variant channel has been modelled and predicted by a Gaussian process regression predictor.
WP3, AI-based control and management for ultra-wideband optical networks
ESR3 (Lareb Zar Khan) has focused on improving the quality of training data for ML based management of optical networks, specifically in the area of failure management. Results show that by using data augmentation techniques, the following is possible:
(i) Reduction in training time of neural networks (NNs) used for failure management.
(ii) Improvement in classification accuracy of NNs.
(iii) Reduction in computational complexity during the inference phase.
ESR4 (Abdennour Ben Terki) has been developing the Routing and Spectrum Assignment (RSA) strategies using reinforcement learning (RL) for MB optical networks to improve RSA and network performance. Such strategies are also driven by the Generalized Signal to Noise Ratio (GSNR) metric in a multi-band scenario that accounts for Stimulated Raman Scattering (SRS), a common nonlinear effect in wideband transmission.
WP4, Distributed intelligence for network surveillance
ESR5 (Mariano Devigili) has investigated the exploitation of a digital twin (DG) of the optical layer for failure management and quality of transmission estimation. The models of such a DT are being exploited to design algorithms for failure detection and severity estimation.
ESR6 (Prasunika Khare) presented a computer model for a non-dispersive multichannel optical communication, providing a solution for a non-dispersive SRS in WDM optical fiber system. This work can be extended to multiband (O, E, S, C, L) optical communication.
WP1 AIPT
We used the so-called knowledge distillation method for the topology modification of the NN equalisers in high-speed optical transmission systems. This enables efficient low-complexity digital signal processing solutions, otherwise impractical due to the high complexity introduced by bi-directional recurrent connections. The use of knowledge distillation has allowed us to circumvent the latter problem.
Using NNs in optical communications improves communication systems, achieving a positive socio-economic impact. NNs are used to improve signal processing, channel coding, and modulation, leading to faster and more reliable data transmission. NNs investigated in WP 1 are used to optimize optical communication systems, such as reducing energy consumption and costs, an clear positive socio-economic impact as it helps improve connectivity.
WP 1 aims to engineer a family of simple and highly-adaptive NN-based equalisers (NNE) with reduced energy consumption, and to test those in real systems.
WP2 DTU
ESR2’s project aims to exploit and apply the ML-based techniques in both free-space optical and fiber optical networks, bridging and tightening the connection of space-air-ground integrated communication networks. The results until the end of ESR2’s project is expected to provide efficient and economic solutions for both of simulation realization and experiment implementations, such as channel prediction and power optimization problems in optical networks.
WP3 SSSA
In the current state-of-the-art research in ML aided failure management, there is a lack of focus on data quality which is essential for obtaining optimal performance from ML models. One issue is the class imbalance among different failure samples within the training dataset for training ML models. Research is addressing this issue by producing high-quality and unbiased datasets for training ML models for failure management. This will not only improve the ML-aided failure management but its practical implementation in the industry, the efficiency of ML models, the reliability and robustness of optical networks. RL-based ML will also impact resource allocation supporting higher throughput.
WP4 UPC
Models and algorithms employed so far will be applied to different scenarios, use cases, and tests to identify the optimal configuration of the algorithms. Work will focus on validating employed models and generating new one for quality of transmission estimation, with the aim of enabling effective network optimization and accurate failure management. These achievements would allow the reduction of network margins and avoid the current network capacity overprovisioning, creating a more efficient and sustainable infrastructure behind the internet.
ESR6 will work on the optical NNEs’ complexity and increase the inference speed for the real-time processing and be able to get closer to the real hardware implementation. The generalization of the NNEs will be investigated as this topic is a concerning issue to the industry when it comes to implementation.
We used the so-called knowledge distillation method for the topology modification of the NN equalisers in high-speed optical transmission systems. This enables efficient low-complexity digital signal processing solutions, otherwise impractical due to the high complexity introduced by bi-directional recurrent connections. The use of knowledge distillation has allowed us to circumvent the latter problem.
Using NNs in optical communications improves communication systems, achieving a positive socio-economic impact. NNs are used to improve signal processing, channel coding, and modulation, leading to faster and more reliable data transmission. NNs investigated in WP 1 are used to optimize optical communication systems, such as reducing energy consumption and costs, an clear positive socio-economic impact as it helps improve connectivity.
WP 1 aims to engineer a family of simple and highly-adaptive NN-based equalisers (NNE) with reduced energy consumption, and to test those in real systems.
WP2 DTU
ESR2’s project aims to exploit and apply the ML-based techniques in both free-space optical and fiber optical networks, bridging and tightening the connection of space-air-ground integrated communication networks. The results until the end of ESR2’s project is expected to provide efficient and economic solutions for both of simulation realization and experiment implementations, such as channel prediction and power optimization problems in optical networks.
WP3 SSSA
In the current state-of-the-art research in ML aided failure management, there is a lack of focus on data quality which is essential for obtaining optimal performance from ML models. One issue is the class imbalance among different failure samples within the training dataset for training ML models. Research is addressing this issue by producing high-quality and unbiased datasets for training ML models for failure management. This will not only improve the ML-aided failure management but its practical implementation in the industry, the efficiency of ML models, the reliability and robustness of optical networks. RL-based ML will also impact resource allocation supporting higher throughput.
WP4 UPC
Models and algorithms employed so far will be applied to different scenarios, use cases, and tests to identify the optimal configuration of the algorithms. Work will focus on validating employed models and generating new one for quality of transmission estimation, with the aim of enabling effective network optimization and accurate failure management. These achievements would allow the reduction of network margins and avoid the current network capacity overprovisioning, creating a more efficient and sustainable infrastructure behind the internet.
ESR6 will work on the optical NNEs’ complexity and increase the inference speed for the real-time processing and be able to get closer to the real hardware implementation. The generalization of the NNEs will be investigated as this topic is a concerning issue to the industry when it comes to implementation.