MENTOR-EID

ESR3 Lareb Zar Khan and ESR4 Abdennour Ben Terki at SSSA worked on Work Package 3.

MAIN ACHIEVEMENTS (ESR3):

• Data augmentation technique was proposed to address data scarcity for training ML models for failure management in optical networks, and findings were shared with the research community through publications in top-tier conferences and a journal.
• Model-centric and data-centric approaches were investigated experimentally to overcome the issue of limited availability of failure data for training ML models, and consequently optimizing the failure recovery process in optical networks. The results were published in a top-tier journal.
• A novel approach was proposed to significantly reduce the computational complexity of neural networks, which in turn reduces the energy consumption and cost of running ML models. The results were presented at the Optical Network Design and Modeling (ONDM) conference in 2024.
• A sophisticated data pre-processing approach was proposed that converts raw alarms data from commercial network management systems to a ML-friendly form. The effectiveness of the proposed approach was presented at the Optical Network Design and Modeling (ONDM) conference in 2023.
• Unsupervised ML technique was investigated to address the issue of unavailability of labelled failure data for distinguishing relevant failure alarms from false alarms. Findings were presented at PSC conference in Japan.

Computational complexity reduction of ML models will impact not just optical networks but the world of AI. AI/ML-based data centres consume one order of magnitude more power than traditional data centres which means that larger ML models consume high power. Our work on computational complexity reduction of ML models can be used to minimize the carbon footprint, analysing and predicting equipment failures, vendor-agnostic solutions for alarm management, and developing ML-based algorithms for failure prediction; there are contributors in maintaining a stable network operation, minimizing downtime. The potential here is to revolutionize optical network management, reducing operational costs and enhancing service provisioning. This contributes to advancing automation and optimization technologies.

The potentials of machine learning (ML) have been investigated to enhance control and management functionalities for optical networks, even assuming multi-band optical networks. Such functionalities span from resource allocation strategies accounting for transmission performance to (soft-)failure management. Several ML models have been proposed, demonstrated, and tested even experimentally. First, reinforcement learning and deep reinforcement learning techniques have been proposed to optimize routing and spectrum assignment in multi-band optical networks, while accounting for transmission performance in wide-band transmission systems in the presence of Stimulated Raman Scattering (SRS). Then, failure management (e.g., failure identification) assisted by neural networks (NNs) has been studied facing with issues coming from live networks (e.g., data scarcity associated to specific failures). ML models have been proposed, also relying on generative AI techniques such as data augmentation. A specific attention has been also given to the NN-complexity reduction to create sustainable networks even in terms of power consumption. Overall, activities in WP3 demonstrated that ML can actually empower control and management functionalities revolutionizing optical networks control and operation.

Key results have been published in:

• L. Z. Khan, J. Pedro, N. Costa, A. Sgambelluri, A. Napoli and N. Sambo, “Model and data-centric machine learning algorithms to address data scarcity for failure identification,” in Journal of Optical Communications and Networking, vol. 16, no. 3, pp. 369-381, March 2024. Open Access link here.
• L. Z. Khan, J. Pedro, N. Costa, L. De Marinis, A. Napoli and N. Sambo, “Data augmentation to improve performance of neural networks for failure management in optical networks,” in Journal of Optical Communications and Networking, vol. 15, no. 1, pp. 57-67, January 2023. Open Access link here.

MAIN ACHIEVEMENTS (ESR4):

• Conducted a comprehensive literature review on routing and spectrum assignment (RSA) in multi-band optical networks using Machine Learning techniques, resulting in a paper at Photonics in Switching and Computing (PSC) Italy, in 2023.
• Proposed novel RSA techniques based on Reinforcement Learning and Deep Reinforcement Learning.
• Simulated and validated the proposed techniques, comparing them against heuristic methods.
• ML model to identify the type, severity, and localization of failures.

ESR4’s work focused on applying Reinforcement Learning (RL) and Deep Reinforcement Learning (DRL) to enhance RSA in multi-band optical networks. This research leveraged the Generalized Gaussian Noise (GGN) model to simulate physical layer impairments accurately, crucial for assessing Quality of Transmission (QoT) and optimizing network performance. Results showed that RL and DRL strategies adapt dynamically to network conditions, effectively minimizing blocking probability and improving throughput across various network topologies compared to k-Shortest Path and First-Fit (k-SP FF) strategy. Then, a multi-stage ML technique has been proposed and demonstrated to infer with high accuracy the type, severity, and localization of failures in an experimental testbed based on commercial equipment.

Key results have been published in:

• A. B. Terki, J. Pedro, A. Eira, A. Napoli and N. Sambo, “Deep Reinforcement Learning for Resource Allocation in Multi-Band Optical Networks,” 2024 International Conference on Optical Network Design and Modelling (ONDM 2024), Madrid, Spain, 2024. Open Access link here.
• A. B. Terki, J. Pedro, A. Eira, A. Napoli and N. Sambo, “Routing and spectrum assignment based on reinforcement learning in multiband optical networks,” Photonics in Switching and Computing (PSC), (Mantova, Italy, 26-29th September 2023. Open Access link here.