Overview

The rapid growth of emerging network applications and their stringent quality-of-service requirements are driving today’s backbone networks to evolve toward multi-domain elastic optical networks (MD-EONs), which can support high-capacity and user-customized end-to-end services across multiple autonomous systems (ASes) [1]. To realize resource-efficient service provisioning in MD-EONs, accurate quality-of-transmission (QoT) modeling techniques are indispensable. In this context, recent studies have reported a number of machine learning (ML)-aided cognitive QoT estimation designs that can model complex network dynamics (e.g., dynamic traffic profiles) and uncertainties (e.g., uncertain device conditions) using big data analytics [2], [3]. However, these existing works focus on single-domain scenarios and cannot be applied to MD-EONs, where only a very limited amount of domain information is available due to domain privacy concerns. Therefore, we lately proposed a hierarchical learning approach for QoT estimation in MD-EONs [4], where domain managers (DMs) work cooperatively with a broker plane by learning domain-level and inter-domain QoT estimators, respectively. Nonetheless, a major challenge remaining unmet is that the approach entails a significant amount of performance monitoring data for every inter-domain lightpath, which can be very costly and unscalable. Fortunately, the invention of transfer learning (TL) has enabled to significantly reduce the amount of efforts required for training a ML task by reusing knowledge learned from relevant tasks [5]. The application of TL for QoT estimation was first studied in [6]. However, the work in [6] only adopts a very simple and straightforward TL scheme, and more importantly, does not address the challenges in MD-EONs.

Current Research

We propose Evol-TL, an evolutionary transfer leaning approach, for enabling scalable QoT estimation in MD-EONs. Evol-TL exploits a broker-based MD-EON architecture, where a broker plane performs end-to-end QoT estimation by collecting encoded features from DMs. Fig. 1 depicts the schematic of Evol-TL in an MD-EON with broker orchestration and the principle of knowledge transfer in Evol-TL. Specifically, A generic algorithm (GA) is designed to enable Evol-TL to determine the proper neural network architectures and the right sets of parameters for transferring through iterative optimizations, whereas the previous works mostly rely on human experiences or brute-force searches. Evaluations with experimental data show that Evol-TL can transfer knowledge between similar tasks and significantly reduce the amount of required training data for new tasks without sacrificing the estimation accuracies.

Fig. 1. (a) Schematic of Evol-TL in an MD-EON with broker orchestration; (b) Principle of knowledge transfer in Evol-TL.
Fig. 2. (a) Two-domain EON testbed implementation; (b) convergence process of Evo-TL; (c-d) loss vs training epochs (c) with Evo-TL and (d) without Evo-TL. (number of training data instances being 30); (e) accuracies with different numbers of data instances; (f) results of required training data instances to threshold (accuracies above 90%) and asymptotic accuracy. Task 1: training of the QoT estimator for the four-node path. Task 2: training of the QoT estimator for the five-node path.

References

[1] X. Chen, R. Proietti, H. Lu, A. Castro and S. J. B. Yoo, “Knowledge-Based Autonomous Service Provisioning in Multi-Domain Elastic Optical Networks,” in IEEE Communications Magazine, vol. 56, no. 8, pp. 152-158, August 2018.

[2] L. Barletta, A. Giusti, C. Rottondi, and M. Tornatore, “QoT Estimation for Unestablished Lighpaths using Machine Learning,” in Optical Fiber Communication Conference, OSA Technical Digest (online) (Optical Society of America, 2017), paper Th1J.1.

[3] Roberto Proietti, Xiaoliang Chen, Kaiqi Zhang, Gengchen Liu, M. Shamsabardeh, Alberto Castro, Luis Velasco, Zuqing Zhu, and S. J. Ben Yoo, “Experimental Demonstration of Machine-Learning-Aided QoT Estimation in Multi-Domain Elastic Optical Networks with Alien Wavelengths,” J. Opt. Commun. Netw. 11, A1-A10 (2019)

[4] Gengchen Liu, Kaiqi Zhang, Xiaoliang Chen, Hongbo Lu, Jiannan Guo, Jie Yin, Roberto Proietti, Zuqing Zhu, and S. J. Ben Yoo, “Hierarchical Learning for Cognitive End-to-End Service Provisioning in Multi-Domain Autonomous Optical Networks,” J. Lightwave Technol. 37, 218-225 (2019)

[5] S. J. Pan and Q. Yang, “A Survey on Transfer Learning,” in IEEE Transactions on Knowledge and Data Engineering, vol. 22, no. 10, pp. 1345-1359, Oct. 2010.

[6] W. Mo et al., “ANN-Based Transfer Learning for QoT Prediction in Real-Time Mixed Line-Rate Systems,” 2018 Optical Fiber Communications Conference and Exposition (OFC), San Diego, CA, 2018, pp. 1-3.

Leave a Reply

Your email address will not be published. Required fields are marked *