EUVIMEDEUVIMED entdecken
Crossref

Uncertainty-guided Bayesian continual learning for robust nonlinear impairment estimation in evolving optical networks

Huazhi Lun, Lei Shi, Haoran Hu, Xingyu Wang, Huicun Yu, Jie Tang, Ying Liu

Journal of Optical Communications and Networking · 01.08.2026 · 18 (8) · 832

DOI: 10.1364/jocn.595882 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Optical Communications and Networking, online erschienen am 21.07.2026. Verfasst von Huazhi Lun und 6 weiteren Autor:innen. Die Arbeit nennt 22 Literaturangaben. Gefördert von: National Natural Science Foundation of China.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Machine learning models for nonlinear impairment (NLI) estimation in optical networks suffer from model aging: as network conditions evolve, catastrophic forgetting causes accumulated performance degradation that neither static retraining nor one-shot fine-tuning can reliably prevent. We propose an uncertainty-guided Bayesian continual learning (CL) framework that enables an NLI estimator to sequentially adapt to new network states while preserving knowledge of prior operating regimes. By modeling network parameters as mean-field Gaussian variational posteriors, the framework enables two synergistic mechanisms from a single uncertainty signal: uncertainty-weighted regularization that softly penalizes changes to high-certainty parameters and uncertainty-based gradient masking that hard-protects the most reliable weights. We validate the framework on a multi-stage lifecycle simulation of the NSFNET topology, spanning mixed modulation formats, multiple symbol rates, and flex-grid channel spacings. The proposed method maintains stable long-term estimation accuracy throughout the sequence, substantially outperforming fine-tuning in both knowledge retention and forgetting suppression. An ablation study confirms that the two mechanisms are synergistic, jointly achieving significantly higher knowledge retention than either component alone. These results demonstrate a practical path toward trustworthy, lifelong ML estimators for autonomous optical network management.

Förderung

  • National Natural Science Foundation of China

Zitieren

APA

Lun, H., Shi, L., Hu, H., Wang, X., Yu, H., Tang, J., & Liu, Y. (2026). Uncertainty-guided Bayesian continual learning for robust nonlinear impairment estimation in evolving optical networks. Journal of Optical Communications and Networking, 18(8), 832. https://doi.org/10.1364/jocn.595882

Vancouver

Lun H, Shi L, Hu H, Wang X, Yu H, Tang J, et al. Uncertainty-guided Bayesian continual learning for robust nonlinear impairment estimation in evolving optical networks. Journal of Optical Communications and Networking. 2026;18(8):832. doi:10.1364/jocn.595882

BibTeX
@article{Lun2026,
  author = {Lun, Huazhi and Shi, Lei and Hu, Haoran and Wang, Xingyu and Yu, Huicun and Tang, Jie and Liu, Ying},
  title = {Uncertainty-guided Bayesian continual learning for robust nonlinear impairment estimation in evolving optical networks},
  journal = {Journal of Optical Communications and Networking},
  year = {2026},
  volume = {18},
  number = {8},
  pages = {832},
  doi = {10.1364/jocn.595882}
}
Kostenlos starten →