Answers are grounded in this paper and cite their sources.
Accurate optical signal-to-noise ratio (OSNR) estimation is important for QoT-aware lightpath provisioning, but machine-learning estimators face scarce target labels and distribution shifts across heterogeneous historical domains. We study multi-source few-sample OSNR regression using the simulation-based Fraunhofer HHI QoT dataset collection. A physics-guided feature audit yields a compact 14-dimensional representation, and domain-discrepancy analysis reveals substantially different source-target compatibility between the CONUS and TSNN domains. We propose RA-MoE, a risk-aware mixture-of-experts framework that preserves source-specific predictors as historical experts and combines them with a target-local expert through fixed reliability-oriented weights derived from source-target discrepancy and target-sample availability. Lightweight target-side calibration further adapts the retained source knowledge using limited labeled target samples. In the cross-topology mismatch regime, RA-MoE achieves the lowest RMSE over the 200–800 few-sample budgets and reduces the RMSE by 69.4% relative to pooled-source transfer learning at the 400-sample budget. Across two complementary transfer regimes and seven few-sample budgets, RA-MoE achieves the best average RMSE rank and the smallest mean and maximum relative gaps to the best method, demonstrating stable performance across the evaluated source-target compatibility conditions.
No takes yet. Share an insight, caveat, or question.
Wang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: