Calibrated Probabilistic QoT Regression for Unestablished Lightpaths in Optical Networks

Quality-of-Transmission (QoT) regression of unestablished lightpaths is a fundamental problem in Machine Learning applied to optical networks. Even though this problem is well-investigated in current literature, many state-of-the-art approaches either predict point-estimates of the QoT or make simplifying assumptions about the QoT distribution. Because of this, during lightpath deployment, an operator might take either overly-aggressive or overly-conservative decisions due to biased predictions. In this paper, we leverage state-of-the-art Gradient Boosting Decision Tree (GBDT) models and recent advances in uncertainty calibration to perform QoT probabilistic regression for unestablished lightpaths. Calibration of a regression model allows for an accurate modeling of the QoT Cumulative Distribution Function (CDF) without any prior assumption on the QoT distribution. In our illustrative experimental results, we show that our calibrated GBDT model's predictions provide accurate confidence interval estimates, even when only few samples per lightpath configuration are available at training time.