Sample-distribution-matching-based transfer learning for QoT estimation in optical networks

Abstract

Machine learning has been introduced to estimate the quality of transmission (QoT) of lightpaths to meet the reliability of optical network transmission. In the early deployment phase of optical networks, it is difficult to collect enough training samples due to the insufficient lightpaths and monitoring equipment, which makes the estimation models inaccurate. Transfer learning (TL) has been demonstrated as a promising technology for improving the accuracy of estimation models. The main idea of the TL method is pre-training the QoT estimation model with the source domain samples, and then using a few target domain samples to fine-tune the models, which are named fine-tuning samples. However, there are many differences in sample distribution between the source and target domain networks, and more fine-tuning samples to train the TL models are required. Thus, how to improve the accuracy of TL-based QoT estimation models with few samples needs to be examined. This paper proposes a sample-distribution-matching-based transfer learning (SDMbTL) method to perform an accurate QoT estimation with fewer samples. The proposed SDMbTL method designs the sample distribution matching model to filter source domain samples to match the distribution of the target domain samples, which makes the pre-training model more suitable for target networks. We also propose three different matching algorithms to accommodate the discriminative features to be matched. The simulation results demonstrate that the performance of the proposed estimation models outperforms the traditional TL models, saving more than 28.5% of fine-tuning samples.