ANN-Based Multi-Channel QoT-Prediction Over a 563.4-km Field-Trial Testbed

Abstract

In this paper, artificial neural network (ANN)-based multi-channel Q-factor prediction is investigated with real-time network operation and configuration information over a 563.4-km field-trial testbed. A unified ANN-based regression model is proposed and implemented to predict Q-factors of all the channels simultaneously. A scenario generator is developed to configure the field-trial testbed with 8 testing channels automatically to generate dynamic scenarios. A network configuration and monitoring database (CMDB) is implemented to collect network configuration and monitoring data that include link information, operational parameters of key optical devices, network configuration state, and real-time Q-factors of the available channels for the generated network scenarios. These collected data are used for training and testing of the developed ANN model. In order to achieve multiple channel predictions, we propose a hot coding method to represent the state of dynamic channel. Besides, an auto-search method is used to search the best hyperparameters of the ANN-based model. The results show that the proposed ANN-based regression model converges quickly, and it can predict the multi-channel's Q-factors with high accuracy. The unified ANN-based multi-channel Qfactor regression model can provide the comprehensive information to assist SDN controller to optimize network configuration for dynamic optical networks.Index Terms-artificial neural network, multi-channel Qfactor prediction, field-trial testbed.

Decentralizing machine-learning-based QoT estimation for sliceable optical networks

“…In this work, we have opted for an MLP that is a class of feedforward artificial NNs [37] with low memory and computational requirements. In general, it was shown that NNs achieve higher accuracy than other ML techniques used for finding QoT models [24], and they were also demonstrated experimentally [38,39] achieving an overall high accuracy.…”

“…In such a case, and as shown in the ML literature, the MLP can be set up to successfully handle noise, i.e., by adding noise to the training input [51]. Also, it has been recently shown through field trials that NNs can achieve high QoT estimation accuracy [39]. Even though both the centralized and distributed approaches are expected to be negatively affected by the use of real data (and hence the MLPs will need to be adjusted accordingly), comparatively, it is expected that still the distributed approach will outperform the centralized approach in accuracy, accuracy per class, CPU usage, and RAM requirements per slice type, especially as the number of diverse slice types increases.…”

Toward Deployment of ML in Optical Networks, Transfer Learning, Monitoring and Modelling

“…The two step process therefore is comprised of the initial training of the source learner ANN on synthetic data, followed by the transferring of parameter knowledge to the target learner which is fine-tuned to practical network data. We built on our previous work in [4] and used the ANN architecture as a starting point for the source learner. Synthetic data was gathered on a simulated version of the national dark fibre facility (NDFF) network using route planning software and the Gaussian noise (GN) analytical model [9] .…”

“…With the unparalleled combination of high accuracy and low computational complexity in inference, Machine Learning (ML) based approaches have been explored to provide promising solutions in QoT estimation with either synthetic data [3] or pre-collected network operation data [4] . These solutions, based on artificial neural networks (ANN), face big challenges in scalability as training and inferencing of ML models are carried out on the same network.…”

ANN-based multi-channel QOT-prediction over A 563.4-km field-trial testbed

“…On the other hand, the increasing ratio of mobile-related traffic drives optical networks to be more dynamic in terms of various service times and on-demand bandwidths [5]. Both high-capacity requests and heterogonous user traffic require a dynamic optical network with automatic network Parts of this work appeared in the proceeding of ECOC 2019, Dublin， Ireland [1].…”