In this work we consider the problem of fault localization in transparent optical networks. We attempt to localize single-link failures by utilizing statistical machine learning techniques trained on data that describe the network state upon current and past failure incidents. In particular, a Gaussian process classifier is trained on historical data extracted from the examined network, with the goal of modeling and predicting the failure probability of each link therein. To limit the set of suspect links for every failure incident, the proposed approach is complemented by the utilization of a graph-based correlation heuristic. The proposed approach is tested on a number of datasets generated for an orthogonal frequency-division multiplexing-based optical network, and demonstrates that the approach achieves a high localization accuracy (91%-99%) that is insignificantly affected as the size of the historical dataset is reduced. The approach is also compared to a conventional fault localization method that is based on the utilization of monitoring information. It is shown that the conventional method significantly increases the network cost, as measured by the number of monitoring nodes required to achieve the same accuracy as that achieved by the proposed approach. The proposed scheme can be used by service providers to reduce the network cost related to the fault localization procedure. As the approach is generic and does not depend on specific network technologies, it can be applied to different network types, e.g., fixed-grid or space-division multiplexing elastic optical networks.

This work proposes a data-driven bandwidth allocation (BA) framework for periodically and dynamically reconfiguring an elastic optical network according to predictive bandwidth allocation (PBA) models. The proposed framework is scalable to the number of network connections and also adaptive to the increasing traffic of each network connection separately and to the overall network load as well. This is achieved by formulating the BA problem as a Partially Observable Markov Decision Process (POMDP), which constitutes are reinforcement learning (RL) model. Specifically, RL is performed continuously and independently (locally) for each network connection according to the most recent data that describe the traffic demand behavior of each network connection. A central controller monitors the network performance that is jointly achieved from all the PBA models and is capable of dynamically modifying the reward function of the POMDP, ensuring that the quality-of-service requirements are met. A reward function R(C) is examined with a clear impact on the network performance when C is modified. For evaluating the network performance, for each R(C), the routing and spectrum allocation (RSA) problem is solved according to an Integer Linear Programming (ILP) algorithm and an RSA heuristic alternative, with both the ILP and the heuristic RSA taking as inputs the outputs of the inferred PBA models. Results indicate that with the appropriate settings of C, bandwidth is efficiently allocated, while ensuring that the QoS requirements are met.

We examine the problem of bandwidth allocation (BA) on flexible optical networks in the presence of traffic demand uncertainty. We assume that the daily traffic demand is given in the form of distributions describing the traffic demand fluctuations within given time intervals. We wish to find a predictive BA (PBA) model that infers from these distributions the bandwidth that best fits the future traffic demand fluctuations. The problem is formulated as a Partially Observable Markov Decision Process and is solved by means of Dynamic Programming. The PBA model is compared to a number of benchmark BA models that naturally arise after the assumption of traffic demand uncertainty. For comparing all the BA models developed, a conventional routing and spectrum allocation heuristic is used adhering each time to the BA model followed. We show that for a network operating at its capacity crunch, the PBA model significantly outperforms the rest on the number of blocked connections and unserved bandwidth. Most importantly, the PBA model can be autonomously adapted upon significant traffic demand variations by continuously training the model as realtime traffic information arrives into the network.

No abstract

We apply deep graph convolutional neural networks for Quality-of-Transmission estimation of unseen network states capturing, apart from other important impairments, the inter-core crosstalk that is prominent in optical networks operating with multicore fibers.