Recent advancements in Hollow Core Fibers (HCF) production are paving the way toward new ground-breaking opportunities of HCF for 6G-and-beyond applications. While Standard Single-Mode Fibers (SSMF) have been the go-to solution in optical communications for the past 50 years, HCF is expected to be a turning point in how next-generation optical networks are planned and designed. Compared to SSMF, in which the optical signal is transmitted in a silica core, in HCF, the optical signal is transmitted in a hollow, i.e., air, core, significantly reducing latency (by 30%), while also decreasing attenuation (as low as 0.11 dB/km) and non-linearities. In this study, we investigate the optimal placement of HCF in latency-constrained optical networks to minimize the number of edge Data Centers (edgeDCs), while also ensuring physical-layer validation. Given the optimized placement of HCF and edgeDCs, we minimize the overall network cost in terms of transponders (TXPs) and Wavelength Selective Switches (WSSes) by optimizing the type, number, and transmission mode of TXPs, and the type and number of WSSes. We develop a Mixed Integer Nonlinear Programming (MINLP) model and a Genetic Algorithm (GA) to solve these problems. We validate the GA against the MINLP model in four synthetically generated topologies and perform extensive numerical evaluations in a realistic 25-node metro aggregation topology and a 22-node national topology. We show that by upgrading 25% of the links to HCF, we can significantly reduce the number of edgeDCs by up to 40%, while also reducing network equipment cost by up to 38%, compared to an SSMF-only network.

This work proposes a pipeline for Software Defined Networking (SDN) that enables natural-language-based flow rule configuration using large language models (LLMs). The system addresses two key challenges: 1) the ambiguity and incompleteness of natural language inputs, and 2) the difficulty of reliably translating them into deployable SDN configurations. To this end, the pipeline integrates: i) an intent recognition module that refines user prompts via iterative clarification, and ii) a retrybased correction mechanism that handles failed configurations by regenerating and resubmitting corrected versions. These components are combined with intermediate YAML generation, documentation-based enrichment, and final translation into OpenFlow-compliant JSON for Ryu controllers. The pipeline is evaluated on flow rule deployment tasks of varying complexity, achieving an accuracy up to 96.7%, while maintaining costefficiency with an estimated API cost of only 0.08 per 100 configurations and remaining model model-agnostic.

The ever-increasing automation of complex decisionmaking processes through Operations Research (OR) raises the need for specialized management systems. To support multiple vertical industries, OR service providers must handle a diverse portfolio of solvers having widely varying resource, computational, and availability requirements. To address this fundamental challenge, we present Octopus, a state-of-the-art cloud-native solver orchestrator. Leveraging Kubernetes, Octopus features event-driven autoscaling to align resources with application demand, dedicated queue management to meet service-level agreements, and robust reliability mechanisms to prevent service interruptions. This demonstration showcases Octopus’s capabilities in a production-like environment, illustrating its full end-to-end workflow and commenting on its real-time performance.

Next-generation networks rely on the network softwarization paradigm to enable faster and more cost-effective deployment of telecommunications services. The ETSI MANO framework plays a critical role in orchestrating these networks, yet it faces challenges such as the hidden state problem, arising from the NFVO's lack of holistic visibility into the internal state of NFVI-PoPs, which can lead to the choice of sub-optimal allocation schemes. This work introduces a novel approach to address the hidden state problem by integrating the Digital Twin (DT) paradigm into the MANO architecture. The proposed DT is a model-based solution employing neural networks to predict orchestration costs and estimate prediction errors, enabling the NFVO to make informed orchestration decisions through what-if analyses while preserving scalability and administrative independence. Performance evaluation demonstrates the DT's ability to mimic the behavior of an NFVI-PoP with high precision, i.e., in 84% of the cases, it returns a prediction that is 5% close to the actual value. Furthermore, the DT-aided NFVO achieves orchestration performance equivalent to approaches that assume full knowledge of the actual allocation costs, while overcoming in the 43% of cases traditional benchmark policies.

With the advent of containerization technologies, microservices have revolutionized application deployment by converting old monolithic software into a group of loosely coupled containers, aiming to offer greater flexibility and improve operational efficiency. This transition made applications more complex, consisting of tens to hundreds of microservices. Designing effective orchestration mechanisms remains a crucial challenge, especially for emerging distributed cloud paradigms such as the Compute Continuum (CC). Orchestration across multiple clusters is still not extensively explored in the literature since most works consider single-cluster scenarios. In the CC scenario, the orchestrator must decide the optimal locations for each microservice, deciding whether instances are deployed altogether or placed across different clusters, significantly increasing orchestration complexity. This paper addresses orchestration in a containerized CC environment by studying a Reinforcement Learning (RL) approach for efficient microservice deployment in Kubernetes (K8s) clusters, a widely adopted container orchestration platform. This work demonstrates the effectiveness of RL in achieving near-optimal deployment schemes under dynamic conditions, where network latency and resource capacity fluctuate. We extensively evaluate a multi-objective reward function that aims to minimize overall latency, reduce deployment costs, and promote fair distribution of microservice instances, and we compare it against typical heuristic-based approaches. The results from an implemented OpenAI Gym framework, named as HephaestusForge, show that RL algorithms achieve minimal rejection rates (as low as 0.002%, 90x less than the baseline Karmada scheduler). Cost-aware strategies result in lower deployment costs (2.5 units), and latency-aware functions achieve lower latency (268–290 ms), improving by 1.5x and 1.3x, respectively, over the best-performing baselines. HephaestusForge is available in a public open-source repository, allowing researchers to validate their own placement algorithms. This study also highlights the adaptability of the DeepSets (DS) neural network in optimizing microservice placement across diverse multi-cluster setups without retraining. The DS neural network can handle inputs and outputs as arbitrarily sized sets, enabling the RL algorithm to learn a policy not bound to a fixed number of clusters.

The Edge-Cloud Continuum represents a paradigm shift in distributed computing, seamlessly integrating resources from cloud data centers to edge devices. However, orchestrating services across this heterogeneous landscape poses significant challenges, as it requires finding a delicate balance between different (and competing) objectives, including service acceptance probability, offered Quality-of-Service, and network energy consumption. To address this challenge, we propose leveraging Multi-Objective Reinforcement Learning (MORL) to approximate the full Pareto Front of service orchestration policies. In contrast to conventional solutions based on single-objective RL, a MORL approach allows a network operator to inspect all possible “optimal” trade-offs, and then decide a posteriori on the orchestration policy that best satisfies the system’s operational requirements. Specifically, we first conduct an extensive measurement study to accurately model the energy consumption of heterogeneous edge devices and servers under various workloads, alongside the resource consumption of popular cloud services. Then, we develop a set-based MORL policy for service orchestration that can adapt to arbitrary network topologies without the need for retraining. Illustrative numerical results against selected heuristics show that our MORL policy outperforms baselines by 30% on average over a broad set of objective preferences, and generalizes to network topologies up to 5x larger than training.

Orchestrating microservice applications deployed on a federation of globally distributed Kubernetes clusters is a challenging and multifaceted optimization problem. It is not only computationally hard, but also requires balancing a delicate trade-off between competing performance metrics, such as latency, deployment cost, and service interruption frequency. Classical approaches in the literature merge multiple objectives into a single one via, e.g., linear combinations. However, in practice, it is complex to express a priori a quantitative preference between heterogeneous objectives, let alone with simple linear combinations. This paper adopts a more comprehensive approach leveraging proper Multi-Objective Optimization (MOO), with the goal of producing multiple solutions from the Pareto Front (PF). Therefore, the orchestrator can inspect a posteriori all possible "optimal" trade-offs and decide on the strategy that best fits their operating requirements. To solve the MOO problem, this paper adopts state-of-the-art Multi-Objective Evolutionary Algorithms and shows their effectiveness in solving the MOO problem. Illustrative results highlight the practical benefits of a MOO formulation, providing several tens of nondominated solutions and evenly covering the objectives’ space.

Machine Learning (ML) models have found numerous applications in the automation of complex network management tasks. More recently, thanks to the introduction of new solutions for data-plane programmability (as the P4 programming language), it has become possible for programmable switches to execute ML-models directly in the data plane, with the great advantage that decisions can now be taken at packet-rate, without the involvement of the control plane. Existing works have shown that tree-based ML models, such as Random Forest, can be implemented on P4 switches, despite strict constraints on the available computational and memory resources. However, the (common, and practical) case when multiple models must be concurrently implemented to perform different tasks is still under-investigated. Assigning separate, dedicated resources (i.e., stages in the packet-processing pipeline) to each model can be very inefficient. In this study we propose a new resource-efficient data-plane implementation of multiple concurrent tree-models that share input features. We focus on the problems of DDoS-attack detection and application-traffic identification and demonstrate high accuracy in both problems while saving up to 40% of the required processing stages.

Optical metro networks currently support various traffic demands with different bit-rates, ranging from low values, e.g., 1 Gbps and 10 Gbps, to high values, e.g., 100 Gbps and 200 Gbps. These traffic demands can be served through coexistence of non-coherent transmission technology (mostly 10 Gbps) or by coherent high-rate technology (100 Gbps and above), characterized by different transmission requirements (e.g., in terms of Signal-to-Noise Ratio (SNR)). To achieve a low-cost metro architecture, various technical directions can be followed: (i) traffic grooming can be employed to decrease the number of line transmission interfaces (at the cost of increased Optical-Transport-Network (OTN) grooming boards), (ii) filterless nodes can reduce the node cost and power consumption by replacing costly Wavelength Selective Switches (WSS) with passive splitters and combiners, and (iii) amplifiers placement can be optimized, benefiting from short distances in metro areas. In this paper, we observe, for the first time to the best of our knowledge, that traffic grooming and amplifier placement are interdependent problems if we aim to achieve overall network cost minimization. Therefore, we propose and compare two cost-effective cross-layer optimization approaches that jointly consider the optical and OTN layers. Precisely, we propose two Quality-of-Transmission (QoT) aware approaches that optimize deployment cost of OTN grooming boards and interfaces in OTN layer while guaranteeing SNR and power on receiver of lightpaths as QoT metrics by considering placement of optical amplifiers along fibers in optical layer. The results indicate that our proposed approaches can save up to 40% compared to real-world baseline solutions.

Efficient service provisioning in the Edge-to-Cloud Continuum is of utmost importance for modern applications. While sensible decisions can be taken if enough monitoring data is collected, maintaining continuous telemetry data streams amidst the continuum’s complexity is challenging. This paper introduces CRISP (reConstructing Resource Information for Service Placement), a solution combining data reconstruction and service placement strategies to optimize decisions despite incomplete monitoring data. CRISP utilizes Convolutional Neural Networks and Long Short-Term Memory models for data reconstruction, integrating them with a heuristic algorithm that selects nodes for service component placement. Numerical results demonstrate CRISP’s efficacy in optimizing service provisioning despite missing data, contributing to enhanced resource utilization and service performance in the considered context.

Microservices have revolutionized application deployment in popular cloud platforms, offering flexible scheduling of loosely-coupled containers and improving operational efficiency. However, this transition made applications more complex, consisting of tens to hundreds of microservices. Efficient orchestration remains an enormous challenge, especially with emerging paradigms such as Fog Computing and novel use cases as autonomous vehicles. Also, multi-cluster scenarios are still not vastly explored today since most literature focuses mainly on a single-cluster setup. The scheduling problem becomes significantly more challenging since the orchestrator needs to find optimal locations for each microservice while deciding whether instances are deployed altogether or placed into different clusters. This paper studies the multi-cluster orchestration challenge by proposing a Reinforcement Learning (RL)-based approach for efficient microservice deployment in Kubernetes (K8s), a widely adopted container orchestration platform. The study demonstrates the effectiveness of RL agents in achieving near-optimal allocation schemes, emphasizing latency reduction and deployment cost minimization. Additionally, the work highlights the versatility of the DeepSets neural network in optimizing microservice placement across diverse multi-cluster setups without retraining. Results show that DeepSets algorithms optimize the placement of microservices in a multi-cluster setup 32 times higher than its trained scenario.

We investigate classifying hardware failures in microwave networks via Machine Learning (ML). Although MLbased approaches excel in this task, they usually provide only hard failure predictions without guarantees on their reliability, i.e., on the probability of correct classification. Generally, accumulating data for longer time horizons increases the model’s predictive accuracy. Therefore, in real-world applications, a trade-off arises between two contrasting objectives: i) ensuring high reliability for each classified observation, and ii) collecting the minimal amount of data to provide a reliable prediction. To address this problem, we formulate hardware failure-cause identification as an As-Soon-As-Possible (ASAP) selective classification problem where data streams are sequentially provided to an ML classifier, which outputs a prediction as soon as the probability of correct classification exceeds a user-specified threshold. To this end, we leverage Inductive and Cross Venn-Abers Predictors to transform heuristic probability estimates from any ML model into rigorous predictive probabilities. Numerical results on a real-world dataset show that our ASAP framework reduces the time-to-predict by 8x compared to the state-of-the-art, while ensuring a selective classification accuracy greater than 95%. The dataset utilized in this study is publicly available, aiming to facilitate future investigations in failure management for microwave networks.

We consider the problem of classifying hardware failures in microwave networks given a collection of alarms using Machine Learning (ML). While ML models have been shown to work extremely well on similar tasks, an ML model is, at most, as good as its training data. In microwave networks, building a good-quality dataset is significantly harder than training a good classifier: annotating data is a costly and time-consuming procedure. We, therefore, shift the perspective from a Model-Centric approach, i.e., how to train the best ML model from a given dataset, to a Data-Centric approach, i.e., how to make the best use of the data at our disposal. To this end, we explore two orthogonal Data-Centric approaches for hardware failure identification in microwave networks. At training time, we leverage synthetic data generation with Conditional Variational Autoencoders to cope with extreme data imbalance and ensure fair performance in all failure classes. At inference time, we leverage Batch Uncertainty-based Active Learning to guide the data annotation procedure of multiple concurrent domain-expert labelers and achieve the best possible classification performance with the smallest possible training dataset. Illustrative experimental results on a real-world dataset show that our Data-Centric approaches allow for training top-performing models with ~4.5x less annotated data, while improving the classifier’s F1-Score by ~2.5% in a condition of extreme data scarcity. Finally, for the first time to the best of our knowledge, we make our dataset (curated by microwave industry experts) publicly available, aiming to foster research in data-driven failure management.

Deep Reinforcement Learning (DRL) is being investigated as a competitive alternative to traditional techniques for solving network optimization problems. A promising research direction lies in enhancing traditional optimization algorithms by offloading low-level decisions to a DRL agent. In this study, we consider how to effectively employ DRL to improve the performance of Local Search algorithms, i.e., algorithms that, starting from a candidate solution, explore the solution space by iteratively applying local changes (i.e., moves), yielding the best solution found in the process. We propose a Local Search algorithm based on lightweight Deep Reinforcement Learning (DeepLS) that, given a neighborhood, queries a DRL agent for choosing a move, with the goal of achieving the best objective value in the long term. Our DRL agent, based on permutation-equivariant neural networks, is composed by less than a hundred parameters, requiring only up to ten minutes of training and can evaluate problem instances of arbitrary size, generalizing to networks and traffic distributions unseen during training. We evaluate DeepLS on two illustrative NP-Hard network routing problems, namely OSPF Weight Setting and Routing and Wavelength Assignment, training on a single small network only and evaluating on instances 2x-10x larger than training. Experimental results show that DeepLS outperforms existing DRL-based approaches from literature and attains competitive results with state-of-the-art metaheuristics, with computing times up to 8x smaller than the strongest algorithmic baselines.

We consider the problem of forecasting the Quality-of-Transmission (QoT) of deployed lightpaths in a Wavelength Division Multiplexing (WDM) optical network. QoT forecasting plays a determinant role in network management and planning, as it allows network operators to proactively plan maintenance or detect anomalies in a lightpath. To this end, we leverage Bayesian Recurrent Neural Networks for learning uncertainty-aware probabilistic QoT forecasts, i.e., for modelling a probability distribution of the QoT over a time horizon. We evaluate our proposed approach on the open-source Microsoft Wide Area Network (WAN) optical backbone dataset. Our illustrative numerical results show that our approach not only outperforms state-of-the-art models from literature, but also predicts intervals providing near-optimal empirical coverage. As such, we demonstrate that uncertainty-aware probabilistic modelling enables the application of QoT forecasting in risk-sensitive application scenarios.

We consider the problem of designing and training a neural network-based orchestrator for fog computing service deployment. Our goal is to train an orchestrator able to optimize diversified and competing QoS requirements, such as blocking probability and service delay, while potentially supporting thousands of fog nodes. To cope with said challenges, we implement our neural orchestrator as a Deep Set (DS) network operating on sets of fog nodes, and we leverage Deep Reinforcement Learning (DRL) with invalid action masking to find an optimal trade-off between competing objectives. Illustrative numerical results show that our Deep Set-based policy generalizes well to problem sizes (i.e., in terms of numbers of fog nodes) up to two orders of magnitude larger than the ones seen during the training phase, outperforming both greedy heuristics and traditional Multi-Layer Perceptron (MLP)-based DRL. In addition, inference times of our DS-based policy are up to an order of magnitude faster than an MLP, allowing for excellent scalability and near real-time online decision-making.

We make a case for in-network Continual Learning as a solution for seamless adaptation to evolving network conditions without forgetting past experiences. We propose implementing Active Learning-based selective data filtering in the data plane, allowing for data-efficient continual updates. We explore relevant challenges and propose future research directions.

Artificial Intelligence (AI) has demonstrated superhuman capabilities in solving a significant number of tasks, leading to widespread industrial adoption. For in-field network-management application, AI-based solutions, however, have often risen skepticism among practitioners as their internal reasoning is not exposed and their decisions cannot be easily explained, preventing humans from trusting and even understanding them. To address this shortcoming, a new area in AI, called Explainable AI (XAI), is attracting the attention of both academic and industrial researchers. XAI is concerned with explaining and interpreting the internal reasoning and the outcome of AI-based models to achieve more trustable and practical deployment. In this work, we investigate the application of XAI for network management, focusing on the problem of automated failure-cause identification in microwave networks. We first introduce the concept of XAI, highlighting its advantages in the context of network management, and we discuss in detail the concept behind Shapley Additive Explanations (SHAP), the XAI framework considered in our analysis. Then, we propose a framework for a XAI-assisted ML-based automated failure-cause identification in microwave networks, spanning model’s development and deployment phases. For the development phase, we show how to exploit SHAP for feature selection and how to leverage SHAP to inspect misclassified instances during model’s development process, and how to describe model’s global behavior based on SHAP’s global explanations. For the deployment phase, we propose a framework based on predictions uncertainty to detect possibly wrong predictions that will be inspected through XAI.