Nicola Di Cicco

We experimentally demonstrate a framework and use cases for full lifecycle management of AI-Agent-assisted digital twin optical networks. We achieve 100% accuracy of API-calling by AI-Agent, 7x speed-up alarm-log analysis, and 83% hardware computation resources reduction through LoRA fine-tuning.

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.

Deep Reinforcement Learning (DRL) is rising as a promising tool for solving optimization problems in optical networks. Though studies employing DRL for solving static optimization problems in optical networks are appearing, assessing strengths and weaknesses of DRL with respect to state-of-the-art solution methods is still an open research question. In this work, we focus on Routing and Wavelength Assignment (RWA), a well-studied problem for which fast and scalable algorithms leading to better optimality gaps are always sought for. We develop two different DRL-based methods to assess the impact of different design choices on DRL performance. In addition, we propose a Multi-Start approach that can improve the average DRL performance, and we engineer a shaped reward that allows efficient learning in networks with high link capacities. With Multi-Start, DRL gets competitive results with respect to a state-of-the-art Genetic Algorithm with significant savings in computational times. Moreover, we assess the generalization capabilities of DRL to traffic matrices unseen during training, in terms of total connection requests and traffic distribution, showing that DRL can generalize on small to moderate deviations with respect to the training traffic matrices. Finally, we assess DRL scalability with respect to topology size and link capacity.