Learning and Adaptation for Efficient and Resilient Internet of Underwater Things Communications

The advent of the Internet of Underwater Things (IoUT) is reshaping communication and computation in marine environments. Unlike terrestrial wireless systems, where abundant spectrum and cloud resources enable complex processing and centralized optimization, UnderWater Acoustic (UWA) networks must operate under extremely harsh and constrained conditions. The acoustic channel suffers from scarce and frequency-dependent bandwidth, long propagation delays, and high variability caused by multipath and Doppler effects. At the same time, network nodes are typically battery-powered devices deployed for long periods without maintenance, thus subject to severe energy and processing limitations. In this context, the efficient management of communication resources is paramount, as it directly impacts reliability, throughput, latency, and energy consumption. Developing both model-based solutions and adaptive learning mechanisms to cope with these challenges is therefore essential to unlock the full potential of IoUT applications.Model-based resource allocation strategies—such as predictive channel models based on Markov chains, Hidden Markov models, or Kalman filters—play a fundamental role in providing valuable insights into the temporal evolution of the channel and in supporting proactive transmission adaptation. However, their effectiveness is limited when dealing with highly irregular and fast-varying dynamics that are difficult to capture through static or purely mathematical models. For this reason, model-based solutions are complemented in this dissertation by learning-based approaches, where intelligent algorithms progressively adapt to network dynamics by exploiting online observations rather than relying on predefined assumptions. In particular, lightweight methods such as Multi-Armed Bandit (MAB) frameworks emerge as a promising alternative to more computationally demanding techniques like Deep Reinforcement Learning (DRL), as they enable fast adaptation while meeting the stringent energy and processing constraints of UWA devices. This makes it possible to design resource allocation mechanisms that efficiently exploit the few available resources while ensuring robustness and scalability in IoUT deployments.Hence, learning-based algorithms—and in particular lightweight Machine Learning (ML) techniques such as MABs—can play a fundamental role in enabling adaptive resource allocation in UWA networks. By exploiting online feedback and progressively refining their decisions, these methods can dynamically adjust transmission parameters to the ever-changing characteristics of the acoustic channel. However, this relation is not one-directional: the effectiveness of such learning approaches also depends on the efficient management of scarce network resources. Only through carefully optimized power usage, modulation selection, and feedback scheduling can ML-driven adaptation remain feasible on battery-powered IoUT devices. This interplay highlights a dual perspective—where learning enhances resource allocation, and resource allocation, in turn, sustains the deployment of learning itself in constrained UWA environments—which constitutes the core paradigm guiding this dissertation.Recently, the evolution of learning paradigms towards distributed approaches—where the decision process is carried out in parallel at multiple devices—has proved particularly appealing for UWA networks. In this setting, each node can autonomously adapt its transmission strategy based on local observations, thereby reducing the computational burden on centralized entities, limiting signaling overhead, and inherently improving system robustness by avoiding single points of failure. This perspective naturally aligns with the design of Multi-Player Multi-Armed Bandit (MP-MAB) frameworks and hierarchical Age of Information (AoI)-aware strategies explored in this dissertation. Nonetheless, the extreme scarcity of energy and communication resources in IoUT nodes requires that such distributed learning processes be carefully supported by appropriate resource management frameworks. Transmission power, modulation schemes, and feedback intervals must therefore be judiciously allocated to ensure that the learning process remains sustainable: fast convergence and reliable adaptation must be achieved without depleting the limited energy budget of the network.According to this perspective, a two-fold paradigm naturally emerges in UWA communications. On the one hand, learning-based techniques such as MABs and AoI-aware frameworks can significantly enhance resource allocation by dynamically adapting modulation, power, and feedback strategies to harsh and time-varying channel conditions. On the other hand, efficient resource allocation is itself a prerequisite for sustaining the learning process, as only careful management of energy, bandwidth, and signaling overhead can make such algorithms feasible on resource-constrained IoUT nodes. This dual interplay between learning and adaptation represents the central theme of this dissertation and guides the organization of its main contributions.Accordingly, the dissertation is structured into three main parts. The first part addresses model-based approaches for UWA communications, presenting predictive channel models derived from real experimental traces through Markov, Hidden Markov, and Kalman filtering techniques, and highlighting their accuracy–complexity trade-offs. The second part focuses on learning-based solutions, where lightweight MAB frameworks are employed for adaptive modulation, distributed decision-making, and AoI-aware feedback scheduling, demonstrating their effectiveness compared to DRL baselines while remaining feasible for resource-constrained IoUT devices. The third part turns to routing strategies, introducing bandit-inspired and reinforcement learning methods to optimize next-hop selection, balance energy consumption, and improve resilience against jamming attacks. Together, these contributions outline a comprehensive path from model-based prediction to learning-driven adaptation, ultimately advancing the efficiency, robustness, and scalability of IoUT networks, and paving the way for future real-world deployments, including testbed validation and the integration of semantic communication paradigms.