POSEIDON: A multi-armed bandit framework for Power and mOdulation SchEme adaptatIon in unDerwater acOustic Networks

Performance optimization of UnderWater (UW) acoustic communications is particularly challenging due to the short coherence time, long propagation delay, and extremely scenario-specific features of the UW channel. These issues clearly call for the implementation of adaptive communication techniques. One effective way to achieve this goal is to rely on Machine Learning (ML), which can easily identify the optimal UW communication parameters to optimize various performance metrics, such as network throughput and energy consumption. However, popular approaches, such as stateful Reinforcement Learning (RL) and Deep Reinforcement Learning (DRL), usually require considerable storage, computing, and energy resources, and may not fit in resource-constrained UW devices. For this reason, in this paper we propose POSEIDON, a framework for transmission power and modulation scheme adaptation based on Multi-Armed Bandit (MAB), a lightweight, stateless alternative to stateful RL and DRL. We evaluate POSEIDON on DESERT, a powerful framework for realistic simulations of UW networks, and compare its performance against two DRL baselines. The results show that, in spite of the lightweight nature of POSEIDON, our framework is able to outperform the DRL baselines by achieving an improvement of up to 72.49% in the network throughput, and up to 77.16% in energy consumption. Moreover, POSEIDON generally exhibits a significantly more stable performance, with fewer oscillations despite the high variability of the UW channel conditions.