Optimizing Source Localization via Reinforcement Learning in Multi-Agent Underwater Networks

In this paper, we propose a novel approach to multi-agent underwater source localization by means of Multi-Agent Reinforcement Learning (MARL). Our framework optimizes the trajectories of two autonomous underwater vehicles, each towing an antenna, to maximize the probability of detection of the source. We implement a shared-parameter MARL strategy with non-synchronous actions to address the challenges posed by non-stationary multi-agent environments. We train a neural network on a simplified simulation environment and evaluate it in a realistic simulation engine, demonstrating robustness to communication losses of up to 60%. Our preliminary results indicate that RL-based trajectory optimization can achieve comparable performance to traditional approaches.