Adaptive Sound Velocity Profile Prediction Method Based on Deep Reinforcement Learning

The construction of an underwater sensor network system necessitates the swift and cost-effective acquisition of precise sound velocity profile data, which is crucial for enhancing underwater positioning accuracy and guaranteeing secure underwater wireless communication. This letter introduces an adaptive sound velocity profile prediction method premised on deep reinforcement learning (DRL-ASP). The DRL-ASP algorithm dynamically interacts with the environment and adjusts its behavior to maximize environmental rewards. During this interactive process, it continuously updates the sound velocity profile, leading to improved accuracy in predicting the sound velocity profile. The efficacy of the DRL-ASP algorithm was evaluated through experiments conducted on datasets from the South China Sea, Arctic Ocean, and Southern Ocean. Its performance in terms of sound velocity error, mean square error, fitting accuracy, and relative error was compared against empirical formula sound velocity prediction methods,
$Q$
-learning-based sound velocity prediction methods, and convolutional-neural-network-based sound velocity prediction methods. The results affirmed that the DRL-ASP algorithm exhibits superior prediction accuracy, enhanced search efficiency, and robust stability.