A Knowledge Distillation Based Reinforcement Learning Scheme for Multi-AUV Trajectory Planning in Network Localization

The effectiveness of data collected by underwater acoustic sensor networks (UASNs) relies on their accurate position information, making localization a fundamental technology in oceanic engineering. However, the harsh underwater environment and high maintenance costs pose significant challenges for UASNs localization. Utilizing an autonomous underwater vehicle (AUV) as a mobile anchor presents a viable solution. Moreover, schemes that use multiple AUVs can perform better than those using a single AUV. Despite this, issues such as low localization accuracy and overlong movement trajectories persist. In this paper, we propose a knowledge distillation for modified multi-agent deep deterministic policy gradient (KD-MMADDPG) scheme for multi-AUV trajectory planning. This scheme not only enhances in oceanic engineering by reducing the length of trajectories, but also improves the localization accuracy of the sensor nodes. With the proposed scheme, the modified multi-agent deep deterministic policy gradient (MMADDPG) algorithm generates training data and trains the decision network. Subsequently, knowledge distillation is applied to simplify the decision network based on MMADDPG. Simulation results show that the KD-MMADDPG scheme outperforms the existing schemes, GSCAN and GTRUN, in terms of localization accuracy and trajectory length.