Multistation Target Localization via Direction-of-Arrival: An Efficient Multiagent Deep Reinforcement Learning Approach

Accurate direction-of-arrival (DOA) estimation and localization are essential for navigation and positioning. However, most existing model-based approaches are restricted to specific scenarios, which results in poor scalability. In addition, deep learning–based methods require extensive data training. To overcome these challenges, we propose a multistation direction-finding (DF) and localization method based on a multiagent deep deterministic policy gradient (MADDPG) algorithm (DF-MADDPG). In the proposed method, agents monitor target positions by accurately selecting signals from the time–frequency spectrum. First, interference is removed from the received signals through space–time high-resolution processing, and DOA is estimated. Specifically, agents are deployed on the time–frequency spectra to establish an efficient collaborative communication policy, thereby improving the signal-to-noise ratio of signal selection. In addition, we design a reward function that employs a multistation cross-localization strategy to guide agents in quickly and accurately selecting signals for precise localization. To further improve environmental awareness under uncertainty, the time–frequency spectra are treated as states and integrated with the MADDPG algorithm to train agents for target localization. Compared to single-agent reinforcement learning approaches, the proposed DF-MADDPG method improves localization efficiency by 9.5% while maintaining consistent localization accuracy.