Attention-Based MAPPO for Large-Scale Sensor Scheduling in Multisource Localisation

Abstract

Large-scale sensor scheduling for multisource localisation is a critical technology in wireless communication control and navigation systems. Most existing heuristic algorithms face challenges in adapting to large-scale sensor systems. To overcome this limitation, we utilise the self-learning capabilities of deep reinforcement learning (DRL) to enable multisource localisation. This paper proposes a large-scale sensor scheduling algorithm based on the multiagent proximal policy optimisation (LSS-MAPPO) framework. We develop a multisource localisation model based on time difference of arrival (TDOA) and design a reward function grounded in the Cramér–Rao lower bound (CRLB). Our approach integrates multihead attention layers into MAPPO to improve the performance of the algorithm. In large-scale sensor scheduling systems, multihead attention mechanisms can effectively handle the high-dimensional state space associated with multisource localisation in multiagent environments. Experimental results under different environments show that LSS-MAPPO improves localisation accuracy compared to the baseline in large-scale sensor scheduling. Notably, it maintains robust performance under partial observability, addressing critical gaps in large-scale dynamic sensor scheduling.