Multiagent Multitarget Search with Multi-Head Attention

In swarm drone applications, the problem of multitarget self-organizing search in unknown environments has demonstrated immense potential. In this problem, agents need to collaboratively search for multiple dynamic targets while only observing their immediate surroundings. However, the computational complexity and uncertainty in traditional multiagent search processes have been critical factors limiting their performance. Therefore, this article focuses on the multitarget search problem. First, a multiagent search environment including agents, targets, and obstacles is constructed. Subsequently, this article proposes a multiagent multitarget search method (MASOS) in complex obstacle environments. This method integrates the Actor–Critic reinforcement learning algorithm and multi-head attention for UAV swarm collaborative control, enabling agents to focus more on critical information such as other agents, targets, and obstacles during observation. It aims to enhance search, collaboration, and obstacle avoidance capabilities in multitarget self-organizing search tasks. Experimental results show that the MASOS method outperforms other commonly used multiagent reinforcement learning algorithms in terms of coordination strategies. In large-scale self-organizing search tasks, the capture success rate of the MASOS method is nearly 100%. Finally, By designing multitask scenario experiments, conducting comparative analyses combining multilayer perceptron and multi-head attention mechanism, and performing ablation studies on key modules, the effectiveness and interpretability of the MASOS method are verified from two aspects: Performance and decision-making logic.