Scholarlink

The study investigates the task allocation problem of heterogeneous drone swarms in dynamic environments, focusing on internal collaboration. Traditional methods based on contract net protocol algorithms excessively rely on communication, making them unsuitable for extreme scenarios. Despite the potential of reinforcement learning in solving task allocation problems, most current deep reinforcement learning methods are based on Q-networks, which lack the ability to capture global features. Consequently, their performance is suboptimal when addressing complex scenarios. This paper proposes an algorithm of a graph neural network, reinforcement learning, and attention mechanisms. Reinforcement learning is used to address the problem of training the graph neural network with labeled data. The graph neural network is employed to perform graph embedding on task points and drones, capturing global features. The attention mechanism serves as the main strategy network to calculate the relevance scores between drones and task points, and the task allocation for the drone swarm is performed according to a greedy strategy, resulting in the task allocation strategy for individual drones. Each drone performs tasks sequentially based on the shortest distance principle. The experimental results show that the algorithm proposed in this paper has an advantage in solution accuracy and exhibits better solution stability in applications.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Addressing Multi-UAV Task Allocation in Dynamic Scenarios Using a Reinforcement Learning-Driven Graph Neural Network and Attention Mechanism

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Hanrui Zhang,
  • Lei Zuo,
  • Fengchen Zhang,
  • Yu Zhang,
  • Jiaming Bai,
  • Xiao Lv

摘要

The study investigates the task allocation problem of heterogeneous drone swarms in dynamic environments, focusing on internal collaboration. Traditional methods based on contract net protocol algorithms excessively rely on communication, making them unsuitable for extreme scenarios. Despite the potential of reinforcement learning in solving task allocation problems, most current deep reinforcement learning methods are based on Q-networks, which lack the ability to capture global features. Consequently, their performance is suboptimal when addressing complex scenarios. This paper proposes an algorithm of a graph neural network, reinforcement learning, and attention mechanisms. Reinforcement learning is used to address the problem of training the graph neural network with labeled data. The graph neural network is employed to perform graph embedding on task points and drones, capturing global features. The attention mechanism serves as the main strategy network to calculate the relevance scores between drones and task points, and the task allocation for the drone swarm is performed according to a greedy strategy, resulting in the task allocation strategy for individual drones. Each drone performs tasks sequentially based on the shortest distance principle. The experimental results show that the algorithm proposed in this paper has an advantage in solution accuracy and exhibits better solution stability in applications.