基于图神经网络的多智能体路径规划方法

禹鑫燚; 刘飞; 欧林林

文章摘要

禹鑫燚,刘飞,欧林林.基于图神经网络的多智能体路径规划方法[J].高技术通讯(中文),2024,34(10):1081~1090

基于图神经网络的多智能体路径规划方法

A multi-agent path planning method based on graph neural network

DOI：10. 3772 / j. issn. 1002-0470. 2024. 10. 007

中文关键词: 路径规划; 多智能体强化学习; 图神经网络(GNN); 多智能体通信

英文关键词: path planning, multi-agent reinforcement learning, graph neural network(GNN), multi-agent communication

基金项目:

作者	单位
禹鑫燚	(浙江工业大学信息工程学院杭州 310023)
刘飞
欧林林

摘要点击次数: 306

全文下载次数: 385

中文摘要:

在多智能体路径规划问题中,每个智能体需要互相协调来完成共同的全局目标,智能体之间通常需要显式的通信策略。传统的多智能体路径规划算法受限于实时性、扩展性、不完全通信等问题,很难适用于复杂的工作环境中。为了解决多智能体工作环境中的通信问题,本文提出了一种基于图神经网络(GNN)的路径规划方法。该方法首先通过卷积神经网络(CNN)在局部观测中采集特征数据,由图神经网络在智能体之间传递这些数据。其次,为了减少智能体的惰性,提出了一种新的奖励函数,鼓励智能体更积极地探索并学习有效的协调策略。接着通过集中式收集数据训练、分布式执行提高学习效率。最后,进行多个环境下的仿真实验评估本文提出的算法,并与其他算法进行对比,验证了算法的有效性和可扩展性。

英文摘要:

For the multi-agent path planning problem, each agent needs to coordinate with each other to accomplish a global goal. Explicit communication strategies are usually required between the agents. The traditional multi-agent path planning algorithms are limited by its insufficiency of instantaneity, scalability, and incomplete communica- tion, which are difficult to be applied in complex environments. In order to solve the communication problem in multi-agent working environment, a path planning method based on graph neural network(GNN) is proposed. The method collects feature data in local observation by convolutional neural network (CNN), and GNN transmits these data among the agents. Second, to reduce the inertia of the agents, a new reward structure is proposed to encourage the agents to explore and learn effective coordination strategies more actively. Then, the learning efficiency is im- proved by centralized data collection for training and distributed execution. Finally, simulation experiments in dif- ferent environments are conducted to evaluate the algorithm proposed in this paper and compare it with other algo- rithms to verify its effectiveness and scalability.

查看全文查看/发表评论下载PDF阅读器

关闭