作者
Komal Yadav,Durga Prasad Palaparthi,Gaurav Garg,Harshitha H B,Thella Preethi Priyanka,S. Kaliappan
摘要
Managing autonomous swarms of drones in dynamic and uncertain conditions is a considerable challenge, especially in high-entropy conditions, where autonomous swarms form disordered formations and unequal spacing, and present a higher likelihood of collisions. To solve this issue, there is a need to have communication effective and uncertainty conscious multi-agent learning schemes that are able to create stability without compromising on scalability. In order to address these shortcomings, this paper presents a new framework called MAPPO-GAT, which is an entropy-sensitive swarm coordination model incorporating Multi-Agent Proximal Policy Optimization (MAPPO) and a Graph Attention Network (GAT)-based communication system. The suggested solution improves the information communication because it allows drones to focus on the most meaningful neighbors, and an entropy-driven reward system punishes disorder and encourages the formation control. The Drone-Swarm-Entropy Dataset, a subset of Kaggle, is used to conduct experiments, and it consists of different flight patterns with varying degrees of disturbance. Quantitative findings prove that MAPPO-GAT is much more successful than MAPPO, MADDPG, QMIX, and COMA, the mission success, coverage, entropy reduced and the rate of collision minimally lowered. Specifically, MAPPO-GAT has the highest success rate of 93%, collision rate of 11% and stable swarm formations than all the baselines. The framework is also highly energy efficient, by reducing the needless drone motions due to attention coordination. On the whole, this paper introduces the first MARL-GAT hybridized system that is tailored to the specific task of entropy-conscious swarm intelligence, creating a high-sized, resilient, and communication-efficient base of autonomous drone systems of the next generation.