Adaptive UAV Swarm Management: Minimizing Energy Consumption with Multi Agent DRL

Abstract

Unmanned Aerial Vehicles (UAVs) are indispensable in disaster scenarios, particularly for the communication and data collection services of a post-disaster field area. Nevertheless, the utility of UAVs is often limited due to constraints on resources and energy consumption. A service that deals most efficiently with data collection and management from terrestrial IoT devices through the UAV swarm for achieving energy optimization and path planning is proposed by this study. We take into account the dynamic swarm behaviour involving UAVs joining and leaving, to accommodate for changing mission conditions. The primary objective is to minimize total energy expenditure while adhering to constraints related to mission duration and transmission power. Given the NP-hard nature of this optimization problem, we introduce a Multi-Agent Double Reinforcement Learning (MA-DRL) approach, which ensures rapid convergence and resilience to sudden environmental changes. Extensive simulations demonstrate that our MA-DRL framework significantly outperforms traditional algorithms, including Multi-Agent Deterministic Policy Gradient (MA-DPG), Deep Q-Learning (DQL), and Q-learning, particularly in energy efficiency and data collection capabilities. Our findings underscore the MA-DRL framework's effectiveness for real-time decision-making in dynamic and mission-critical contexts, paving the way for improved UAV swarm operations in disaster recovery and other challenging scenarios.