Hierarchical Learning Framework for Swarm Coordination

The proposed architecture addresses the complexity of search and rescue operations by decomposing decision-making into three distinct levels. This hierarchical approach prevents the computational bottlenecks typically associated with centralized control while avoiding the chaotic behavior of purely decentralized systems.

• Level 1 (Individual Agent Level): Focuses on local perception, obstacle avoidance, and basic navigation. This level uses reinforcement learning to ensure each UAV can react to its immediate environment in real-time without waiting for global instructions.
• Level 2 (Swarm/Group Level): Manages local coordination between sub-groups of UAVs. This layer handles task allocation and collision avoidance within clusters, ensuring that agents working in the same sector don't duplicate efforts.
• Level 3 (Global Mission Level): Oversees the entire swarm's objective, such as area coverage, target identification, and resource management. This level processes high-level mission parameters and distributes sub-goals down to the lower levels.

Balancing Autonomy and Global Objectives

The core innovation of this architecture is the feedback loop between levels. By allowing the global mission level to dynamically adjust the reward functions of the individual agents, the system can adapt to changing conditions in a search and rescue environment—such as discovering a new area of interest or losing an agent to battery failure. This structure allows the swarm to maintain mission continuity even when communication with a central base is intermittent or degraded. The authors argue that this multi-tiered approach significantly improves search efficiency compared to flat, non-hierarchical learning models, as it reduces the state-space complexity that each individual agent needs to navigate.