A transformer-based architecture proposal to drive a self-organizing UAV swarm communication network in emergency scenarios

Abstract

In the aftermath of a natural disaster, telecommunication infrastructures' integrity could be compromised, leading to an unreliable communication service or even a complete blackout in the affected zone. This undesirable outcome increases victims' distress and panic levels. Quick response deployed systems to alleviate the crisis have the potential to save numerous people in danger. Therefore, this work proposes to deploy a multi-UAV swarm network for network coverage services to victims and rescue members, by connecting it with an infrastructure-based telecommunication system through an access point. This work was done in partial fulfillment of the requirements for the master's degree in Computer Science. This study explores solutions to handle UAV's partial observability constraints. Moreover, an encoder-decoder architecture known as Soft Transformer Recurrent Graph Networks (STRGN) was proposed from the insights of the transformer model, typically used for machine translation tasks, and a novel model called Soft Deep Recurrent Graph Networks (SDRGN). The proposal considers information from the agent's subgraph, including current neighbor and ground user positions, to determine the optimal actions to improve ground user coverage, fairness, and network connectivity. Extensive analysis of a variant of the proposed model, Soft Transformer Graph Networks (STGN), demonstrates its effectiveness in solving the Ground User Coverage Problem, outperforming the benchmarked state-of-the-art models. Additionally, we analyze the scalability of our proposed models for various environmental configurations.