PARouting: Prediction-supported adaptive routing protocol for FANETs with deep reinforcement learning

Abstract

Flying Ad-hoc Networks (FANETs) are becoming increasingly popular for various applications. Effective routing protocols for FANETs are essential yet challenging due to the high dynamic nature of Unmanned Aerial Vehicles (UAVs). Most existing routing protocols require the periodic broadcast of Hello packets to maintain neighbor tables that store the locations of neighbors, mobility patterns, etc. However, the frequent exchange of Hello packets leads to a large routing overhead in FANETs. This paper proposes PARouting, a prediction-supported adaptive routing protocol with Deep Reinforcement Learning, which introduces a novel UAV mobility prediction algorithm using Deep Learning (DL-UMP) to estimate the locations of UAVs. Based on DL-UMP, we design an adaptive Hello packet mechanism to realize on-demand broadcasting of Hello packets, which reduces routing overhead. The routing process is formulated as a Partially Observable Markov Decision Process, and a new Q-network structure is proposed to select the optimal next hop. Simulation results confirm the accuracy of the DL-UMP and show that PARouting outperforms benchmark routing protocols in terms of packet delivery rate, end-to-end delay, and routing overhead.

Keywords

• Mobility prediction
• Routing protocol
• Deep reinforcement learning
• Flying Ad-hoc networks
• Partially observable markov decision process
• Routing overhead