Decentralized multi-hop data processing in UAV networks using MARL

Unmanned Aerial Vehicles (UAVs) have become integral to numerous applications, prompting research towards enhancing their capabilities. For time-critical missions, minimizing latency is crucial; however, current studies often rely on sending data to ground station or cloud for processing due to their limited onboard capacities. To leverage the networking capabilities of UAVs, recent research focuses on enabling data processing and offloading within the UAV network for coordinated decision-making. This paper explores a multi-hop data offloading scheme designed to optimize the task processing and resource management of UAVs. The proposed distributed strategy uses multi-agent reinforcement learning, where UAVs, each with varying computational capacities and energy levels, process and offload tasks while managing energy consumption and latency. The agents, represented as actor-critic models, learn and adapt their actions based on current state and environment feedback. The study considers a consensus-based method to update learning weights, promoting cooperative behavior among the agents with minimum interaction. Through multiple training episodes, the agents improve their performance, with the overall system achieving faster convergence with high rewards, demonstrating the viability of decentralized data processing and offloading in UAV networks.