Cooperative DNN Partitioning in Energy-Harvesting and MEC-Enabled UAV Networks

Unmanned Aerial Vehicles (UAVs) are critical in modern emergency response due to their high mobility. However, limited computing resources and energy supplies necessitate the use of UAV networks for collaborative inference. UAV intelligent tasks are often Deep Neural Networks (DNN)-based, with DNN partitioning enabling collaborative inference. However, executing DNN partitioning in a highly dynamic UAV network faces two challenges that have not been addressed in existing research: the time gap between the state sampling and the execution of the corresponding action based on that state, and the unknown trajectories in advance. The time gap requires predictive action decision-making. To address this, we model DNN partitioning and edge offloading with hybrid action decisions in dynamic, energy-harvesting UAV networks as a Predictive Markov Decision Process (P-MDP). The rapidly changing and previously unknown network topology significantly impacts channel and data transmission energy consumption, affecting DNN partitioning decisions. To better solve the action prediction problem, we use the Transformer module to extract motion features from recent time slots in the proposed Transformer-enhanced Multi-Agent Hybrid Action Proximal Policy Optimization (TE-MHAPPO) framework. Simulation results show that TE-MHAPPO reduces the reward which comprehensively considers task delay and energy consumption, by at least 12.1% compared to the state-of-theart MHAPPO. Additionally, its reward performance degradation with the increase in prediction time is at most 55.2% of that observed in the baseline.