LE-MHAPPO-Enhanced DNN Task Partitioning in Energy-Harvesting Heterogeneous UAV Swarms

In recent years, multi-Unmanned Aerial Vehicle (UAV) swarms have been extensively deployed as mobile edge computing (MEC) systems to support delay-sensitive applications in emergency response, often integrated within Internet of Things (IoT) networks to enhance data collection and processing capabilities. Deep neural network (DNN) partition-assisted collaborative inference has garnered significant attention due to its ability to mitigate the computing and energy limitations of individual UAVs. However, most studies on UAV collaborative inference overlook the challenges posed by high-speed and unpredictable topology changes within UAV swarms. These dynamics require UAVs to make effective action decisions for future time periods based on current conditions. To tackle this issue, we formulate the multi-dimensional action decision problem of DNN partitioning in highly dynamic and energy-harvesting heterogeneous UAV swarms as a predictive Markov decision process (P-MDP) framework. By leveraging the long short-term memory (LSTM)-enhanced multi-agent hybrid action proximal policy optimization (LE-MHAPPO) algorithm, each UAV predicts hybrid action decisions using centralized training and distributed execution (CTDE). The simulation results demonstrate that the proposed algorithm lowers the weighted sum of task delay and energy consumption by 13.8% compared to the state-of-the-art MHAPPO. Furthermore, its performance degradation over time is only 27.7% of that observed in the baseline.