Deep Reinforcement Learning-Based Task Offloading With Collaborative Inference in UAV-Assisted Mobile Edge Computing Networks

Intelligent air-ground integration communication is an emerging technology. Uncrewed aerial vehicles (UAVs) serve as mobile edge computing (MEC) servers in large-scale Internet of Things (IoT) applications, alleviating the computational load on ground users. Existing multi-UAV MEC approaches struggle with the complex computation and large data sizes of deep neural network tasks. To address these challenges, we propose a Deep Reinforcement Learning (DRL)-based DNN Partitioning and Dynamic Trajectory Selection (DPDTS) method, which reduces end-to-end latency and system energy consumption through task offloading and collaborative inference. Specifically, we propose an Optimal Partition Point Selection (OPPS) algorithm to minimize transmission overhead by selecting optimal partition points for DNN tasks. Then, we design a fairness-based matching algorithm to optimize user offloading and resource allocation. Finally, OPPS and matching algorithms are integrated to optimize UAV flight trajectories and user transmission power via DRL. The simulation results show that DPDTS outperforms existing benchmark methods in terms of delay and energy efficiency.