Hybrid Reinforcement Learning for Resource Allocation in VQA-Oriented UAV Semantic Offloading

Visual question answering (VQA) plays an important role in enhancing situational awareness, while unmanned aerial vehicles (UAV) provide flexible platforms to support such semantic communication tasks. This paper investigates resource allocation for VQA in UAV-assisted semantic communication networks. In the proposed framework, both UAV and user equipment (UE) adopt deep semantic communication (DeepSC)-VQA to encode semantic features, which are then offloaded to a multi-access edge computing (MEC) server for inference. To jointly optimize the energy consumption and the latency of the whole system, we develop a hybrid proximal policy optimization (HPPO) algorithm with a tailored reward function that coordinates the UAV trajectory, radio resource allocation, and semantic offloading. Simulation results show that the HPPO-based scheme consistently outperforms conventional reinforcement learning benchmarks, achieving a superior balance between the energy consumption and the latency.