Resource allocation for UAV-assisted anti-jamming semantic D2D networks: A graph reinforcement learning approach

Semantic communication (SemCom) breaks the limitations of traditional communication methods in semantic understanding and processing, and provides more efficient and intelligent information exchange for wireless networks. In this paper, we investigate an unmanned-aerial-vehicle (UAV)-assisted anti-jamming device-to-device (D2D) network with SemCom, aiming to maximize the quality of experience (QoE) of mobile users (MUs) in the presence of malicious jammers. The UAV serves as a relay to improve the space-extensibility of D2D networks. The formulated problem is challenging to be solved due to its non-convex and stochastic nature. Therefore, we model the problem as a Markov decision process and address it by designing a proximal policy optimization (PPO)-based deep reinforcement learning (DRL) framework to tackle the high-dimensional hybrid action space. To address the irregular and dynamic network topologies in D2D networks, we introduce heterogeneous graph neural networks (GNNs) into the DRL agent to enhance its feature extraction capability over the wireless links. Extensive numerical results demonstrate that the proposed GPPO approach outperforms the multi-layer-perception-based PPO scheme, and effectively maximizes the QoE of MUs under various scenarios.