Joint Optimization for Semantic-Awared Communication and Control: A GDM-Empowered DRL Approach
Lijie Zhou, Shiyi Gu, Yu Zhou, Wenjing Li · 2025
With the advancement of industrial intelligence, control systems are increasingly demanding higher real-time performance and accuracy, particularly in the face of growing network data volumes. Semantic communication has shown considerable promise in improving transmission efficiency and minimizing latency, making it a promising approach for future communication systems. However, due to the lack of a unified and comprehensive theoretical framework, the relationship between semantic communication and control performance remains unexplored, as well as semantic-aware resource allocation. To deal with these challenges, we first analyze and derive the relationship between communication delay and control stability and define the transmission delay in semantic communication systems using the DeepSC model as an example. Then, we formulate a semantic-aware resource allocation problem aimed at maximizing semantic similarity through joint optimization of channel assignment, power allocation, and semantic compression ratio across multiple devices. To overcome the inefficiencies of traditional mathematical methods and deep reinforcement learning (DRL) algorithms in solving complex optimization problems, we propose a two-stage Generative Diffusion Model (GDM)empowered DRL algorithm framework which decouples the solution space by leveraging GDM to generate resource allocation scheme and optimal semantic compression parameter is solved subsequently with exhausted searching. The simulation results confirm the effectiveness of the proposed two-stage approach and highlight the advantages of the two-stage GDM-enhanced DRL algorithm over both the single-stage approach and the standalone DRL algorithm.