Research on Multi-Agent Semantic Communication Framework Based on Comparative Learning Joint Optimization.

With the rapid development of intelligent services, communication objectives are shifting from humans to multi-agent (MA) systems. This transition necessitates new communication paradigms capable of supporting real-time perception, decision-making, and collaboration among agents. Semantic communication (SeC) focuses on the efficient transmission and accurate understanding of information "meaning" and is well-suited to meet the needs of Mas, such as collaborative perception, reasoning, and decision-making. However, the transmission of semantic information is still constrained by dynamic environments and the diversity of MA tasks. To address these challenges, this work proposes a COmparative learning Joint Optimal (COJO) SeC framework. This work makes three main contributions: first, it jointly optimizes the image reconstruction and classification functions designed for multi-task semantic objectives under different channel conditions, thereby improving the overall task performance of the system; second, based on input image features, compression ratio, task requirements, and channel conditions, an enhanced further compressor is designed, which obtains a training-based mask to significantly reduce the volume of transmitted data; finally, to prevent the loss of key semantic information in multi-task scenarios under channel constraints, it designs a task-driven end-to-end semantic communication training scheme.