Joint Semantic-Physical Layer Framework for Goal-Oriented Communication
A novel framework has been developed that combines semantic and physical layers to facilitate goal-oriented communication, safeguarding task-relevant data through a unified constellation design. This system employs a vector quantized-variational autoencoder to identify discrete latent concepts, while a semantic criticality indicator (SCI) evaluates each concept's relevance to the task. Additionally, a deep reinforcement learning agent dynamically chooses transmission subsets based on prevailing channel conditions. In the physical layer, a learned semantic-aware M-QAM constellation allocates symbol positions based on joint co-occurrence statistics and SCI ratings, moving away from uniform spacing and Gray coding. This strategy enhances the protection of critical symbols against channel errors, thereby improving task performance compared to existing decoupled approaches.
Key facts
- Proposed joint semantic-physical layer framework for goal-oriented communication
- Uses vector quantized-variational autoencoder for discrete latent concept extraction
- Semantic criticality indicator (SCI) scores each concept by task relevance
- Deep reinforcement learning agent selects transmission subset based on channel conditions
- Learned semantic-aware M-QAM constellation assigns symbol positions using co-occurrence statistics and SCI scores
- Departs from uniform spacing and Gray coding of standard constellations
- Aims to protect task-relevant information from channel errors more effectively
- Published on arXiv with ID 2605.14940
Entities
Institutions
- arXiv