Context-Aware Embedding Masking Based on Reinforcement Learning for Semantic Multiplexing

doi:10.1109/lwc.2026.3718953

Context-Aware Embedding Masking Based on Reinforcement Learning for Semantic Multiplexing

2026-08-03

0

PRE

AI

DOI：10.1109/lwc.2026.3718953

Original

Original request for help

Save

Abstract

En 中文

In shared-embedding (SE) multiple access, <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$U$ </tex-math></inline-formula> users share a single space and user-specific masks provide the attention-based separation that determines the multiplexing capacity. The original SE design uses state-agnostic masks. We instead learn a context-aware masking policy that maps <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$(\mathrm {SNR},U)$ </tex-math></inline-formula> to a mask matrix through a low-rank actor trained by proximal policy optimization, rewarding per-user cosine similarity and mask orthogonality. On real BERT embeddings of AG News, it improves cosine similarity at every SNR and cuts the training-time mask-orthogonality penalty tenfold. Compared with a fixed-orthogonal scheme, the method shows that the gain originates from context-aware adaptation rather than orthogonality, while preserving recovery.

Keywords:

Shared embedding

semantic communications

embedding masking

multiple access

context-aware reinforcement learning

proximal policy optimization

Context-Aware Embedding Masking Based on Reinforcement Learning for Semantic Multiplexing

Abstract

Journal

Organization