发表机构
University of Macau; Shenzhen Institute of Artificial Intelligence and Robotics for Society (AIRS), School of Science and Engineering, The Chinese University of Hong Kong (CUHK), Shenzhen（澳门大学; 香港大学深圳校区科学与工程学院人工智能与机器人研究院）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
该研究针对时延与能量受限无线视觉上行链路，提出依赖感知可靠性分配方法，通过语义依赖拉格朗日块算法优化分组参数，在多数实例达全局最优，性能优于多种对比方法。
AI 中文摘要
无线视觉感知上行链路越来越多地传输结构化语义分组，而非原始图像或不透明特征张量。在开放词汇场景图中，只有当索引的主语-关系-宾语三元组的分组及其所有必备词汇增量分组均被成功恢复时，该三元组才具有可解释性，这一依赖结构是传统分组丢失、层优先级或语义特征HARQ规则无法捕捉的。我们在期望帧时延和帧能量约束下，研究此类分组的依赖感知可靠性分配问题。该模型将词汇增量分组与索引三元组分离开，构建它们的必备图，并推导Chase合并HARQ残余失败、感知反馈的时延与能量，以及依赖感知语义失真。所得有限离散问题需联合选择每个分组的调制与编码方案、发射功率和最大HARQ深度。与固定优先级UEP不同，词汇分组会获得一个共享的下游启用值，该值取决于其他必备分组和三元组分组的当前可靠性。利用所得的多仿射失真结构，所提出的语义依赖拉格朗日块方法对每个分组更新执行精确的有限候选最小化。在对30个精简菜单实例的小规模穷尽审计中，该方法在28个实例中达到全局最优，观测到的平均和最坏情况相对间隙分别为0.052%和1.10%。通过分析链路预算扫描、公开场景图轨迹、多会话与多种子验证，以及查表BLER测试，结果表明，与依赖无关HARQ、语义优先级HARQ、优先级UEP、仅UEP和均匀HARQ相比，该方法能降低词汇诱导的语义失败率，提高关键查询的成功率。
英文摘要
Wireless visual uplinks increasingly carry structured representations for edge inference, making packet reliability part of task-aware link adaptation. In open-vocabulary scene-graph transmission, an indexed triplet is usable only if both its triplet packet and the vocabulary packets defining any newly introduced tokens are recovered. This prerequisite coupling makes the marginal value of packet reliability depend on neighboring packet reliabilities. We formulate a dependency-aware semantic distortion and jointly optimize finite choices of modulation and coding scheme (MCS), transmit power, and Chase-combining hybrid automatic repeat request (HARQ) depth under expected delay and energy constraints. The distortion is multi-affine in packet failure probabilities and cannot, in general, be reduced to static separable unequal error protection (UEP) weights when prerequisites are active. This structure yields a state-dependent reliability coefficient and explicit switching thresholds among wireless actions. A Lagrangian block method performs exact per-packet finite-action updates for fixed multipliers. On reduced instances, it matches exhaustive optimization in 28 of 30 cases, with a worst gap of 1.095%. On GQA traces using a table-driven block error rate (BLER) abstraction, it reduces mean semantic distortion by 58.71% and grounded-query failure by 57.03% relative to dependency-agnostic HARQ under the same budgets.
CommentsSubmitted to IEEE Transactions on Wireless Communications
