Freshness-Aware Robust Cooperative Task Offloading under Stale Observations in Dynamic SAGINs
Freshness-Aware Robust Cooperative Task Offloading under Stale Observations in Dynamic SAGINs
Lei Tao 1Shihe Xu 2Xinming Zhang2
作者信息
- 1. Shenzhen Polytechnic University
- 2. University of Science and Technology of China
- 折叠
摘要
Space--air--ground integrated networks (SAGINs) enable computation-constrained ground users to access heterogeneous computing resources on terrestrial base stations, unmanned aerial vehicles, and low-Earth-orbit satellites. Although recent offloading studies address dynamic topology and delayed system state, receiver-specific asynchronous resource reports remain a practical source of partial observability: a user may act on stale or missing server load, computing-capacity, and bandwidth information. We study cooperative indivisible task offloading under this information imperfection. We explicitly model report delay, loss, cached state, missingness, and age of information (AoI), and formulate the sequential decision process as a decentralized partially observable Markov decision process. We then develop FARO, a freshness-aware robust offloading framework that augments a local heterogeneous graph with AoI and validity features, summarizes observation histories into a decision-oriented recurrent belief, and learns cooperative policies under centralized training and decentralized execution. Rewards are evaluated from the true environment state, whereas each actor uses only its local stale observation. Experiments isolate observation delay, report loss, resource-state volatility, component contributions, and user-population shift without test-time fine-tuning. At a 40\% status-report loss rate, FARO achieves 0.34 utility and an 88\% task-success ratio, compared with 0.26 and 80\% for the strongest non-oracle ablation. A policy trained with 30 GUs also generalizes to 50 GUs without fine-tuning, attaining 0.24 per-GU utility and 80\% success with 56.0~ms batched actor latency.
Abstract
Space--air--ground integrated networks (SAGINs) enable computation-constrained ground users to access heterogeneous computing resources on terrestrial base stations, unmanned aerial vehicles, and low-Earth-orbit satellites. Although recent offloading studies address dynamic topology and delayed system state, receiver-specific asynchronous resource reports remain a practical source of partial observability: a user may act on stale or missing server load, computing-capacity, and bandwidth information. We study cooperative indivisible task offloading under this information imperfection. We explicitly model report delay, loss, cached state, missingness, and age of information (AoI), and formulate the sequential decision process as a decentralized partially observable Markov decision process. We then develop FARO, a freshness-aware robust offloading framework that augments a local heterogeneous graph with AoI and validity features, summarizes observation histories into a decision-oriented recurrent belief, and learns cooperative policies under centralized training and decentralized execution. Rewards are evaluated from the true environment state, whereas each actor uses only its local stale observation. Experiments isolate observation delay, report loss, resource-state volatility, component contributions, and user-population shift without test-time fine-tuning. At a 40\% status-report loss rate, FARO achieves 0.34 utility and an 88\% task-success ratio, compared with 0.26 and 80\% for the strongest non-oracle ablation. A policy trained with 30 GUs also generalizes to 50 GUs without fine-tuning, attaining 0.24 per-GU utility and 80\% success with 56.0~ms batched actor latency.关键词
space--air--ground integrated networks/computation offloading/age of information/stale observations/multi-agent reinforcement learning/Dec-POMDPKey words
space--air--ground integrated networks/computation offloading/age of information/stale observations/multi-agent reinforcement learning/Dec-POMDP引用本文复制引用
Lei Tao,Shihe Xu,Xinming Zhang.Freshness-Aware Robust Cooperative Task Offloading under Stale Observations in Dynamic SAGINs[EB/OL].(2026-08-30)[2026-09-01].https://chinaxiv.org/abs/202608.00185.学科分类
航空航天技术
