PUMA-Former: Uncertainty-Aware Agentic Computation Offloading in MEO-LEO Satellite Networks With Incomplete Information
Organizations: Department of Electronic Engineering, Tsinghua University, Beijing 100084, China · State Key Laboratory of Space Network and Communications, Tsinghua University, Beijing 100084, China, and also with the Beijing National Research Center for Information Science and Technology, Tsinghua University, Beijing 100084, China
Abstract
Satellite edge computing distributes mobile workloads across moving processors whose service capacity must be coordinated over time. In medium Earth orbit (MEO)--low Earth orbit (LEO) networks, periodic reports leave CPU states stale or missing, making offloading sensitive to uncertain capacity and shared-destination congestion. We propose PUMA-Former, an uncertainty-aware agentic framework for cooperative offloading under incomplete information that balances completed work, service delay, and execution cost. A MEO coordinator maintains distinct-report memory and closes the loop between available evidence, capacity estimates, allocation commands, and execution feedback. Latent-delay attention weights plausible measurement-age hypotheses, and historical innovations refine CPU estimates around the latest received report. Queue-error-aware fusion incorporates auxiliary pressure while accounting for backlog uncertainty. The inferred resource moments condition a structured Transformer allocator for joint destination, partition, bandwidth, CPU, and power decisions. Bounded risk-aware calibration then adjusts workload splits to reflect uncertainty and shared destination load. In traffic-driven development evaluation, PUMA-Former increases admitted compute-work completion by 8.62%, reduces mean modeled service delay by 34.78%, and reduces mean episode-P95 delay by 71.08% relative to ATLA-adapt. It also improves completion and normalized energy over fixed-half round-robin offloading. Service gains over the learning baselines involve higher normalized energy.
Figures & tables
| Direction | Established focus | Focus of this work |
|---|---|---|
| Delayed satellite state [ 9 ] | Delay-MDP and augmented state | Masked, age-aware state distribution and auxiliary update |
| Robust computation [ 10 ] | Uncertain task-computation requirements | Hidden available CPU and queue state |
| GATE / ATLA [ 12 , 13 ] | Structured encoders and hybrid resource decisions | Reception-delay inference and bounded offload correction |
| Contact / routing [ 16 , 17 , 18 ] | Visibility and diverse routing settings | Fluid communication/computation execution |
| Symbol | Meaning |
|---|---|
| Demand-window length and reachable neighbors | |
| New input bits, new CPU cycles, retained cycles | |
| Available fraction, nominal and executed CPU rate | |
| Proposed command, calibrated command, realized offload fraction | |
| Destination, partition, bandwidth, CPU, power controls | |
| Transferred bits, executed cycles, dropped cycles |
| Quantity | Online access | Role |
|---|---|---|
| Arrived resource values | Yes; possibly stale | Reports and memory |
| Reception age, masks, status | Yes | Evidence interpretation |
| Demand descriptors, topology | Yes | Tasks and candidate masks |
| Noisy pressure, validity | Yes; channel assumption | Auxiliary fusion |
| Current true CPU and queue | Hidden | Execution and training labels |
| Report generation time, true delay | Hidden | Channel simulation |
| Parameter | Value |
|---|---|
| Orbit planes / satellites per plane | 12 / 6 |
| Altitude / inclination | 550 km / |
| Minimum elevation / ISL neighbors | / at most 4 |
| Nominal orbital period | 5730 s |
| Demand window / episode length | 300 s / 288 windows |
| Nominal CPU / available fraction | 50 Gcycles/s / 0.35–1 |
| Reference | Completion (pp) | Mean delay (s) | P95 delay (s) |
|---|---|---|---|
| Fixed-half RR | 0.73 [0.51, 0.96] | [ , 0.04] | [ , 0.08] |
| GATE-adapt | 3.59 [3.13, 4.05] | [ , ] | [ , ] |
| ATLA-adapt | 6.99 [6.26, 7.71] | [ , ] | [ , ] |
| Prepared backbone | 0.62 [0.51, 0.74] | [ , ] | [ , 0.69] |
| Backbone + correction | 0.21 [0.18, 0.25] | [ , ] | [ , 0.20] |
| Variant | CPU NLL | Coverage (%) | CPU MSE |
|---|---|---|---|
| Prepared backbone | 95.87 | 2.31 | |
| Full | 91.03 | 1.11 | |
| No-delay | 89.76 | 1.16 | |
| No-history | 88.35 | 1.11 |