Source-linked AI summary
SatCraft: Agentic Optimization of Scheduling Policies for Satellite-Cellular Edge Services
Pengcheng Luo, Dengke Deng, Zenghui Zhang, Genke Yang, Jian Chu
TL;DR
Joint satellite–cellular edge-service scheduling must account for dependencies, uncertain access, and shared resource limits, while fixed policy structures offer limited scope to address failures in scheduling logic. SatCraft revises executable policies using complete-run feedback and selects among retained alternatives based on access observations and resource conditions. It achieves the highest completed-service value across four resource profiles, with gains over ADPRL, EdgeOPT, and TF-DDRL.
Problem
Joint scheduling must coordinate computation and transmission across dependent services and changing access conditions, while fixed policy structures offer limited scope to correct failures rooted in scheduling logic.
Source: B. LLM-Driven Policy Construction and… · p. 3, C. Shared Computing and Dual-Access… · p. 4Method
SatCraft uses LLM-guided execution feedback to revise policy code, evaluates and retains valid alternatives, and selects a policy using initial access observations and known resource conditions.
Source: A. Background and Motivation · p. 2, B. Contributions · p. 2Results
SatCraft achieves the highest completed-service value across four resource profiles, with gains of 3.80%–30.16% over ADPRL, EdgeOPT, and TF-DDRL.
Source: B. Contributions · p. 2Takeaways & Limitations
A varied policy library and context-dependent selection contribute to completed-service value, while current repair rounds provide a smaller incremental benefit in the validation ablations.
Source: D. Ablation of the Policy Library and… · p. 11
Abstract
from arXiv · showhide
Mobile edge services can combine local computation with satellite and cellular access to deliver time-sensitive results in areas with intermittent terrestrial coverage. Meeting their deadlines and quality requirements calls for joint computation and transmission scheduling that accounts for task dependencies, changing link capacity, and shared resource budgets. However, optimizing resource allocations or tuning parameters within a fixed policy structure offers limited scope to correct failures rooted in the scheduling logic. We propose SatCraft, a framework in which a large language model (LLM) agent uses execution feedback to revise the code of joint scheduling policies. Each revision undergoes policy checks and complete-run evaluation to assess its validity, runtime, and effects on service completion. Because policy performance varies with access conditions, SatCraft retains valid alternatives in a policy library and selects one for each episode using initial access observations and the known resource configuration. Experiments using Starlink and cellular measurement data show that SatCraft improves cumulative completed-service value by 3.61%-16.89% over baselines, including recent state-of-the-art scheduling methods ADPRL, EdgeOPT, and TF-DDRL. It achieves the highest completed-service value in all four resource profiles and averages 25.02 ms per scheduling call. These results demonstrate how LLM-guided policy optimization can improve complete service delivery under the computing, communication, and quality constraints of satellite-cellular networks.
I. INTRODUCTION
SatCraft addresses scheduling failures that arise from the interacting effects of processing choices, service dependencies, and shared resource use. It revises executable policies from complete-run feedback and selects among retained alternatives using access observations and resource conditions.
A. Background and Motivation S
- Joint scheduling must balance processing delay and reduced transmission demand while accounting for dependencies, changing access conditions, and limited cellular use.Execution feedback can expose failures such as delayed evidence, postponed priorities, or premature cellular activation.Source: A. Background and Motivation · pp. 1–2
B. Contributions
- SatCraft uses an LLM agent to revise joint-scheduling policy code from complete-service execution feedback, then checks and evaluates revisions before retaining valid alternatives.The evaluation assesses validity, runtime, and completed-service value.Source: A. Background and Motivation · p. 2, B. Contributions · p. 2
- The scheduling model evaluates processing and transmission decisions by whether services meet acceptance conditions before their deadlines, including downstream effects.It represents output timing, result quality, task dependencies, and satellite and cellular resource limits.Source: B. Contributions · p. 2
- After an initial access-observation period, SatCraft selects a policy using observed access and known resource conditions; the policy then adapts actions as tasks and resources change.The selected policy responds to arriving tasks and revealed quality.Source: B. Contributions · p. 2
- SatCraft achieves the highest cumulative completed-service value among ADPRL, EdgeOPT, and TF-DDRL, with gains of 3.80%–30.16% across four resource profiles.The evaluation uses measured Starlink and cellular rates and modeled service workloads.Source: B. Contributions · p. 2
II. RELATED WORK
Prior work connects edge-service outcomes to computing and communication resources, and LLM-driven methods make executable policy revisions testable. SatCraft’s setting adds the need to evaluate complete service workflows and match policies to access conditions.
A. Edge-Service Scheduling with Satellite and Terrestrial Access
- Prior edge-service work coordinates computation and communication, while quality adaptation and workflow scheduling connect resource decisions to output quality and downstream progress.The cited approaches span satellite–ground inference, adaptive data quality, and dependency-aware workflows.Source: A. Edge-Service Scheduling with… · p. 3
B. LLM-Driven Policy Construction and Optimization
- LLM-driven program search uses execution feedback to revise code, while evaluation ties the measured value of candidates to the workloads and constraints used to assess them.Related systems also motivate retaining complementary policies and selecting among them according to instance features or resource demands.Source: B. LLM-Driven Policy Construction and… · p. 3
- For satellite–cellular services, policy revisions can alter output readiness, downstream task needs, and resources left for later arrivals, so evaluation must follow complete-service outcomes and selection must use deployment context.The relevant context includes access observations and resource conditions.Source: B. LLM-Driven Policy Construction and… · p. 3
III. DUAL-ACCESS EDGE-SERVICE MODEL AND SCHEDULING OBJECTIVE
- The service model treats an assessment as a deliverable that may require multiple products or accept alternative combinations, with products sharing resources with routine uploads.Products may be transmitted raw or processed first, and delivered quality can affect downstream tasks or service completion.Source: §III · p. 4
A. Service Requests and Completion Conditions
Service completion depends on delivery records satisfying dependency and acceptance conditions, including quality requirements and deadlines. Product mode and shared resource constraints determine when outputs can be delivered and which services remain feasible.
A. Service Requests and Completion Conditions
- A service earns its value once, at its first time that its delivery and quality acceptance predicate holds by the deadline.A product may complete a service directly or make downstream work feasible.Source: A. Service Requests and Completion… · p. 4
B. Product Processing and Output Timing
- Raw outputs can enter transmission immediately, whereas processed outputs require computing slots and become transmissible only after computation finishes.Processing reduces transmission volume but consumes computing capacity and can change delivered quality; mode remains fixed until delivery or expiry.Source: B. Product Processing and Output Timing · p. 4
C. Shared Computing and Dual-Access Resources
- Products share one transmission queue and compete for local computation, Starlink capacity, and cellular capacity constrained by startup delay and usage budgets.Queue ordering sets transmission priority, while cellular activation timing must account for output availability and remaining allowance.Source: C. Shared Computing and Dual-Access… · pp. 4–5
D. Online Decisions and Objective
The policy maps available observations and memory to joint scheduling actions, while the objective values services completed on time. Candidate policies are assessed through complete-run evaluation, under the stated historical-capacity assumption.
D. Online Decisions and Objective
- The policy observes arrived-product progress, past link capacities, revealed qualities, remaining budgets, cellular state, and known resource configuration, but not current or future capacity or undelivered quality.A policy may retain memory of previous observations and decisions.Source: D. Online Decisions and Objective · p. 5
- The scheduling objective aggregates the value of services completed on time, with candidate policies evaluated through complete runs and fixed before evaluation-set comparison.The evaluated policy set is constrained to causal policies that satisfy dependency, timing, and resource rules.Source: D. Online Decisions and Objective · p. 5
- The assumed historical service-capacity estimates for both access paths require deployment calibration of the measurement or probing cost.The evaluation prototype updates samples for both links at every step, including when cellular transmission is disabled.Source: D. Online Decisions and Objective · p. 5
IV. SATCRAFT: POLICY OPTIMIZATION AND EXECUTION
SatCraft combines revision of executable joint-scheduling policies with context-dependent selection among valid alternatives. Its LLM agent uses execution feedback to revise code, while complete-run evaluation checks candidates before retention.
A. Framework and Optimization Decomposition
- SatCraft jointly optimizes executable scheduling policies and selects among valid alternatives according to access conditions and resource configuration.Policies map observations and internal memory to joint actions; the selector uses initial access features and a known resource configuration.Source: §IV · p. 5, A. Framework and Optimization… · p. 6
B. Execution-Feedback-Driven Policy Optimization
- The LLM agent receives policy code, case outcomes, and execution traces in a budgeted, multi-turn revision process.The interface and repair direction specify the policy constraints and focus revisions on an observed scheduling failure.Source: B. Execution-Feedback-Driven Policy… · p. 6
- Complete-run evaluation and policy checks determine whether a revised candidate is retained, while condition-specific utility guides its deployment.A candidate can remain useful under particular access conditions even if its aggregate value is lower.Source: B. Execution-Feedback-Driven Policy… · pp. 6–7
- Candidate policies revise contention prediction, supplementary-access activation, and quality-dependent downstream task handling.These structural changes are assessed through the same complete-run evaluation interface.Source: B. Execution-Feedback-Driven Policy… · p. 7
C. Context-Dependent Policy Selection
SatCraft fits a shallow policy selector for each known resource configuration, using initial access observations to choose among library policies by development-set utility. It then freezes the selectors and referenced policy code for evaluation and deployment.
C. Context-Dependent Policy Selection
- For each resource configuration, SatCraft fits a shallow tree that maps initial access observations to a policy selected by development-set utility.The feature vector summarizes link capacities, variation, trends, zero-capacity fractions, and final-to-mean capacity ratios.Source: C. Context-Dependent Policy Selection · p. 7
- The fitter recursively selects the largest positive-gain split subject to leaf-size constraints, with maximum depth three and at least three development traces per leaf.Candidate thresholds lie between adjacent observed feature values.Source: C. Context-Dependent Policy Selection · p. 7
- SatCraft compares a newly fitted tree with three preset tree structures relabeled using the current policy library, then selects among four trees by development utility.The chosen selectors and referenced policy code are exported and frozen.Source: C. Context-Dependent Policy Selection · p. 7
D. Policy Deployment and Online Joint Scheduling
At the start of each episode, SatCraft uses initial access observations and the known resource configuration to choose one frozen policy. That policy then adapts its actions and memory to changing task and resource states at each scheduling slot.
D. Policy Deployment and Online Joint Scheduling
- The selector chooses one policy for the episode after observing the initial access period, and the selected policy remains fixed for that episode.The policy code and selector are frozen after offline optimization.Source: D. Policy Deployment and Online Joint… · pp. 7–8
- At each slot, the selected policy uses current observations to produce joint scheduling actions and update its memory.The executor applies computing and transmission actions, then updates task progress, budgets, delivery records, and completed services.Source: D. Policy Deployment and Online Joint… · p. 8
- Within an episode, the policy adapts to arrivals, dependencies, delivered quality, cellular state, and remaining budgets.The resulting progress and delivery outcomes become the next observation.Source: D. Policy Deployment and Online Joint… · p. 8
V. PERFORMANCE EVALUATION
The evaluation pairs measured Starlink and cellular access traces with modeled workloads that include task dependencies and quality requirements. It compares SatCraft with five scheduling baselines under shared cases and resource constraints.
1) Measured Access and Service Workloads:
- The study uses 60 selected NUWiNS drive traces pairing Starlink with AT&T or Verizon throughput measurements, replayed in 119 slots.The selected segments preserve run boundaries, and eligible traces average at least 1 Mbit/s on Starlink during the first eight slots.Source: 1) Measured Access and Service Workloads · p. 8
- Each modeled case contains six overlapping task batches, 24 product tasks, and 12 services sharing one processor and satellite-cellular access.Tasks may be sent raw or processed locally before their smaller outputs are transmitted; processed-result quality is revealed after delivery.Source: 1) Measured Access and Service Workloads · p. 8
- The 960 cases from 60 measured runs are divided into development, validation, and test sets, with 30 distinct measured traces in the test set.All methods face identical tasks, link sequences, quality realizations, and computing, transmission, and cellular-resource constraints.Source: 1) Measured Access and Service Workloads · p. 8
2) Compared Scheduling Methods:
- Five baselines cover deadline priorities, rolling-horizon planning, resource pricing, and reinforcement learning, including ADPRL, EdgeOPT, and TF-DDRL.The three published edge-computing methods are adapted to the common service model, joint-action interface, and completed-service-value objective.Source: 2) Compared Scheduling Methods · pp. 8–9
3) Implementation Details:
The evaluation uses fixed policies under shared test conditions, measures completed-service value and service fulfillment, and finds that SatCraft leads five baselines in cumulative value and complete delivery.
3) Implementation Details:
- ADPRL and TF-DDRL are trained for approximately 1.5 million transitions per seed, and each seed’s checkpoint is selected by development-set completed-service value.Their rewards combine incremental completed-service value with shaping based on observable task progress.Source: 3) Implementation Details · p. 9
4) Metrics and Statistical Procedure:
- Cumulative completed-service value is the primary metric; completed services and fully served cases are also reported, with the latter requiring all 12 services to meet deadlines and quality requirements.ADPRL and TF-DDRL results average three training seeds, with 95% confidence intervals from 20,000 paired bootstrap resamples clustered by measured trace.Source: 4) Metrics and Statistical Procedure · p. 9, B. Overall Service-Completion… · p. 10
3) Implementation Details:
- All methods use the same 360 test cases, link sequences, arrivals, quality realizations, resource budgets, and fixed evaluation configurations.Complete test runs use a 2-s callback limit and a 90-s case limit.Source: B. Overall Service-Completion… · pp. 9–10, 3) Implementation Details · p. 9
B. Overall Service-Completion Performance
- SatCraft achieves the highest cumulative completed-service value, 23,672, exceeding five baselines by 3.61%–16.89%.It completes 3,935 of 4,320 requested services (91.09%) and fully serves 270 of 360 cases (75.0%), versus 72–223 fully served cases for the baselines.Source: B. Overall Service-Completion… · p. 10, Figure 5 · p. 10, Table III · p. 10
C. Performance Under Resource Pressure
SatCraft leads across all four resource profiles; ablations show gains from its varied policy library and context-dependent selection, while repair and scheduling incur measured offline and online costs.
C. Performance Under Resource Pressure
- SatCraft achieves the highest cumulative completed-service value in every resource profile, with all 20 relative gains positive and gains of 3.80%–30.16% over ADPRL, EdgeOPT, and TF-DDRL.The advantage persists as constraints shift among computation, transmission, and cellular activation; its largest gains over those three methods occur under network pressure.Source: C. Performance Under Resource Pressure · p. 10
D. Ablation of the Policy Library and Context-Dependent Selection
- SatCraft’s validation value of 14,792 exceeds all five ablations; the full library adds 1.62% over parameter-only policies, and context-dependent selection adds 1.90% over fixing one policy per profile.These comparisons support generated policy structures and context-dependent selection when processed-result qualities change on the same access traces.Source: D. Ablation of the Policy Library and… · p. 11, Table V · p. 11
- Removing all current repairs reduces value by 104, a 0.71% SatCraft gain; removing future-load or new link-repair policies separately reduces value by 32 or 44.Because the selector is refitted after policy removal, these differences measure library changes and fitted selection jointly.Source: D. Ablation of the Policy Library and… · p. 11
E. Policy Optimization Cost and Runtime Overhead
- Across both repair branches, 11 candidate policies were evaluated, nine passed validity checks, and the process used 79 model responses and 218,353 noncached tokens.The best future-load and link-repair scores rose by 28 and 72 over the common starting point; branch record windows were 16.73 and 14.81 minutes.Source: E. Policy Optimization Cost and… · p. 11, Figure 7 · p. 11, Table VI · p. 11
- SatCraft’s mean scheduling callback takes 25.02 ms, and its longest takes 1.180 s, below the shared 2-s limit.EDF, ADPRL, and TF-DDRL have lower mean callback times of 0.11, 0.51, and 2.67 ms, respectively.Source: E. Policy Optimization Cost and… · pp. 11–12, Table VII · p. 12
F. Repair Mechanisms and Execution Traces
Paired execution traces show how future-load anticipation and cellular-activation repairs change resource use and service completion, while development-set results show that their effects vary across cases. The conclusion situates these repairs within SatCraft’s broader policy-optimization framework and identifies future-work scope.
F. Repair Mechanisms and Execution Traces
- The paired cases hold code, tasks, measured links, and quality outcomes fixed except for the component switch; Fig. 8 plots resource states and completed value.Case A compares future-load anticipation, while case B compares cellular activation; the cases share an access trace and computing-pressure profile but use different quality seeds.Source: F. Repair Mechanisms and Execution… · p. 12, Figure 8 · p. 12
- At slot 88, disabled anticipation leaves 2.28 Mbit and five activation slots versus 325.21 Mbit and 21 slots when enabled.The enabled variant completes two additional services, raising value from 20 to 36 while using less cellular traffic and computing time overall; the gain comes from allocation across batches.Source: Figure 8 · p. 12, F. Repair Mechanisms and Execution… · p. 12, Table VIII · p. 12
- Cellular repair lowers traffic from 513.30 to 350.16 Mbit and activation use from 60 to 54 slots, with 43 computing slots and value rising from 40 to 48.The disabled variant activates cellular access during slots 72–74 without transmitting; the enabled variant avoids those activations and changes subsequent transmission scheduling.Source: Figure 8 · p. 12, F. Repair Mechanisms and Execution… · pp. 12–13, Table VIII · p. 12
- Future-load anticipation adds 240 in cumulative value, improving 71 development cases and worsening 49; cellular-activation repair adds 96 with 19 improvements and four regressions.Quality-path handling adds only eight, with most cases unchanged; the effects depend on case conditions.Source: Table IX · p. 13, F. Repair Mechanisms and Execution… · p. 13
VI. CONCLUSION
- SatCraft combines LLM-guided policy optimization with context-dependent selection, and the authors identify calibration, operational costs, and workflow testing as future work.Experiments using Starlink and cellular measurement data report that SatCraft outperforms all baselines in test-set cumulative completed-service value.Source: §VI · p. 13
APPENDIX A REPRODUCIBLE PROMPT DESCRIPTIONS FOR THE LLM
Appendix A describes initial and continuation prompt templates for revising executable scheduling policies. The templates specify the task and interface, then use complete-run feedback to guide bounded revisions.
APPENDIX A REPRODUCIBLE PROMPT DESCRIPTIONS FOR THE LLM
- The appendix presents two parameterized prompt templates: an initial template and a continuation template that incorporates execution feedback.The initial template specifies the task, policy interface, repair direction, and budgets.Source: Appendix A · p. 13
A. Task Interface and Repair Direction
- The initial optimization task is to maximize cumulative value from services delivered by their deadlines with acceptable quality, subject to the model and resource budgets..Source: A. Task Interface and Repair Direction · p. 13
- The policy interface returns mode choices, a computing task, transmission-queue order, cellular activation, and a cellular sending cap, while requiring available observations and memory.Static checks cover imports, file access, and dynamic evaluation.Source: A. Task Interface and Repair Direction · p. 13
- The initial prompt directs repairs toward future demand and contention or cellular activation and output timing, with up to six submissions and 12 diagnostics per branch.It also sets policy callback and episode limits at 2 and 90 s.Source: A. Task Interface and Repair Direction · p. 13
B. Continuation After Execution Feedback
- The continuation template keeps the task and interface fixed while adding execution feedback to guide the next revision..Source: B. Continuation After Execution Feedback · p. 13
- Continuation feedback includes value, validity, resource, and runtime summaries, relevant cases and traces, errors, timeouts, and remaining budgets.The prompt asks the agent to distinguish individual action changes from service-completion outcomes and reduce execution complexity after errors or timeouts.Source: B. Continuation After Execution Feedback · p. 13
- Each continuation returns revised policy code, the reason for revision, and corresponding complete-evaluation results..Source: B. Continuation After Execution Feedback · p. 13