MAPCON: Multi-Agent Predictive Coordination Network for Autonomous Interference Management in Mega-Constellation Satellite Systems
Accelerating Space Commerce, Exploration, and New Discovery Conference, ASCEND 2026, Wasington, DC, Amerika Birleşik Devletleri, 19 - 21 Mayıs 2026, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.2514/6.2026-3087
- Basıldığı Şehir: Wasington, DC
- Basıldığı Ülke: Amerika Birleşik Devletleri
- İstanbul Üniversitesi-Cerrahpaşa Adresli: Evet
Özet
The densification of Low Earth Orbit mega-constellations now exceeds 7,000 operational Starlink satellites, with industry projections surpassing 100,000 spacecraft by the mid-2030s. The resulting co-channel interference in shared Ku-band spectrum cannot be accommodated by static ITU coordination frameworks. This paper introduces MAPCON (Multi-Agent Predictive Coordination Network), a hierarchical multi-agent reinforcement learning framework for autonomous interference management across competing constellation operators. The architecture comprises decentralized satellite agents that exploit the deterministic predictability of orbital mechanics to forecast interference over a 600-second horizon and execute joint frequency, power, and beam allocation via Dueling Double DQN; constellation coordinators that perform QMIX value decomposition with Graph Neural Network-based conflict resolution over the evolving inter-satellite link topology; and a ground-based orchestrator that enables cross-operator federated aggregation under CKKS homomorphic encryption, allowing coordination without exposing proprietary spectrum strategies. Validated on a Walker-Delta constellation simulation with SGP4 dynamics and ITU-R S.1528 interference modeling, MAPCON jointly optimizes SINR, spectral efficiency, inter-operator fairness, and handover stability across a 400-dimensional action space, achieving a 20% composite-objective improvement over the strongest reactive baseline (game-theoretic best-response) and 44% over centralized water-filling. The framework generalizes zero-shot from 10 to 500 satellites, 10× beyond the training configuration, at ∼350 KB per-satellite communication per federated round under top sparsification, comparable to existing satellite federated learning frameworks. Hierarchical federation contributes an additional 8.5% gain over the non-federated variant, and ablation isolates the GNN conflict resolver as the dominant coordination mechanism, with a 2.8× larger performance loss on removal than any other module. A broader methodological finding, validated by a five-point server-side learning-rate sweep, is that cooperative MARL under QMIX requires conservative federated aggregation at =0.01: standard FedAvg at =1.0 induces a 6.5× collapse from the centralized checkpoint, a constraint applicable to any distributed system that combines federated learning with cooperative value factorization.