Joint Partitioning, Allocation, and Transmission Optimization for Federated Learning in Satellite Constellations via Multi-Task MARL

Orbital edge computing (OEC) is crucial for supporting space intelligence applications within satellite networks. However, individual satellites face resource constraints, and implementing distributed processing techniques, such as federated learning (FL), across multiple satellites introduces significant scheduling complexity. To address these challenges, we first model the key factors influencing complex satellite networks, including satellite constellations, regional resource demands, inter-satellite communication and routing, energy consumption, and battery aging—a novel aspect invoked by OEC operations. We propose an adaptive aggregation method to fundamentally improve communication efficiency in OEC-based FL. To enhance scheduling performance, we formulate a unified optimization problem that jointly considers data partitioning, resource allocation, and aggregation transmission tasks within a decentralized partially observable Markov decision process (Dec-POMDP) framework. Furthermore, we introduce an episodic-phase-recalling reward shaping (EPRS) method to correlate the influences across these phases. Inspired by multi-task learning, we propose an efficient multi-agent reinforcement learning (MARL) algorithm featuring a multi-head actor-critic (MH-AC) network structure and task-equalized adaptation (TEA) technology, designed to optimize latency, energy consumption, network traffic, and battery aging. Extensive experiments validate the effectiveness of the proposed method, showing a 29.9% reduction in total training time, an 11.5% reduction in network traffic, and superior overall performance compared to rule-based methods.