Leader Federated Learning Optimization Using Deep Reinforcement Learning for Distributed Satellite Edge Intelligence

The deployment of satellite mobile edge computing (SMEC) incorporating artificial intelligence (AI) in low Earth orbit (LEO) constitutes satellite edge intelligence (SEI), which is promising to achieve autonomous processing of space missions on board driven by massive data. However, individual satellites with constrained resources and insufficient samples learn inefficiently, while the spatio-temporal constraints of large-scale LEO networks make collaborative training difficult. In this paper, a leader federated learning (FL) architecture for distributed SEI (SELFL) is proposed. By evaluating the connectivity and load of the dynamic constellation, the global and local parameters of the shared AI model are transmitted and updated continuously between the elected leader and other follower satellites based on the established inter-satellite link, which realizes efficient self-evolution of SELFL independent of the ground. Also we introduce a deep reinforcement learning-based resource allocation strategy for SELFL, which leverages the distributed proximal policy optimization (DPPO) to optimize the computing capability and transmit power of satellites for accelerating FL and reducing energy consumption. This method not only updates stably utilizing adaptive learning steps, but also improves sample efficiency with multiple parallel workers. The simulation results demonstrate the proposed SELFL optimization scheme effectively reduces the total energy consumption and training time by ensuring the AI model accuracy, and outperforms the benchmark algorithms. IEEE