Joint Optimization of Flexible Inter-Satellite Link Scheduling and Routing for Hybrid Constellations via Attention-Enhanced MADRL

Abstract

This paper investigates the joint optimization of Flexible Inter-Satellite Link (FISL) scheduling and routing under tight coupling and mutual constraints. Traditional sequential methods break the equivalence of the original problem through decoupling, thereby leading to suboptimal performance. To overcome this limitation, a Joint FISL Scheduling and Routing Optimization (JFRO) framework is proposed to simultaneously determine FISL scheduling and routing, with a theoretical guarantee that its optimal objective value is no worse than that of sequential solutions, albeit at the cost of prohibitively high computational complexity. To address this challenge, an Attention-Enhanced Multi-Agent Deep Reinforcement Learning (AMADRL) solution is developed with a centralized training and decentralized execution paradigm and an attention mechanism to balance computational efficiency, decision performance, adaptability, and scalability, while theoretically guaranteeing convergence under consistency conditions. Simulation results demonstrate that AMADRL significantly outperforms baseline methods by achieving 57% higher weighted throughput than sequential baseline methods, reducing task delay, and enhancing FISL resource utilization, while maintaining lower inference latency than single-agent approaches. Notably, although the attention mechanism introduces additional computational overhead compared with MADRL without attention, the resulting performance gains make this cost entirely acceptable.